Intelligent diagnosis method for multi-dimensional health state of roadbed

Abstract

The invention discloses an intelligent diagnosis method for the multi-dimensional health state of a roadbed, and belongs to the technical field of structure health monitoring, and the method comprises the steps: obtaining an internal structure radar image of the roadbed, multi-measurement-point time sequence strain data, and the real-time position information of a moving load; based on an internal structure radar image, a structure reflection abnormal area and a structure reflection normal area are identified, in combination with real-time position information of a moving load, a load strain dynamic response waveform and a reference dynamic response waveform are extracted from multi-measuring-point time sequence strain data for quantitative comparison, and a response differentiation feature vector is generated; and carrying out roadbed disease diagnosis by using the response differentiation feature vector to obtain a roadbed health state diagnosis report containing a disease type and a risk level. By fusing the internal structure radar image and the multi-measuring-point time sequence strain data and performing differential comparison and model analysis, the roadbed disease type can be diagnosed, the risk level can be evaluated, and the roadbed health state can be efficiently and reliably evaluated.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, and in particular to an intelligent diagnostic method for the multi-dimensional health status of roadbeds. Background Technology

[0002] As the fundamental load-bearing structure of highways and railways, the long-term service performance and stability of the roadbed directly affect the safety, comfort, and durability of transportation infrastructure. Under the long-term combined effects of traffic loads and the natural environment, the roadbed is prone to various defects such as voids, loosening, and uneven settlement. Therefore, timely and accurate detection and assessment of the roadbed's health status is a crucial step in ensuring traffic safety and extending the service life of roads.

[0003] In existing technologies, the health status detection of roadbeds mainly relies on non-destructive testing (NDT) and sensor monitoring technologies. For example, ground-penetrating radar (GPR), a commonly used NDT method, detects the internal structure of the roadbed by emitting electromagnetic waves and receiving their reflected signals in the underground medium. It can be used to detect structural anomalies such as cavities and water-rich areas. Simultaneously, embedding sensors such as strain gauges in the roadbed allows for direct measurement of the dynamic mechanical response of the roadbed under moving loads, reflecting its stress performance.

[0004] However, the interpretation of ground-penetrating radar (GPR) images is complex and heavily reliant on the professional experience of the inspectors. It struggles to identify early, weak signs of structural damage and only provides static structural information. Furthermore, using strain sensors alone results in dynamic response data that is severely affected by various non-structural factors such as the speed and weight of moving loads and environmental temperature and humidity. This makes it difficult to accurately isolate response changes caused by structural damage, leading to unreliable diagnostic results. In addition, current technologies generally lack a systematic method for effectively integrating static structural information with dynamic response information, resulting in often simplistic diagnostic conclusions that fail to provide detailed assessments of damage type and risk level. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an intelligent diagnostic method for the multi-dimensional health status of roadbeds. By fusing internal structural radar images with time-series strain data from multiple measurement points, and performing differential comparison and model analysis, it can diagnose roadbed disease types and assess risk levels, achieving efficient and reliable assessment of roadbed health status.

[0006] The above objectives can be achieved through the following approach: A method for intelligent diagnosis of the multi-dimensional health status of a roadbed includes acquiring radar images of the roadbed's internal structure, time-series strain data from multiple measuring points, and real-time location information of moving loads; identifying abnormal and normal structural reflection zones based on the radar images of the internal structure; extracting load-strain dynamic response waveforms and benchmark dynamic response waveforms from the time-series strain data from the multiple measuring points, based on the real-time location information of the moving loads and corresponding to the abnormal and normal structural reflection zones; quantitatively comparing the load-strain dynamic response waveforms with the benchmark dynamic response waveforms to generate response-differential feature vectors; and using the response-differential feature vectors to diagnose roadbed defects, obtaining a roadbed health status diagnosis report containing defect types and risk levels.

