A method and system for detecting subgrade disease based on active phased array radar

Abstract

The application relates to the technical field of nondestructive testing of highway subgrade diseases, and discloses a subgrade disease detection method and system based on an active phased array radar, which realizes continuous acquisition of different buried depths, different scales and different types of subgrade diseases by utilizing the multi-frequency synchronous detection capability and multi-channel beam synthesis capability of the active phased array radar, improves the integrity of underground disease response information and the detection coverage capability, realizes integrated output of nondestructive detection of diseases by utilizing a YOLOv10s model, a pixel-physical scale mapping relationship and a Beidou positioning binding mechanism, and improves the standardization degree, spatial positioning accuracy and engineering application value of subgrade disease detection results.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for roadbed defects, and in particular to a method and system for detecting roadbed defects based on active phased array radar. Background Technology

[0002] As the fundamental load-bearing structure in highway engineering, the roadbed is continuously affected by various factors during long-term service, including rainwater infiltration, temperature and humidity cycles, adverse geological conditions, repeated traffic loads, and construction disturbances. This makes it highly susceptible to defects such as cracks, softening, settlement, and cavities. These defects often have the characteristics of being insidious in their formation, gradually expanding, and difficult to identify through early visual observation. If they are not detected and addressed in a timely manner, they will not only reduce the overall service life of the road but may also further induce pavement damage, structural instability, and increased operational safety risks. With the continuous expansion of the highway network and the increasing workload of maintenance and inspection, how to achieve rapid, accurate, and continuous detection of roadbed defects has become a key technical issue in highway maintenance management.

[0003] While manual inspections are useful for verifying localized excavations and making experience-based judgments, they are heavily influenced by the inspectors' experience, the working environment, and traffic conditions. This results in low efficiency, high randomness, insufficient repeatability, and potential disruption to traffic flow, making it difficult to meet the practical needs of large-scale, high-frequency, and preventative inspections. To improve inspection efficiency and coverage, mounting inspection equipment on vehicles for onboard non-destructive testing has become an important development direction for highway subgrade inspection. Existing vehicle-mounted inspection equipment mainly includes vehicle-mounted lidar, vehicle-mounted rolling electrodes, and ground-penetrating radar. Among them, vehicle-mounted lidar has weak penetration capability and is mainly suitable for detecting superficial defects such as road surface cracks and surface damage, but it is difficult to effectively identify defects inside the roadbed. Vehicle-mounted rolling electrode detection is greatly affected by road smoothness, contact state and medium resistivity, and the detection speed is relatively slow and the resolution is limited, making it more suitable for detecting local and shallow defects. Ground penetrating radar, on the other hand, has the advantages of fast detection speed, strong adaptability, high accuracy and continuous operation, and has become the mainstream non-destructive testing method for detecting defects inside the roadbed.

[0004] Furthermore, most existing ground-penetrating radar systems still employ passive phased array radar or operate with relatively fixed frequency configurations. This type of technology provides a relatively limited range of detection information within the same detection period, and is prone to insufficient response to weak echo targets formed at different burial depths, scales, and media conditions. Simultaneously, the long-term stability and reliability of related components are also constrained under high-power operation. In contrast, active phased array radar can transmit multiple detection waves of different frequencies simultaneously and achieve rapid scanning of different detection angles, depths, and types of defects through coordinated operation of different channels. It offers advantages such as good beam directivity, low signal transmission loss, strong equipment stability, and small detection blind spots, making it more suitable for continuous inspection of roadbed defects in complex scenarios.

[0005] When the detection vehicle travels at high speed, especially when passing through bridge joints, undulating road sections, abrupt changes in semi-rigid base layers, or sections with strong short-term vibrations, the relative attitude between the detection system and the road surface changes continuously within a short period. This causes rapid fluctuations in the relative distance, incident angle, and equivalent coupling state between the radar antenna array and the measured medium. Simultaneously, the vehicle's longitudinal and vertical vibrations increase the spatial interval between adjacent sampling times and produce instantaneous shifts. This causes the echo of the same defect, which should be stably distributed along a continuous trajectory, to be dispersed across multiple adjacent spatiotemporal frames. Because ground-penetrating radar echoes are affected by the inhomogeneity of the underground medium, interlayer reflection, and local scattering effects, when these dynamic disturbances are superimposed with weak defect echoes, the energy distribution of the same defect in adjacent image frames tends to become dispersed and non-centralized. That is, the response energy that should have been continuously concentrated in a local area is split across multiple sampling positions. Meanwhile, due to the drift of sampling centers between adjacent frames, the projection position of the lesion boundary in the image will also exhibit irregular jumps. For lesions that are small in scale and have weak contrast, the effective echo energy within a single frame will be weakened, resulting in short-term weakening or even intermittent disappearance of local textures. In other words, in this type of scenario, it is not the lesion itself that changes, but rather the continuity of the target response in the temporal and spatial dimensions is disrupted during the detection process. This leads to higher instability in subsequent lesion identification, boundary localization, and geometric parameter extraction based on single-frame features. If only single-frame images are relied upon for lesion anomaly judgment, it is easy to cause problems such as missed detection of minor lesions, boundary jitter of the same lesion, and increased deviations in the estimation of lesion size and burial depth, thereby affecting the consistency of detection results and the engineering application value.

[0006] Therefore, how to achieve roadbed defect detection based on multi-frequency detection of active phased array radar, so as to improve the detection accuracy, efficiency, adaptability to complex scenarios, detection robustness and intelligence level, while improving the response stability, boundary continuity and recognition consistency of defect targets in continuous image frames under complex dynamic acquisition scenarios, remains a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] This invention provides a method and system for detecting roadbed defects based on active phased array radar, in order to solve the problems of poor detection effect, unstable expression of defect targets in adjacent frames, boundary drift, and increased probability of missing small defects in existing roadbed defect detection methods under dynamic fluctuation conditions.

[0008] To achieve the above objectives, the present invention provides a method for detecting roadbed defects based on active phased array radar, the method comprising the following steps: Acquire the multi-frequency raw echo signal corresponding to the target road segment, and perform multi-channel beamforming processing on the multi-frequency raw echo signal to generate synthesized echo data; The synthesized echo data is subjected to analog-to-digital conversion, spatiotemporal frame accumulation, and auxiliary information embedding processing to generate a labeled signal frame sequence; wherein, the auxiliary information includes at least BeiDou positioning information and device detection speed information; The labeled signal frame sequence is subjected to clutter suppression, noise suppression and data reconstruction processing in sequence, and the reconstructed data is converted into a B-Scan image sequence; The B-Scan image sequence is subjected to image enhancement, invalid frame removal and multi-scale spatiotemporal co-stabilization processing to generate an image sequence to be identified. The image sequence to be identified is input into the pre-trained YOLOv10s disease identification model to generate disease identification results. Based on the Beidou positioning information, equipment detection speed information and disease identification results, physical coordinate mapping, disease level classification, result binding and detection report output are performed.

[0009] Optionally, the multi-frequency raw echo signal corresponding to the target road segment is acquired, and multi-channel beamforming processing is performed on the multi-frequency raw echo signal to generate synthesized echo data, specifically including: Acquire the original echo signals corresponding to each channel of the active phased array radar, and establish the correspondence between the original echo signals and the array element positions according to the channel number; Time delay compensation is performed on the original echo signals of each channel based on the beam pointing angle, and multi-channel beamforming processing is performed to output the final synthesized output signal as synthesized echo data.

[0010] Optionally, the synthesized echo data undergoes analog-to-digital conversion, spatiotemporal frame accumulation, and auxiliary information embedding processing to generate a labeled signal frame sequence, specifically including: The synthesized echo data is subjected to analog-to-digital conversion processing according to the sampling time to obtain single-frame digital signal data; Multiple consecutive single-frame digital signal data are accumulated in the sampling order to form a spatiotemporal digital signal frame corresponding to the detection process; Read the BeiDou positioning information, device detection speed information, and acquisition time information corresponding to each of the aforementioned spatiotemporal digital signal frames; The BeiDou positioning information, equipment detection speed information, and acquisition time information are written into the corresponding spatiotemporal digital signal frames to obtain a labeled signal frame sequence. The labeled signal frame sequence is sorted and stored according to the frame number.

[0011] Optionally, clutter suppression, noise suppression, and data reconstruction processing are sequentially performed on the labeled signal frame sequence, and the reconstructed data is converted into a B-Scan image sequence, specifically including: The data corresponding to each point in the B-Scan image is extracted and subjected to mean filtering to obtain clutter-suppressed data. After clutter suppression, the data is standardized, the standardized correlation coefficient matrix is ​​calculated, and the characteristic equation of the correlation coefficient matrix is ​​solved to obtain... n Each eigenvalue and its corresponding eigenvector; The number of principal components is determined based on the cumulative contribution rate, and the normalized matrix is ​​reconstructed based on the principal components to generate a denoised B-Scan image sequence.

[0012] Optionally, image enhancement and invalid frame removal processes are performed on the B-Scan image sequence, specifically including: The CLAHE algorithm is used to perform adaptive contrast enhancement processing on the B-Scan image sequence; After adaptive contrast enhancement, Gaussian blur is applied to the image, followed by Laplacian sharpening. After feature enhancement processing, the images are uniformly converted into standard grayscale images and uniformly scaled to the preset model input size, while the pixel values ​​are standardized. The images in the unified format are filtered frame by frame to remove all black frames, all white frames, frames with signal interruption, and blurry frames with a resolution less than a preset value, resulting in a preprocessed image sequence.

[0013] Optionally, multi-scale spatiotemporal co-stabilization processing is performed on the B-Scan image sequence to generate an image sequence to be identified, specifically including: For adjacent image frames acquired continuously along the detection direction, the longitudinal displacement between the current frame and each adjacent frame is calculated based on the device detection speed information and the sampling time interval. Based on the longitudinal displacement, position alignment processing is performed on the candidate abnormal regions in each adjacent frame, and the local response intensity, edge continuity and texture aggregation degree after alignment are extracted under different scale windows to obtain multi-scale alignment response data. Calculate the collaborative stability score based on the current frame candidate anomaly region score and multi-scale alignment response data; The scale consistency constraint is applied to the collaborative stability score to obtain the weights of stable candidate regions; When the weight of the stable candidate region reaches a preset threshold, the corresponding candidate abnormal region is retained, and its boundary range is smoothly updated to generate an image sequence to be identified.

[0014] Optionally, when the weight of the stable candidate region reaches a preset threshold, the corresponding candidate abnormal region is retained, and its boundary range is smoothly updated to generate an image sequence to be identified, specifically including: Extract the initial boundary parameters of the candidate anomaly region in the current frame, read the boundary parameters of the corresponding regions in the adjacent frames before and after it that are aligned with its position, and calculate the boundary update result; The updated boundary parameters are written back to the candidate anomaly region of the current frame to obtain a boundary-stable image sequence to be identified.

[0015] Optionally, the image sequence to be identified is input into a pre-trained YOLOv10s disease identification model to generate disease identification results, specifically including: Training, validation, and test sets were constructed using historical B-Scan disease image samples. Data augmentation processing was performed through random rotation, brightness adjustment, Gaussian blur, mosaic enhancement, and small object cropping to obtain a pre-trained YOLOv10s disease recognition model. The image sequence to be identified is input into the pre-trained YOLOv10s disease identification model, which outputs the disease category, bounding box coordinates, and confidence score to generate disease identification results. Based on the pixel physical scale mapping relationship, the bounding box pixel coordinates are converted into actual physical coordinates, and the actual width, height and depth parameters of the lesion are calculated; Diseases are classified into different levels based on their size, burial depth, and confidence level, generating structured disease result data.

[0016] Optionally, based on the BeiDou positioning information, equipment detection speed information, and disease identification results, physical coordinate mapping, disease level classification, result binding, and detection report output are performed, specifically including: The structured disease result data is associated one by one with the corresponding B-Scan image identifier, Beidou positioning coordinates, equipment detection speed information and acquisition time information to form a disease positioning data structure; Based on the disease location data structure, output the denoised B-Scan image and disease identification results, and compare the denoised B-Scan image with the original B-Scan image to generate verification marker data; Upon receiving the correction result generated based on the verification marker data, the structured disease result data is updated to obtain the verified disease result data; A standardized roadbed defect detection report is generated based on the reviewed defect data.

[0017] Furthermore, to achieve the above objectives, the present invention also provides a roadbed defect detection system based on an active phased array radar, comprising: The acquisition module is used to acquire the multi-frequency raw echo signal corresponding to the target road segment, and perform multi-channel beamforming processing on the multi-frequency raw echo signal to generate synthesized echo data. The generation module is used to perform analog-to-digital conversion, spatiotemporal frame accumulation, and auxiliary information embedding processing on the synthesized echo data to generate a labeled signal frame sequence; wherein, the auxiliary information includes at least BeiDou positioning information and device detection speed information; The reconstruction module is used to sequentially perform clutter suppression, noise suppression and data reconstruction processing on the labeled signal frame sequence, and convert the reconstructed data into a B-Scan image sequence; The processing module is used to perform image enhancement, invalid frame removal and multi-scale spatiotemporal co-stabilization processing on the B-Scan image sequence to generate an image sequence to be identified; The output module is used to input the image sequence to be identified into the pre-trained YOLOv10s disease identification model, generate disease identification results, and perform physical coordinate mapping, disease level classification, result binding and detection report output based on the Beidou positioning information, equipment detection speed information and disease identification results.