[0007] Optionally, identifying structural reflection anomalous areas and structural reflection normal areas based on the internal structure radar image includes: performing image analysis on the internal structure radar image to generate an image feature map; using a self-attention mechanism to enhance the weights of regions related to structural defects in the image feature map, performing semantic segmentation, and outputting a region mask marked with location information; and determining structural reflection anomalous areas and structural reflection normal areas based on the region mask.

[0008] Optionally, the step of extracting the load-strain dynamic response waveform and the reference dynamic response waveform from the multi-point time-series strain data based on the real-time location information of the moving load, corresponding to the structural reflection anomaly zone and the structural reflection normal zone, includes: mapping the real-time location information of the moving load to a preset roadbed coordinate system to determine the start and end time windows of the moving load acting in the structural reflection anomaly zone and the structural reflection normal zone; intercepting the multi-point time-series strain data according to the start and end time windows to form an initial dynamic response waveform; performing filtering and noise reduction processing on the initial dynamic response waveform to generate a smooth response waveform, and extracting the load-strain dynamic response waveform and the reference dynamic response waveform based on the smooth response waveform.

[0009] Optionally, the quantitative comparison of the load-strain dynamic response waveform with the reference dynamic response waveform to generate a response differentiation feature vector includes: calculating the dynamic time warping distance between the load-strain dynamic response waveform and the reference dynamic response waveform; extracting morphological features from the two waveforms respectively, and calculating the relative difference of the morphological features; and combining the dynamic time warping distance with the relative difference to form a response differentiation feature vector.

[0010] Optionally, the method further includes: acquiring multi-point temperature and humidity sensing data of the structural reflection anomaly zone and extracting temperature and humidity parameters from it; and using the temperature and humidity parameters to correct the response differential feature vector.

[0011] Optionally, the step of using the response differential feature vector to diagnose subgrade diseases and obtain a subgrade health status diagnosis report containing disease type and risk level includes: constructing a subgrade health status diagnosis model; inputting the modified response differential feature vector into the subgrade health status diagnosis model, performing reasoning analysis, and generating a preliminary diagnosis result containing disease type and risk level information; and generating a subgrade health status diagnosis report based on the preliminary diagnosis result.

[0012] Optionally, the construction of the roadbed health status diagnostic model includes: obtaining response differential feature vectors with known disease type labels, and constructing a historical training set; using the historical training set to perform supervised training on the probabilistic classification model, so that the model learns the mapping relationship from feature vectors to disease types and corresponding confidence probabilities, thereby obtaining the roadbed health status diagnostic model; wherein, the roadbed health status diagnostic model includes multiple heterogeneous base classifiers.

[0013] Optionally, the step of inputting the response differential feature vector into the roadbed health status diagnosis model for inference analysis to generate a preliminary diagnosis result containing information on disease type and risk level includes: inputting the corrected response differential feature vector into the multiple heterogeneous base classifiers respectively to obtain multiple parallel diagnostic sub-results; and performing weighted fusion of the multiple parallel diagnostic sub-results according to a preset decision fusion rule to generate a preliminary diagnosis result.

[0014] Optionally, the method includes: extracting disease location and risk level information from the roadbed health status diagnosis report; rendering the disease location and risk level information onto a three-dimensional digital model of a roadbed; and generating a three-dimensional visualized risk map based on the rendering results.

[0015] Based on the same inventive concept, this invention also provides an intelligent diagnostic system for the multi-dimensional health status of a roadbed. The system includes: a data acquisition module for acquiring radar images of the roadbed's internal structure, time-series strain data from multiple measuring points, and real-time location information of moving loads; a region identification module for identifying abnormal and normal structural reflection areas based on the radar images of the internal structure; a dynamic response extraction module for extracting load-strain dynamic response waveforms and baseline dynamic response waveforms from the time-series strain data from multiple measuring points, based on the real-time location information of the moving loads and corresponding to the abnormal and normal structural reflection areas; a differential feature generation module for quantitatively comparing the load-strain dynamic response waveforms with the baseline dynamic response waveforms to generate response differential feature vectors; and an intelligent diagnostic module for using the response differential feature vectors to diagnose roadbed defects and obtain a roadbed health status diagnostic report containing defect types and risk levels.

[0016] Compared with the prior art, the present invention has the following advantages: 1. This invention achieves multi-dimensional analysis from static structural detection to dynamic mechanical behavior verification by fusing radar images of the internal structure of the roadbed with time-series strain data from multiple measurement points. It also uses the response of the normal reflection zone of the structure as a dynamic benchmark and performs differential comparison on the response of the abnormal zone, which can eliminate common interference caused by differences in moving loads and some environmental factors, thereby improving the signal-to-noise ratio of defect identification and the accuracy of diagnostic results. 2. This invention achieves full automation and intelligence of the entire process from automatic identification of abnormal areas in radar images and extraction of differential features of dynamic response waveforms to inference and diagnosis of disease type and risk level through deep learning image analysis, self-attention mechanism and probabilistic classification model. It gets rid of the dependence on human experience interpretation and ensures the objectivity, consistency and efficiency of the diagnosis process. 3. This invention not only determines whether there are defects in the roadbed, but also provides more refined diagnostic conclusions including specific defect types and risk levels. By introducing temperature and humidity parameters to correct the feature vector, the stability of the diagnostic results under different environmental conditions is further enhanced, providing richer and more reliable decision-making basis for the refined, preventive maintenance and full life-cycle management of the roadbed.

[0017] This invention renders abstract diagnostic report information onto a three-dimensional digital model of the roadbed, generating an intuitive and interactive three-dimensional visual risk map. This visualization method makes the spatial location, distribution pattern, and severity of defects immediately apparent, greatly enhancing the interpretability and operability of the diagnostic results. It facilitates management and technical personnel in quickly grasping the overall health status of the roadbed and formulating scientific maintenance and repair plans.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating an intelligent diagnostic method for the multi-dimensional health status of a roadbed according to an embodiment of the present invention.

[0021] Figure 2This is a schematic diagram of region recognition based on radar images according to an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of an intelligent diagnostic system for the multi-dimensional health status of a roadbed according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Reference Figure 1 One embodiment of the present invention proposes an intelligent diagnostic method for the multi-dimensional health status of roadbeds. By fusing internal structural radar images with time-series strain data from multiple measurement points and performing differential comparison and model analysis, the method can diagnose the types of roadbed defects and assess the risk level, thereby achieving efficient and reliable assessment of the health status of roadbeds.

[0025] The method described in this embodiment specifically includes: Acquire radar images of the roadbed's internal structure, time-series strain data from multiple measuring points, and real-time location information of moving loads; Based on the radar image of the internal structure, areas with abnormal structural reflection and areas with normal structural reflection are identified. Based on the real-time location information of the moving load, and corresponding to the abnormal reflection area and the normal reflection area of ​​the structure, the load strain dynamic response waveform and the reference dynamic response waveform are extracted from the multi-point time-series strain data, respectively. The load strain dynamic response waveform is quantitatively compared with the reference dynamic response waveform to generate a response difference feature vector; The roadbed disease diagnosis is performed using the aforementioned response differential feature vector, resulting in a roadbed health status diagnosis report that includes disease type and risk level.

[0026] This invention achieves precise location and qualitative and quantitative analysis of roadbed defects by integrating static structural information and dynamic mechanical response information. First, the roadbed is spatially divided into zones using internal structural radar images, establishing anomaly zones with suspected structural defects and normal reflection zones serving as a healthy reference. Then, by using real-time location information of the moving load, precise spatiotemporal synchronization of load action and multi-point time-series strain data is achieved, enabling the extraction of dynamic load-strain response waveforms and baseline dynamic response waveforms induced by the moving load acting in the two different zones. This method does not analyze the response of the anomaly zone in isolation; instead, it introduces the response of the normal zone as a dynamic benchmark. By quantitatively comparing these two waveforms, a response-differentiated feature vector is generated that filters out common interferences such as load variations and highlights the differences in mechanical behavior caused by structural defects. Finally, the feature vector is input into a pre-trained intelligent diagnostic model, which automatically infers and maps the features to defect types and risk levels.