[0018] The beneficial effects of this invention are as follows: Detection accuracy and efficiency are significantly improved. Utilizing multi-frequency synchronous detection and multi-channel beamforming technology of the active phased array radar, the radar's beam directivity is greatly enhanced, enabling precise differentiation between normal roadbed layers and millimeter-level micro-diseases. A customized YOLOv10s model, combined with a CLAHE adaptive contrast enhancement and Laplacian sharpening preprocessing strategy, further improves defect identification accuracy and significantly reduces the missed detection rate of small target defects. Simultaneously, the active phased array radar supports vehicle-mounted mobile detection from 0-80km / h, and combined with YOLOv10s' high-speed processing capability of tens of frames per second, continuous and uninterrupted detection of the entire roadbed cross-section is achieved.

[0019] The system exhibits superior adaptability to complex scenarios and robustness in detection. Utilizing the multi-frequency electromagnetic wave synchronous transmission technology of active phased array radar, it can adapt to different road surface types such as asphalt, concrete, and gravel, as well as different geological conditions such as soft soil, sand, and rock, solving the problem of poor adaptability of traditional single-frequency radar to complex geological conditions. Employing a combined denoising technique of mean filtering and principal component analysis, it effectively filters various interference signals from underground stray media and equipment vibrations, preserving clear defect features even in complex scenarios with strong interference. The use of CLAHE adaptive contrast enhancement, Gaussian blur denoising, and Laplacian sharpening for small target enhancement preprocessing significantly improves the robustness of the customized YOLOv10s model in defect identification under low contrast and complex backgrounds.

[0020] Dynamic scene detection offers advantages. By utilizing adaptive contrast enhancement, small target feature enhancement, and multi-scale spatiotemporal collaborative stabilization processing, the energy expression, boundary position, and texture features of disease targets remain highly consistent across consecutive image frames. This improves the accuracy of small target disease identification and the stability of boundary extraction in complex dynamic scenes. It addresses the problems of unstable disease target expression, boundary drift, and increased probability of missed detection of minor diseases in adjacent frames under dynamic fluctuation conditions in existing roadbed disease detection methods.

[0021] The level of intelligence has been significantly improved. The equipment can be directly mounted on existing highway inspection vehicles, powered by the vehicle's onboard power supply, achieving full automation of the inspection process and greatly reducing the operational threshold. The customized YOLOv10s model has a powerful ability to simultaneously identify multiple types of defects, eliminating the need for separate model training for different defects, thus improving the intelligent identification capability of defects. It also supports lightweight model deployment and subsequent algorithm upgrades, and can be continuously optimized according to new needs and scenarios in roadbed defect detection. The system can automatically perform associated archiving of "defect type + geometric parameters + geographic coordinates + radar image", generating standardized, multi-format inspection reports. It can seamlessly integrate with existing highway maintenance management systems, enabling standardized input and management of defect data, and providing core data support for the construction of an intelligent highway maintenance management system. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the roadbed defect detection method based on active phased array radar of the present invention. Figure 2 This is a schematic diagram of the detection process of the present invention in one specific embodiment; Figure 3 This is a schematic diagram illustrating the principle of the roadbed defect detection architecture based on active phased array radar of the present invention. Figure 4 This is a schematic diagram of the roadbed defect detection system based on active phased array radar according to the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] This invention provides a method for detecting roadbed defects based on active phased array radar, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the roadbed defect detection method based on active phased array radar according to an embodiment of the present invention.

[0025] In this embodiment, a method for detecting roadbed defects based on active phased array radar is provided, the method comprising the following steps: S1: Obtain the multi-frequency raw echo signal corresponding to the target road segment, and perform multi-channel beamforming processing on the multi-frequency raw echo signal to generate synthesized echo data.

[0026] Specifically, the original echo signals corresponding to each channel of the active phased array radar are acquired, and the correspondence between the original echo signals and the array element positions is established according to the channel number; the time delay compensation of the original echo signals of each channel is performed based on the beam pointing angle, and multi-channel beamforming processing is performed, and the final synthesized output signal is output as synthesized echo data.

[0027] In this embodiment of the invention, step S1 is mainly used to establish the original data foundation for the subsequent entire disease detection process. Because different diseases within the roadbed exhibit significant differences in burial depth, scale, media boundaries, and geometry, echo signals acquired under single-frequency or single-channel conditions can often only characterize underground anomaly responses in a limited dimension, making it difficult to simultaneously address the detection needs of different targets such as shallow fine cracks, localized softening zones, and deeper cavities. Therefore, this invention employs an active phased array radar as the core of data acquisition, enabling multiple T / R components to synchronously transmit radio frequency detection signals of different frequencies during the same detection period, with corresponding echoes received by the corresponding channels. This allows for joint detection of the target area from different frequency response dimensions and different array channel dimensions. This approach not only improves the adaptability to different disease types but also provides a more abundant source of original information for subsequent beamforming.

[0028] Specifically, in one executable implementation, the radar control subsystem can uniformly schedule the antenna array subsystem, T / R component subsystem, and signal processing subsystem according to a preset detection task, so that each array element transmits detection waves at the corresponding frequency configuration at the set detection time and receives echo signals in the target area. Then, the original echo signals corresponding to each channel can be read sequentially according to the channel number, and a correspondence between the original echo signals of each channel and the array element position, sampling time, and frequency configuration can be established. Since there are spatial distance differences between different array elements and the reference array element, the arrival times of echoes from different channels usually have a certain offset. If these echo signals are directly superimposed, it will introduce beam energy diffusion and target response defocusing problems. Therefore, before beam combining, this invention first performs time delay compensation processing on the original echo signals of each channel based on the beam pointing angle, and then performs multi-channel beam combining. The specific expression is: ; in, For the final synthesized output signal, Where t is the beam pointing angle, t is time, and n is the total number of channels. Let i be the original echo signal of the i-th channel. Let be the distance from the i-th array element to the reference array element, and v be the propagation speed of underground electromagnetic waves. Using the above formula, each channel achieves time alignment and superposition under a unified reference condition, concentrating the target responses originally dispersed in different channels into the same beam direction, thus forming synthetic echo data with more concentrated energy and clearer directionality. It should be further noted that the selection method of the reference array element is not fixed; as long as time delay compensation for different channels can be performed under the same reference standard, the technical effect of this invention can be achieved. Simultaneously, the propagation speed of underground electromagnetic waves can be preset based on known geological conditions or determined through calibration before detection.