[0027] This invention improves the accuracy, reliability, and intelligence of roadbed health status diagnosis. By combining radar imagery and dynamic strain response, it achieves a leap from static structural detection to dynamic behavior verification, making the diagnostic conclusions more comprehensive and reliable. The differential comparison method using normal zone response as a dynamic benchmark can resist interference from varying moving load speeds and weights, as well as some environmental factors, enhancing the stability and robustness of the diagnostic results. By generating quantified response difference feature vectors and analyzing them using an intelligent diagnostic model, this method can achieve refined identification of disease types and objective assessment of risk levels, surpassing the limitations of traditional methods that can only determine the presence or absence of abnormalities. Ultimately, the automatically generated health status diagnosis report provides a technical basis for preventive maintenance and scientific decision-making in roadbeds.

[0028] Optionally, identifying areas of abnormal structural reflection and areas of normal structural reflection based on the radar image of the internal structure includes: Image analysis is performed on the radar image of the internal structure to generate an image feature map; The weights of regions related to structural defects in the image feature map are enhanced by using a self-attention mechanism, and semantic segmentation is performed to output a region mask labeled with location information. Based on the region mask, the structural reflection anomalous region and the structural reflection normal region are determined.

[0029] Specifically, such as Figure 2First, radar images of the roadbed's internal structure are acquired and input. These images, presented as two-dimensional grayscale or pseudo-color images, visually reflect the intensity of electromagnetic wave reflection signals at different depths and locations of the roadbed. Subsequently, these internal structure radar images are input into a pre-constructed deep learning image analysis network model. The front end of this network model is a feature extraction backbone network, such as a convolutional neural network. It processes the input radar image layer by layer through multiple convolutional and pooling operations, extracting structural information from low-level textures and edges to high-level abstractions from the original pixels, ultimately generating one or more image feature maps. The image feature map is a compact representation of the original radar image across different dimensions, containing rich structural semantic information. To improve the accuracy of identifying structural defect regions, a self-attention mechanism is introduced for feature enhancement based on the image feature maps. The self-attention mechanism can calculate the interdependence between features at any two locations in the image feature map, thereby assigning a global context-aware weight to each feature location. For unique reflection patterns associated with structural defects, such as cavities, loose areas, or water-rich regions, this mechanism can automatically learn and assign them higher weights, while suppressing the feature weights of normal, homogeneous structural regions. This process can be understood as the model autonomously focusing its "attention" on the areas most likely to contain defects. After weighted enhancement by the self-attention mechanism, features related to structural defects are highlighted in the image feature map.

[0030] Next, the enhanced image feature map is fed into the semantic segmentation module of the network. The semantic segmentation module is responsible for classifying each pixel in the feature map. It outputs a region mask of the same size as the original radar image. This region mask is a label map, where each pixel is assigned a specific category label; for example, "1" represents a structural reflection anomaly area, and "0" represents a structural reflection normal area. Since the region mask corresponds one-to-one with the original radar image in spatial location, it not only labels the categories of different regions but also accurately marks the location and boundary contours of these regions. Finally, based on the pixel values ​​of the region mask, the system can directly determine the specific spatial range of structural reflection anomaly areas and structural reflection normal areas, providing precise location guidance for subsequent strain data analysis.

[0031] The above method enables the interpretation of radar images of internal structures. By leveraging the powerful feature extraction capabilities of deep learning models and combining them with the self-attention mechanism to enhance key defect features, the methods overcome the problems of strong subjectivity, susceptibility to noise interference, and insensitivity to subtle anomalies inherent in traditional manual interpretation or simple image processing algorithms. This improves the accuracy and robustness of identifying defects in the internal structure of the roadbed, enabling precise segmentation of potential disease areas from complex radar signal backgrounds. It provides a reliable spatial positioning foundation for subsequent multi-dimensional data fusion diagnosis, thereby enhancing the intelligence level of the entire diagnostic method and the reliability of the diagnostic results.