[0029] S2: Perform analog-to-digital conversion, spatiotemporal frame accumulation, and auxiliary information embedding processing on the synthesized echo data to generate a labeled signal frame sequence; wherein, the auxiliary information includes at least BeiDou positioning information and device detection speed information.

[0030] Specifically, the synthesized echo data undergoes analog-to-digital conversion processing according to the sampling time to obtain single-frame digital signal data; multiple consecutive single-frame digital signal data are accumulated according to the sampling order to form a spatiotemporal digital signal frame corresponding to the detection process; the BeiDou positioning information, equipment detection speed information, and acquisition time information corresponding to each spatiotemporal digital signal frame are read; the BeiDou positioning information, equipment detection speed information, and acquisition time information are written into the corresponding spatiotemporal digital signal frame to obtain a labeled signal frame sequence; the labeled signal frame sequence is sorted and stored according to the frame sequence number.

[0031] In this embodiment of the invention, step S2 is used to convert the synthetic echo data obtained in the previous stage into a digital signal form that is convenient for subsequent continuous processing. During this stage, auxiliary information closely related to spatial position and motion state during the detection process is synchronously embedded into the corresponding data frame. For vehicle-mounted continuous detection scenarios, radar echo acquisition, vehicle movement, position changes, and time progression occur synchronously. Therefore, if position matching of the detection results is performed after defect identification, inaccurate result mapping can easily occur due to timing misalignment, speed fluctuations, or position sampling errors. Based on this, the present invention directly embeds BeiDou positioning information, equipment detection speed information, and acquisition time information into the spatiotemporal digital signal frame, ensuring that subsequent signal processing, spatiotemporal stabilization, image recognition, geometric parameter mapping, and report generation processes are always based on the same set of spatiotemporal data benchmarks.

[0032] Specifically, in one executable implementation, the synthesized echo data output in step S1 can first undergo analog-to-digital conversion processing according to the sampling time to convert the analog signal into single-frame digital signal data. Subsequently, multiple consecutive single-frame digital signal data can be accumulated according to the sampling order to form a spatiotemporal digital signal frame corresponding to the detection process. This spatiotemporal frame not only reflects the echo intensity information at a single moment but also reflects the spatial sampling process formed by the vehicle's movement along the detection route through the sequence relationship between consecutive frames. Further, the BeiDou positioning information, equipment detection speed information, and acquisition time information corresponding to each spatiotemporal digital signal frame can be read synchronously, and this auxiliary information can be written into the corresponding spatiotemporal digital signal frame to generate a labeled signal frame sequence. After the labeled signal frame sequence is generated, it can be sorted and stored according to the frame number to serve as a unified input data source for subsequent clutter suppression, noise suppression, multi-scale spatiotemporal collaborative stabilization, and result binding processes.

[0033] It should be noted that the equipment detection speed information in this invention is not only used to record the detection process, but will also be used to calculate the longitudinal displacement between adjacent image frames; similarly, after the Beidou positioning information is embedded in this step, it is used to directly achieve spatial positioning binding when generating subsequent disease identification results.

[0034] S3: Perform clutter suppression, noise suppression and data reconstruction processing on the labeled signal frame sequence in sequence, and convert the reconstructed data into a B-Scan image sequence.

[0035] Specifically, the data corresponding to each point in the B-Scan image is extracted and subjected to mean filtering to obtain clutter-suppressed data; the clutter-suppressed data is then standardized, the standardized correlation coefficient matrix is ​​calculated, and the characteristic equation of the correlation coefficient matrix is ​​solved to obtain n eigenvalues ​​and corresponding eigenvectors; the number of principal components is determined based on the cumulative contribution rate, and the standardized matrix is ​​reconstructed according to the principal components to generate a denoised B-Scan image sequence.

[0036] In this embodiment of the invention, the main function of step S3 is to perform layered suppression of background clutter and random noise in the original labeled signal frame sequence, and on this basis, reconstruct a B-Scan image sequence suitable for image representation. Because ground-penetrating radar electromagnetic waves are affected by factors such as interlayer medium differences, local scatterers, equipment vibration, and complex environmental noise during underground propagation, the received echo signal contains not only the target echoes truly related to the disease, but also background clutter and random noise of a certain intensity. If these interference components are not processed, the resulting B-Scan image will exhibit phenomena such as enhanced background band interference, prominent local artifacts, and blurred disease boundaries, especially for weak-response targets such as microcracks and shallow cavities, which are more easily masked by background components. Therefore, this invention adopts a processing approach of first removing clutter, then suppressing noise, and finally reconstructing the data to improve the clarity of the effective features of the disease in the B-Scan image.

[0037] Specifically, in one executable implementation, the labeled signal frame sequence can first be mapped to an initial B-Scan data matrix, and the data corresponding to the (i,j)th point in the B-Scan image can be extracted. Where i∈[1,m], j∈[1,n], m is the total number of rows in the B-Scan, and n is the total number of columns in the B-Scan. Then, mean filtering is performed on the data according to the following formula: ; in, The data is after mean filtering, and q is the column index for summation. Through this processing, the average background component in the same row of data can be subtracted from each column, thus achieving an initial reduction in the impact of background clutter. It should be noted that mean filtering primarily targets relatively stable background components extending along a certain direction in the B-Scan image, rather than directly altering the local abrupt changes in the disease target itself. Therefore, it can reduce background interference while preserving the target's abnormal information.

[0038] After mean filtering, this invention further performs principal component analysis on the processed data to suppress random noise and retain the main effective features. Specifically, the clutter-suppressed data can first be standardized according to the following formula: ; in, For standardized data, Let be the mean of the sample in column j. Let be the standard deviation of the sample in column j. After standardization, calculate the standardized correlation coefficient matrix using the following formula: ; Where R is the correlation coefficient matrix. Let Z be the transpose of the normalized matrix. Then, solve for the characteristic equation of the correlation coefficient matrix R: ; This yields n eigenvalues ​​and their corresponding eigenvectors. After orthogonalization, the projection basis required for subsequent principal component extraction is formed. To determine how many principal components should be retained to balance noise suppression and effective feature retention, this invention further uses a cumulative contribution rate for control, expressed as follows: ; Where B represents the cumulative contribution rate. Number of main components For the first There are several characteristic values. When the cumulative contribution rate reaches a preset range, it can be considered that the top [values] are [values]. The principal components are sufficient to characterize the effective information of the target. Furthermore, the preceding parameters can be obtained according to the following formula. Principal components: ; in, For the first Principal components, For the first Unit orthogonalized eigenvectors. Then, based on the previous... By reconstructing the original normalized matrix using principal components, a denoised B-Scan image sequence can be generated. Through the combined processing of mean filtering and principal component analysis, not only can background clutter and random noise be suppressed in layers, but the low-order effective features of the diseased targets can also be more clearly preserved in the reconstructed B-Scan image, laying a reliable foundation for subsequent image enhancement and automatic recognition.