[0032] Optionally, the extraction of load strain dynamic response waveforms and reference dynamic response waveforms from the multi-point time-series strain data based on the real-time location information of the moving load, corresponding to the structural reflection anomaly zone and the structural reflection normal zone, includes: The real-time location information of the moving load is mapped to a preset roadbed coordinate system to determine the start and end time windows of the moving load's action in the abnormal reflection zone and the normal reflection zone of the structure. The time-series strain data of the multiple measurement points are extracted according to the start and end time windows to form the initial dynamic response waveform; The initial dynamic response waveform is filtered and denoised to generate a smooth response waveform, and the load strain dynamic response waveform and the reference dynamic response waveform are extracted based on the smooth response waveform.

[0033] Specifically, the first step is to establish a unified roadbed coordinate system, typically a three-dimensional Cartesian coordinate system, to accurately describe the spatial location of any point within the roadbed and the moving load on the ground. The locations and extents of the structural reflection anomaly zones and normal structural reflection zones identified in previous steps are also calibrated within this coordinate system. Simultaneously, a vehicle-mounted high-precision positioning system, such as Real-Time Kinematic GPS (RTK-GPS), continuously acquires the real-time location information of the moving load on the roadbed surface. This information includes timestamps and corresponding planar coordinates.

[0034] Next, spatiotemporal mapping is performed to determine the start and end time windows of the load application. The real-time position information sequence of the moving load over time is projected onto a pre-defined roadbed coordinate system to form its trajectory on the roadbed surface. By comparing this trajectory with the geometric boundaries of the pre-defined structural reflection anomaly zone and structural reflection normal zone, the moments when the moving load enters and leaves these two specific areas can be accurately determined. Thus, the start and end time windows of the moving load's action in the structural reflection anomaly zone and the structural reflection normal zone are obtained, respectively. This step ensures that the strain data in subsequent analyses correspond to events where the moving load passes through specific health condition areas.

[0035] Based on a defined start and end time window, relevant response segments are extracted from pre-collected multi-point time-series strain data. This multi-point time-series strain data is recorded in real-time by strain sensors, such as fiber optic strain gauges, deployed within the roadbed, particularly near or inside areas of structural reflection anomalies and normal zones. According to the time window, the system extracts signals within the corresponding time period from the continuous strain data stream, thus constructing an initial dynamic response waveform representing the load effect. For example, data corresponding to the time window in the structural reflection anomaly zone constitutes the initial waveform for that region, while data corresponding to the time window in the structural reflection normal zone constitutes the initial waveform for the normal zone.

[0036] Finally, the extracted initial dynamic response waveform undergoes signal purification processing. Since the original strain signal may contain high-frequency noise introduced by factors such as electromagnetic interference and environmental vibration, it needs to be filtered and denoised to generate a smoother response waveform that better reflects the true stress state of the structure. Commonly used filtering methods include Butterworth low-pass filtering, wavelet transform denoising, or moving average. After processing, the smooth response waveform originating from the structural reflection anomaly zone is formally defined as the load-strain dynamic response waveform, which characterizes the mechanical behavior of potentially damaged areas under load. The smooth response waveform originating from the structural reflection normal zone is defined as the baseline dynamic response waveform, serving as a reference standard for healthy subgrade structures under the same load.

[0037] This method enables the extraction of dynamic response signals directly related to the health status of specific subgrade areas from monitoring data. It effectively aligns and correlates information across four dimensions: load, location, time, and strain, thus capturing the effects of load. Using the response of the normal area as a dynamic benchmark provides a high-fidelity reference system for subsequent comparative analysis, eliminating interference from external variables such as load velocity and weight. This allows the analysis to focus more on the differences in response caused by variations in the internal structure of the subgrade, improving the accuracy and reliability of subsequent disease diagnosis.

[0038] Optionally, the quantitative comparison of the load-strain dynamic response waveform with the reference dynamic response waveform to generate a response difference feature vector includes: Calculate the dynamic time warp distance between the load strain dynamic response waveform and the reference dynamic response waveform; Morphological features are extracted from the two waveforms respectively, and the relative difference between the morphological features is calculated. The dynamic time-warped distance is combined with the relative difference to form a response-differentiated feature vector.

[0039] Specifically, the dynamic time warping algorithm is first used to calculate the overall morphological similarity between the load-strain dynamic response waveform and the reference dynamic response waveform. Dynamic time warping is an effective algorithm for measuring the similarity between two time series, especially suitable for situations where the two series have nonlinear distortions or phase differences on the time axis. This algorithm finds an optimal warping path that minimizes the cumulative distance between corresponding points of the two waveforms; this minimum cumulative distance is the dynamic time warping distance. The smaller this distance value, the more similar the overall shapes of the two waveforms are; conversely, the larger the value, the greater the difference. This scalar value, as the first dimension of the differential characteristics, reflects the fundamental difference in the stress response patterns between the damaged area and the normal area.

[0040] Secondly, to capture the response differences more precisely, it is necessary to extract a set of key morphological features from each of the two waveforms and calculate their relative changes. Morphological features are quantitative indicators describing the local or global geometric properties of a waveform. The morphological features extracted in this method may include, but are not limited to, peak values ​​(the maximum amplitude of the strain response); trough values ​​(the minimum amplitude or maximum reverse amplitude of the strain); response duration (the time elapsed from the start of the response to recovery to a steady state); and waveform energy, which is usually approximated by the absolute value or sum of squares of the area enclosed by the waveform curve and the time axis. For each pair of morphological features of the same name extracted from the load-strain dynamic response waveform and the reference dynamic response waveform, their relative difference is calculated using the following formula: , in, The relative difference representing a certain morphological feature. This characteristic value is extracted from the dynamic response waveform under load and strain. This is the corresponding characteristic value extracted from the reference dynamic response waveform. This relative difference is a dimensionless pure number that standardizes the differences in characteristics, making it unaffected by absolute factors such as load magnitude, and directly reflecting the relative degree of change in mechanical response characteristics caused by structural anomalies.

[0041] Finally, the dynamic time warp distance calculated in the preceding steps is combined with the relative differences of all morphological features to construct a multi-dimensional response differentiation feature vector. For example, if three morphological features—peak value, response duration, and waveform energy—are extracted, the generated response differentiation feature vector can be represented as [dynamic time warp distance, relative difference of peak value, relative difference of response duration, relative difference of waveform energy]. This vector comprehensively encapsulates the differences in strain response between the affected area and the normal area in multiple physical dimensions, such as overall shape, peak size, response speed, and energy dissipation.

[0042] The above method transforms the dynamic waveform comparison problem into a feature vector construction problem. This method not only overcomes the challenge of misaligned response timing under actual loads through dynamic time warping, ensuring robustness of the comparison, but also reveals the intrinsic mechanism of the impact of defects on the mechanical behavior of the roadbed through the relativistic comparison of multi-dimensional morphological features. The generated response-differentiated feature vectors have high information entropy and strong discriminative power, providing input for subsequent intelligent diagnostic models and improving the accuracy and refinement of defect type identification and risk level assessment.

[0043] Optionally, the method further includes: Acquire multi-point temperature and humidity sensing data of the structural reflection anomaly zone, and extract temperature and humidity parameters from them; The temperature and humidity parameters are used to correct the response differential feature vector.

[0044] Specifically, the first step is to collect real-time temperature and humidity data at different depths within the pre-determined structural reflection anomaly zone using multi-point temperature and humidity sensors, either pre-positioned or simultaneously deployed. These sensors can be thermistors, thermocouples, or humidity sensors. After acquiring the discrete temperature and humidity sensor data, effective temperature and humidity parameters representing the environmental conditions during the load application period need to be extracted. Since the time window for the moving load to act on a specific area is short, the acquisition of temperature and humidity data should be time-aligned with the extraction of the load strain dynamic response waveform. From the multi-point temperature and humidity sensor data, all data points within the start and end time window of the moving load acting on the structural reflection anomaly zone are extracted. Then, the temporal or spatial average values ​​of these data points are calculated to obtain representative temperature and humidity parameters.