[0039] S4: Perform image enhancement, invalid frame removal, and multi-scale spatiotemporal co-stabilization processing on the B-Scan image sequence to generate an image sequence to be identified.

[0040] Specifically, the CLAHE algorithm is used to perform adaptive contrast enhancement processing on the B-Scan image sequence; Gaussian blur processing is performed on the image after adaptive contrast enhancement processing, followed by Laplacian sharpening processing; the image after feature enhancement processing is uniformly converted into a standard grayscale image and uniformly scaled to the preset model input size, while pixel values ​​are standardized; the image after standardization is frame by frame to perform validity screening, removing all-black frames, all-white frames, signal interruption frames, and blurred frames with a sharpness less than a preset value, thus obtaining the preprocessed image sequence.

[0041] Based on this, for adjacent image frames continuously acquired along the detection direction, the longitudinal displacement between the current frame and each adjacent frame is calculated based on the device detection speed information and sampling time interval. Position alignment processing is performed on candidate anomaly regions in each adjacent frame based on the longitudinal displacement, and the aligned local response intensity, edge continuity, and texture aggregation are extracted under different scale windows to obtain multi-scale alignment response data. A collaborative stabilization score is calculated based on the candidate anomaly region score of the current frame and the multi-scale alignment response data. Scale consistency constraints are applied to the collaborative stabilization score to obtain stable candidate region weights. When the stable candidate region weights reach a preset threshold, the corresponding candidate anomaly region is retained, and its boundary range is smoothly updated to generate an image sequence to be identified.

[0042] Furthermore, when the weight of the stable candidate region reaches a preset threshold, the corresponding candidate abnormal region is retained, and its boundary range is smoothly updated to generate an image sequence to be identified. Specifically, this includes: extracting the initial boundary parameters of the candidate abnormal region in the current frame, reading the boundary parameters of the corresponding regions of the adjacent frames before and after which the region is aligned with its position, and calculating the boundary update result; and writing the updated boundary parameters back to the candidate abnormal region in the current frame to obtain a stable image sequence to be identified.

[0043] In this embodiment of the invention, step S4 is located between radar signal processing and automatic disease identification. Its purpose is to further improve the distinguishability of disease regions in B-Scan images and enhance the expression stability of the same disease target in consecutive image frames. Although the B-Scan image sequence processed in step S3 has completed background clutter suppression and random noise suppression, in actual complex scenarios, the images may still have problems such as insufficient local contrast, unclear texture of small targets, unclear boundary transitions, and unstable abnormal responses between consecutive frames. If such image sequences are directly input into the recognition model, on the one hand, it is easy to make the difference between the disease boundary and the background region insufficient, and on the other hand, it may also cause the same disease in consecutive frames to show inconsistent abnormal expressions in different frames, thereby affecting the output stability of the recognition model. Therefore, this invention further introduces multi-scale spatiotemporal collaborative stabilization processing to enhance the response continuity and boundary consistency of disease targets in consecutive image frames.

[0044] Specifically, in one executable implementation, the CLAHE algorithm is first used to perform adaptive contrast enhancement processing on the B-Scan image sequence. To achieve a more balanced local grayscale distribution and avoid over-enhancement, the image is divided into 8×8 rectangular sub-blocks, and the contrast limit threshold is set to 2.0. This processing enhances grayscale differences in local areas, providing a certain degree of contrast improvement to lesion areas that are originally close to the background grayscale. Afterward, Gaussian blurring is performed on the image after adaptive contrast enhancement to smooth high-frequency isolated noise points, followed by Laplacian sharpening. The Laplacian sharpening kernel used is: ; By performing convolution operations on the image using this sharpening kernel, the grayscale difference between the central pixel and its neighboring pixels can be enhanced, thereby improving the recognizability of cracks, hole edges, and local abnormal textures. After the enhancement process, the image is uniformly converted to a standard grayscale image and uniformly scaled to a model input size of 640×640. Simultaneously, pixel values ​​are standardized to ensure that subsequent recognition models have consistent requirements for the size and numerical distribution of the input image. Furthermore, this invention performs validity screening frame-by-frame on the standardized image, removing all-black frames, all-white frames, signal interruption frames, and blurry frames with a resolution less than 50, thereby preventing invalid frames from participating in subsequent spatiotemporal collaborative processing and automatic recognition.

[0045] After completing the image enhancement and invalid frame removal processes described above, this invention further performs multi-scale spatiotemporal collaborative stabilization processing. It should be noted that multi-scale spatiotemporal collaborative stabilization refers to, within a continuous image frame sequence, taking the current frame as the center, comprehensively utilizing response information related to the current candidate anomaly region in the temporal dimension, and jointly evaluating this response information through different scale windows in the spatial dimension, thereby maintaining higher consistency in the expression of the same defect target across consecutive frames. Specifically, this can be achieved first by using the device detection speed information corresponding to the current frame. and the sampling time interval between adjacent frames Calculate the vertical displacement between the current frame and the kth adjacent frame: ; in, Here, k represents the vertical displacement between the current frame and adjacent frames, where k is the adjacent frame number. After obtaining the vertical displacement, position alignment processing can be performed on candidate anomaly regions in adjacent frames based on this displacement, enabling these candidate regions in adjacent frames to correspond with the candidate regions in the current frame under a unified reference system. Since the degree of texture aggregation and edge continuity exhibited by disease targets varies at different scales, this invention is not limited to alignment responses at a single scale. Instead, it further extracts the aligned local response intensity, edge continuity, and texture aggregation at different scale windows to form multi-scale alignment response data. In this way, it can identify anomalies such as small cracks, which are more sensitive at small scale windows, while also taking into account disease types such as softened areas and cavities, which are more likely to form stable responses at larger scale windows.

[0046] After obtaining the multi-scale alignment response data, this invention further calculates the collaborative stabilization score based on the candidate anomaly region score of the current frame and the region scores after alignment with the preceding and following adjacent frames. The expression is as follows: ; in, To coordinate and stabilize the scoring, Score the candidate anomaly regions in the current frame. and , where represents the region score after positional alignment in the preceding and following frames, is the weight of the current frame, and K is the number of adjacent frames participating in the collaborative stabilization. Through the above scoring calculation, the abnormal response in the current frame is no longer determined solely by a single frame, but rather by the combined support of the preceding and following frames. In this way, even if the defective response in a certain frame is weakened due to short-term vibration, attitude disturbance, or local energy attenuation, as long as it maintains a relatively consistent response distribution in the preceding and following frames, it can be reflected in the collaborative stabilization score.