[0045] The next core step is to use these two environmental parameters to correct the response differentiation feature vector generated in the previous stage. The elastic modulus, Poisson's ratio, and other mechanical properties of the subgrade fill material are highly sensitive to temperature and humidity. For example, increased temperature or moisture content typically leads to a decrease in the stiffness of the subgrade material, resulting in a larger strain response under the same load. This environmentally induced response change can be confused with response changes caused by structural defects. The purpose of the correction is to decouple these two effects. This correction process is based on a pre-established correction model that describes the quantitative relationship between the components of the response differentiation feature vector and the temperature and humidity parameters. The correction process can be expressed as: , in, It is the corrected response differential feature vector. It is the uncorrected, original response differential feature vector. and These are the extracted temperature and humidity parameters. This represents one or a set of correction functions or models. The model... This model can be constructed through systematic laboratory research or numerical simulation. For example, by conducting mechanical tests on roadbed material samples under different temperature and humidity conditions, the variation of their stress-strain relationship with temperature and humidity can be obtained. Then, combined with numerical simulation methods such as the finite element method, the dynamic response of the roadbed under different temperature, humidity, and damage conditions can be simulated, thereby calibrating the dependence of each component of the eigenvector on temperature and humidity, and finally solidifying it into a modified model. In actual diagnosis, by substituting the V, T, and H data obtained on-site into the model, the result after eliminating environmental impacts can be calculated. .

[0046] This method effectively isolates the interference of environmental temperature and humidity changes on the mechanical response of the roadbed, solving the problem that traditional strain-response-based diagnostic methods are susceptible to environmental factors such as seasons, diurnal temperature variations, and rainfall. By specifically modifying the response-differentiated feature vector, this vector can more purely and accurately reflect the abnormal mechanical behavior caused by the roadbed's internal structural defects. This enhances the stability, reliability, and all-weather adaptability of the diagnostic results, reduces the risk of false alarms and missed alarms due to environmental changes, and enables the intelligent diagnostic system to maintain a high level of diagnostic performance under a wider range of complex working conditions.

[0047] Optionally, the step of using the response differential feature vector to diagnose subgrade diseases and obtain a subgrade health status diagnosis report including disease type and risk level includes: Construct a roadbed health status diagnostic model; The modified response differential feature vector is input into the roadbed health status diagnosis model for inference analysis, and a preliminary diagnosis result containing information on disease type and risk level is generated. Based on the preliminary diagnostic results, a roadbed health status diagnostic report is generated.

[0048] Specifically, the first step is to construct a roadbed health status diagnostic model. This model is an intelligent classification and evaluation system pre-trained with a large amount of historical data. Essentially, it is a mapping function designed to learn the complex relationship between input features and specific disease conclusions.

[0049] During the diagnostic phase, the response differential feature vector, which has already undergone temperature and humidity correction in the previous steps, is input into this pre-built roadbed health status diagnostic model. Upon receiving this vector, the model performs inference analysis. Inference analysis is the process by which the model applies its internally stored knowledge to calculate and judge the new input data. The final output is the most likely type of disease and its corresponding risk level. This output is called the preliminary diagnostic result.

[0050] Finally, based on the preliminary diagnostic results, the system will automatically generate a roadbed health status diagnostic report. This report is not just a simple list of the preliminary diagnostic results; it also integrates all the key information from the diagnostic process, including the location information of the defects (derived from the coordinates of the structural reflection anomaly zone), the specific values ​​of the response differential feature vectors on which the diagnosis is based, the types of defects diagnosed, the assessed risk level, and the confidence level of the diagnostic conclusions.