[0047] To further improve the consistency of candidate anomaly region judgment at different scales, this invention also applies a scale consistency constraint to the collaborative stability score and calculates the weight of stable candidate regions according to the following formula: ; in, To stabilize the weights of candidate regions, The weights for score fusion are S, where S is the number of scale windows. This is the alignment response evaluation value of the current frame at the s-th scale window. Only when the weight of a stable candidate region reaches a preset threshold is the corresponding candidate anomalous region retained for subsequent identification. This process avoids misjudging occasional noise anomalies at a certain scale as real defects, and also prevents real defects from being incorrectly rejected due to weak responses at individual scales. Furthermore, after retaining candidate anomalous regions, this invention performs a smooth update process on their boundary range to reduce the jitter of the same defect boundary in consecutive frames, as expressed below: ; in, For the updated boundary parameters, These are the initial boundary parameters for the candidate anomaly regions in the current frame. and These are the boundary parameters after positional alignment of adjacent frames, and γ is the weight of the boundary parameters in the current frame. Through this boundary smoothing update process, the boundary range of the same lesion can maintain higher consistency in consecutive image frames, so that the subsequent recognition model faces a more stable spatiotemporal representation and smoother boundary image sequence when reading the target area.

[0048] S5: Input the image sequence to be identified into the pre-trained YOLOv10s disease identification model to generate disease identification results, and perform physical coordinate mapping, disease level classification, result binding and detection report output based on the Beidou positioning information, equipment detection speed information and disease identification results.

[0049] Specifically, a training set, validation set, and test set are constructed using historical B-Scan disease image samples. Data augmentation processing is performed through random rotation, brightness adjustment, Gaussian blur, mosaic enhancement, and small object cropping to obtain a pre-trained YOLOv10s disease recognition model. The image sequence to be recognized is input into the pre-trained YOLOv10s disease recognition model, which outputs the disease category, bounding box coordinates, and confidence score to generate disease recognition results. Based on the pixel physical scale mapping relationship, the bounding box pixel coordinates are converted into actual physical coordinates, and the actual width, height, and depth parameters of the disease are calculated. Based on the disease size, burial depth, and confidence score, the disease is classified into levels to generate structured disease result data.

[0050] Furthermore, the structured disease result data is associated one-to-one with the corresponding B-Scan image identifier, BeiDou positioning coordinates, equipment detection speed information, and acquisition time information to form a disease location data structure. Based on the disease location data structure, a denoised B-Scan image and disease identification results are output, and the denoised B-Scan image is compared with the original B-Scan image to generate verification marker data. When a correction result based on the verification marker data is received, the structured disease result data is updated to obtain verified disease result data. A standardized roadbed disease detection report is generated based on the verified disease result data.

[0051] In this embodiment of the invention, step S5 is used to complete the conversion of the disease target from image-level representation to engineering result representation, that is, to automatically extract the disease category, location range, and confidence information from the image sequence to be identified, and further obtain the actual geometric parameters, spatial location, and standardized output results of the disease. In order to make the recognition model more adaptable to the characteristics of roadbed disease B-Scan images, such as many small targets, complex background, limited gray-level differences, and weak texture continuity, this invention preferably uses the YOLOv10s model as the disease recognition model, and improves the model's adaptability to various diseases such as cracks, softening, settlement, and cavities through scene-based datasets and targeted enhancement methods during the model training stage.

[0052] Specifically, in one feasible implementation, B-Scan image samples of roadbed defects can be pre-collected as training data and divided into training, validation, and test sets according to a preset ratio. To improve the model's generalization ability in complex scenarios, data augmentation processing such as random rotation, brightness adjustment, Gaussian blur, mosaic enhancement, and small object cropping can be performed on the samples, and the defect categories and bounding box positions can be labeled using the YOLO standard annotation format. After the training data preparation is completed, the augmented samples can be input into the YOLOv10s model training environment for training. During the training process, the anchor boxes, loss function, and training termination strategy are optimized in combination with the characteristics of the defect scene to obtain a pre-trained model suitable for B-Scan image recognition of roadbed defects. Then, the image sequence to be recognized output in step S4 can be input into the pre-trained YOLOv10s defect recognition model, and the defect category, bounding box coordinates, and confidence score are output, thereby generating defect recognition results in txt format.

[0053] After the disease identification results are generated, this invention further converts the bounding box pixel coordinates into actual physical coordinates based on the pixel-to-physical scale mapping relationship, and calculates the actual width, height, and depth of the disease accordingly. It should be noted that the bounding box positions in the B-Scan image are essentially pixel representations in the image coordinate system, and there is no natural one-to-one correspondence between them and the actual roadbed spatial positions. Therefore, the conversion must be completed using a pre-established pixel-to-physical scale mapping relationship. Through this processing, the identification results no longer stop at the level of abnormal areas in the image, but can further obtain dimensional parameters that can be directly used for disease grading and maintenance decisions in engineering. Subsequently, the disease size, burial depth, and identification confidence level can be comprehensively considered to classify the disease into grades, forming structured disease result data.

[0054] Furthermore, since the BeiDou positioning information, equipment detection speed information, and acquisition time information have already been bound to the spatiotemporal digital signal frame in step S2, the structured disease result data can be directly associated with the corresponding B-Scan image identifier, BeiDou positioning coordinates, equipment detection speed information, and acquisition time information in this step to form a disease location data structure. Based on this disease location data structure, a denoised B-Scan image and disease identification results in txt format can be output. By comparing the denoised B-Scan image with the original B-Scan image, verification marker data is generated.

[0055] Upon receiving the correction results based on the verification marker data, the structured disease result data can be updated to obtain verified disease result data. Finally, a standardized roadbed disease detection report can be generated based on the verified disease result data. This report may include disease type, disease level, actual width, height, depth, geographic coordinates, image index, and verification status, etc., so that it can be directly integrated into the highway maintenance management system for standardized entry, archiving, and scheduling of disease results.

[0056] In one specific implementation, such as Figure 2As shown, a portable active phased array radar is installed on a highway inspection vehicle to perform continuous inspection of a road section to be inspected. After the inspection begins, the radar control subsystem uniformly schedules the antenna array subsystem and the T / R component subsystem, so that multiple array elements transmit radio frequency detection signals of different frequencies and synchronously receive corresponding echoes during the same inspection period. The system first reads the original echo signals of each channel and performs multi-channel beamforming processing based on the array element position relationship and beam pointing angle to generate synthetic echo data. Subsequently, the synthetic echo data undergoes analog-to-digital conversion and spatiotemporal frame accumulation processing, and BeiDou positioning coordinates, equipment detection speed information, and acquisition time information are simultaneously written in to obtain a labeled signal frame sequence. Afterward, the system performs mean filtering and principal component analysis on the labeled signal frame sequence to remove background clutter and random noise, and reconstructs and generates a denoised B-Scan image sequence. Next, the system performs CLAHE algorithm enhancement, Gaussian blurring, Laplacian sharpening, format unification, and invalid frame removal on the B-Scan image sequence. When the detection vehicle passes through undulating road sections or sections with significant short-term vibration, the system calculates the longitudinal displacement between adjacent image frames based on the equipment's detection speed and sampling time interval. Candidate abnormal regions in adjacent frames are then aligned, multi-scale response extracted, collaboratively scored and fused, and their boundaries smoothed and updated to obtain a more stable and continuous image sequence for identification. Finally, the image sequence is input into a pre-trained YOLOv10s disease identification model, which outputs the disease category, bounding box coordinates, and confidence score. The actual geometric parameters of the disease are obtained by combining the pixel-physical scale mapping relationship. Spatial and temporal binding of the disease results is achieved using BeiDou positioning coordinates and acquisition time information, ultimately generating a standardized roadbed disease detection report. This processing flow maintains high consistency of disease targets in consecutive image frames under complex dynamic conditions, thereby improving the stability of minor disease identification, the continuity of boundary extraction, and the reliability of geometric parameter calculation.

[0057] In one specific implementation, a roadbed defect detection architecture based on active phased array radar is proposed, such as... Figure 3 As shown, the architecture includes a portable active phased array radar system (its module is positioned as a signal data transmission / reception module, and its module function is used for multi-frequency electromagnetic wave transmission, electromagnetic wave signal acquisition and multi-channel beamforming), a positioning system based on the Beidou satellite navigation system (its module is positioned as a positioning information module, and its module function is used for positioning information acquisition), a radar signal processing system (its module is positioned as a processing module, and its module function is used for signal noise reduction, detection image display, defect identification and result storage and output), and a computer system (its module is positioned as a system carrier module, and its module function is used as a system execution platform).

[0058] Furthermore, in practical applications, a portable active phased array radar system mainly includes an antenna array subsystem, a T / R module subsystem, a signal processing subsystem, a radar control subsystem, and a power supply and heat dissipation subsystem. The antenna array subsystem, as the foundation of the overall hardware integration, is used to transmit and receive electromagnetic wave signals. The T / R module subsystem consists of miniaturized RF transceivers, used to support independent transmission and reception of RF signals at different frequencies by each radiating element. The signal processing subsystem performs preliminary processing on the received raw echo signals and forms standardized input data. The radar control subsystem converts detection tasks and control commands into unified scheduling of each subsystem. The power supply and heat dissipation subsystem ensures stable operation of the equipment under continuous detection conditions. The positioning system based on the BeiDou satellite navigation system mainly consists of an embedded BeiDou receiver, a positioning antenna, an auxiliary positioning device, and a data storage unit, used to synchronously acquire positioning information during radar detection and feed the positioning information back to the subsequent processing module. The radar signal processing system receives the raw echo signal output from the antenna array and performs processing such as multi-channel beamforming, digital signal conversion, clutter suppression, noise suppression, and B-Scan image construction. The computer system, as the control core and processing platform of the entire detection system, uniformly controls the various subsystems and further performs image enhancement, multi-scale spatiotemporal collaborative stabilization, YOLOv10s defect identification, physical coordinate mapping, defect level classification, result binding, verification and updating, and detection report output. Through this system architecture, fully automated processing from raw echo acquisition to standardized defect report output can be achieved.

[0059] Reference Figure 4 , Figure 4 This is a schematic diagram of the roadbed defect detection system based on active phased array radar according to an embodiment of the present invention.

[0060] like Figure 4 As shown, the roadbed defect detection system based on active phased array radar proposed in this embodiment of the invention includes: The acquisition module 10 is used to acquire the multi-frequency raw echo signal corresponding to the target road segment, and perform multi-channel beamforming processing on the multi-frequency raw echo signal to generate synthesized echo data. The generation module 20 is used to perform analog-to-digital conversion, spatiotemporal frame accumulation and auxiliary information embedding processing on the synthesized echo data to generate a labeled signal frame sequence; wherein, the auxiliary information includes at least BeiDou positioning information and device detection speed information; The reconstruction module 30 is used to sequentially perform clutter suppression, noise suppression and data reconstruction processing on the labeled signal frame sequence, and convert the reconstructed data into a B-Scan image sequence; Processing module 40 is used to perform image enhancement, invalid frame removal and multi-scale spatiotemporal co-stabilization processing on the B-Scan image sequence to generate an image sequence to be identified; The output module 50 is used to input the image sequence to be identified into the pre-trained YOLOv10s disease identification model, generate disease identification results, and perform physical coordinate mapping, disease level classification, result binding and detection report output based on Beidou positioning information, equipment detection speed information and disease identification results.

[0061] Other embodiments or specific implementations of the roadbed defect detection system based on active phased array radar of the present invention can be referred to the above-described method embodiments, and will not be repeated here.

[0062] It is understood that in the description of this specification, references to terms such as "one embodiment," "another embodiment," "other embodiments," or "first embodiment to Nth embodiment," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0063] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0064] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for detecting roadbed defects based on active phased array radar, characterized in that, The method includes the following steps: Acquire the multi-frequency raw echo signal corresponding to the target road segment, and perform multi-channel beamforming processing on the multi-frequency raw echo signal to generate synthesized echo data; The synthesized echo data is subjected to analog-to-digital conversion, spatiotemporal frame accumulation, and auxiliary information embedding processing to generate a labeled signal frame sequence; wherein, the auxiliary information includes at least BeiDou positioning information and device detection speed information; The labeled signal frame sequence is subjected to clutter suppression, noise suppression and data reconstruction processing in sequence, and the reconstructed data is converted into a B-Scan image sequence; The B-Scan image sequence is subjected to image enhancement, invalid frame removal and multi-scale spatiotemporal co-stabilization processing to generate an image sequence to be identified; The image sequence to be identified is input into the pre-trained YOLOv10s disease identification model to generate disease identification results. Based on the Beidou positioning information, equipment detection speed information and disease identification results, physical coordinate mapping, disease level classification, result binding and detection report output are performed.

2. The method for detecting roadbed defects based on active phased array radar as described in claim 1, characterized in that, Acquire the multi-frequency raw echo signal corresponding to the target road segment, and perform multi-channel beamforming processing on the multi-frequency raw echo signal to generate synthesized echo data, specifically including: Acquire the original echo signals corresponding to each channel of the active phased array radar, and establish the correspondence between the original echo signals and the array element positions according to the channel number; Time delay compensation is performed on the original echo signals of each channel based on the beam pointing angle, and multi-channel beamforming processing is performed to output the final synthesized output signal as synthesized echo data.

3. The method for detecting roadbed defects based on active phased array radar as described in claim 1, characterized in that, The synthesized echo data undergoes analog-to-digital conversion, spatiotemporal frame accumulation, and auxiliary information embedding processing to generate a labeled signal frame sequence, specifically including: The synthesized echo data is subjected to analog-to-digital conversion processing according to the sampling time to obtain single-frame digital signal data; Multiple consecutive single-frame digital signal data are accumulated in the sampling order to form a spatiotemporal digital signal frame corresponding to the detection process; Read the BeiDou positioning information, device detection speed information, and acquisition time information corresponding to each of the aforementioned spatiotemporal digital signal frames; The BeiDou positioning information, equipment detection speed information, and acquisition time information are written into the corresponding spatiotemporal digital signal frames to obtain a labeled signal frame sequence. The labeled signal frame sequence is sorted and stored according to the frame number.

4. The method for detecting roadbed defects based on active phased array radar as described in claim 1, characterized in that, The labeled signal frame sequence is sequentially subjected to clutter suppression, noise suppression, and data reconstruction processing, and the reconstructed data is converted into a B-Scan image sequence, specifically including: The data corresponding to each point in the B-Scan image is extracted and subjected to mean filtering to obtain clutter-suppressed data. After suppressing clutter, the data is standardized, the standardized correlation coefficient matrix is ​​calculated, and the characteristic equation of the correlation coefficient matrix is ​​solved to obtain n eigenvalues ​​and corresponding eigenvectors. The number of principal components is determined based on the cumulative contribution rate, and the normalized matrix is ​​reconstructed based on the principal components to generate a denoised B-Scan image sequence.

5. The method for detecting roadbed defects based on active phased array radar as described in claim 1, characterized in that, The B-Scan image sequence is subjected to image enhancement and invalid frame removal processing, specifically including: The CLAHE algorithm is used to perform adaptive contrast enhancement processing on the B-Scan image sequence; After adaptive contrast enhancement, Gaussian blur is applied to the image, followed by Laplacian sharpening. After feature enhancement processing, the images are uniformly converted into standard grayscale images and uniformly scaled to the preset model input size, while the pixel values ​​are standardized. The images in the unified format are filtered frame by frame to remove all black frames, all white frames, frames with signal interruption, and blurry frames with a resolution less than a preset value, resulting in a preprocessed image sequence.

6. The method for detecting roadbed defects based on active phased array radar as described in claim 5, characterized in that, Multi-scale spatiotemporal co-stabilization processing is performed on the B-Scan image sequence to generate an image sequence to be identified, specifically including: For adjacent image frames acquired continuously along the detection direction, the longitudinal displacement between the current frame and each adjacent frame is calculated based on the device detection speed information and the sampling time interval. Based on the longitudinal displacement, position alignment processing is performed on the candidate abnormal regions in each adjacent frame, and the local response intensity, edge continuity and texture aggregation degree after alignment are extracted under different scale windows to obtain multi-scale alignment response data. Calculate the collaborative stability score based on the current frame candidate anomaly region score and multi-scale alignment response data; The scale consistency constraint is applied to the collaborative stability score to obtain the weights of stable candidate regions; When the weight of the stable candidate region reaches a preset threshold, the corresponding candidate abnormal region is retained, and its boundary range is smoothly updated to generate an image sequence to be identified.

7. The method for detecting roadbed defects based on active phased array radar as described in claim 6, characterized in that, When the weight of the stable candidate region reaches a preset threshold, the corresponding candidate abnormal region is retained, and its boundary range is smoothly updated to generate a sequence of images to be identified, specifically including: Extract the initial boundary parameters of the candidate anomaly region in the current frame, read the boundary parameters of the corresponding regions in the adjacent frames before and after it that are aligned with its position, and calculate the boundary update result; The updated boundary parameters are written back to the candidate anomaly region of the current frame to obtain a boundary-stable image sequence to be identified.

8. The method for detecting roadbed defects based on active phased array radar as described in claim 1, characterized in that, The image sequence to be identified is input into a pre-trained YOLOv10s disease identification model to generate disease identification results, specifically including: Training, validation, and test sets were constructed using historical B-Scan disease image samples. Data augmentation processing was performed through random rotation, brightness adjustment, Gaussian blur, mosaic enhancement, and small object cropping to obtain a pre-trained YOLOv10s disease recognition model. The image sequence to be identified is input into the pre-trained YOLOv10s disease identification model, which outputs the disease category, bounding box coordinates, and confidence score to generate disease identification results. Based on the pixel physical scale mapping relationship, the bounding box pixel coordinates are converted into actual physical coordinates, and the actual width, height and depth parameters of the lesion are calculated; Diseases are classified into different levels based on their size, burial depth, and confidence level, generating structured disease result data.

9. The method for detecting roadbed defects based on active phased array radar as described in claim 8, characterized in that, Based on the BeiDou positioning information, equipment detection speed information, and defect identification results, physical coordinate mapping, defect level classification, result binding, and detection report output are performed, specifically including: The structured disease result data is associated one by one with the corresponding B-Scan image identifier, Beidou positioning coordinates, equipment detection speed information and acquisition time information to form a disease positioning data structure; Based on the disease location data structure, output the denoised B-Scan image and disease identification results, and compare the denoised B-Scan image with the original B-Scan image to generate verification marker data; Upon receiving the correction result generated based on the verification marker data, the structured disease result data is updated to obtain the verified disease result data; A standardized roadbed defect detection report is generated based on the reviewed defect data.

10. A roadbed defect detection system based on active phased array radar, characterized in that, The system includes: The acquisition module is used to acquire the multi-frequency raw echo signal corresponding to the target road segment, and perform multi-channel beamforming processing on the multi-frequency raw echo signal to generate synthesized echo data. The generation module is used to perform analog-to-digital conversion, spatiotemporal frame accumulation, and auxiliary information embedding processing on the synthesized echo data to generate a labeled signal frame sequence; wherein, the auxiliary information includes at least BeiDou positioning information and device detection speed information; The reconstruction module is used to sequentially perform clutter suppression, noise suppression and data reconstruction processing on the labeled signal frame sequence, and convert the reconstructed data into a B-Scan image sequence; The processing module is used to perform image enhancement, invalid frame removal and multi-scale spatiotemporal co-stabilization processing on the B-Scan image sequence to generate an image sequence to be identified; The output module is used to input the image sequence to be identified into the pre-trained YOLOv10s disease identification model, generate disease identification results, and perform physical coordinate mapping, disease level classification, result binding and detection report output based on the Beidou positioning information, equipment detection speed information and disease identification results.

Images