[0051] Optionally, the construction of the roadbed health status diagnostic model includes: Obtain response differential feature vectors with known disease type labels and construct a historical training set; The historical training set is used to conduct supervised training on the probabilistic classification model, so that the model learns the mapping relationship from feature vectors to disease types and corresponding confidence probabilities, and obtains a roadbed health status diagnosis model. The roadbed health status diagnostic model includes multiple heterogeneous base classifiers.

[0052] Specifically, the first step is to construct a historical training set. This training set is built by collecting a large amount of roadbed sample data with clearly known defect type labels. For each sample, such as a roadbed location confirmed as "void" through core sampling or excavation, the technique described in the preceding steps of this patented method is used to generate its corresponding response differential feature vector through radar imagery and strain response analysis. This feature vector is then paired with the real label "void" to form a training sample. This process is repeated to collect and process numerous samples covering various states such as "void," "loose," "settlement," and "healthy and normal," ultimately forming a historical training set containing a large number of (feature vector, defect label) data pairs.

[0053] Next, a supervised training process is performed on a probabilistic classification model containing multiple heterogeneous base classifiers based on this historical training set. Heterogeneous base classifiers refer to models composed of various different types of machine learning algorithms; for example, support vector machines (SVM), decision trees, and gradient boosting machines (GBM) can be integrated simultaneously. The training goal is for the model to learn an accurate mapping from response differential feature vectors to disease types and their corresponding confidence probabilities. During training, feature vectors from the historical training set are input into the model, and the model outputs a predicted probability distribution for disease types. This mapping can be abstractly represented as: , in, It is the modified response differential feature vector of the input; This represents the diagnostic model to be trained. It is a probability vector ,in This represents the confidence probability that the model predicts the sample belongs to the i-th disease type, where k is the total number of all predefined disease types. Training algorithms, such as backpropagation or gradient descent, calculate a loss function by comparing the model's predicted probability distribution P with the sample's true label, and iteratively adjust the parameters of the heterogeneous base classifiers within the model to minimize this loss function. Training is complete when the model reaches the preset performance metrics on the validation set. The final model, the roadbed health status diagnostic model, encapsulates complex nonlinear knowledge ranging from multi-dimensional mechanical response characteristics to specific disease types.

[0054] Optionally, the step of inputting the response differential feature vector into the roadbed health status diagnosis model for inference analysis to generate a preliminary diagnosis result containing information on disease type and risk level includes: The corrected response differential feature vectors are input into the multiple heterogeneous base classifiers to obtain multiple parallel diagnostic sub-results; Based on preset decision fusion rules, multiple parallel diagnostic sub-results are weighted and fused to generate preliminary diagnostic results.

[0055] Specifically, the response differential feature vectors, which have already undergone temperature and humidity correction in the previous steps, are first used as a unified input and simultaneously fed into multiple parallel heterogeneous base classifiers contained within the roadbed health status diagnosis model. These heterogeneous base classifiers are instances of different algorithm models, such as support vector machines, decision trees, and neural networks, which have each learned the mapping knowledge from features to defects during the training phase.

[0056] Each heterogeneous base classifier independently analyzes the differential feature vector of the input response and outputs a diagnostic sub-result. This sub-result is a probability vector with a dimension equal to the number of predefined disease types. Each element in the vector represents the confidence probability that the classifier classifies the input sample as belonging to the corresponding disease type. Next, the system performs a weighted fusion of these parallel diagnostic sub-results according to a pre-defined decision fusion rule. This fusion process can be represented by the following formula: , in, It is the final integrated probability vector obtained after fusion; It is the preset weight of the j-th heterogeneous base classifier. This weight is usually determined based on the performance of the classifier on the historical validation set. The classifier with better performance is given a higher weight, and the sum of all weights is 1. This is the diagnostic sub-result output by the j-th heterogeneous base classifier, i.e., its predicted probability vector. The comprehensive probability vector is then calculated. Then, the model will select the category with the highest probability value as the final disease type diagnosis conclusion, and combine this probability value with the disease type to output the corresponding risk level, which together constitutes a complete preliminary diagnosis result.

[0057] By utilizing multiple heterogeneous base classifiers for parallel diagnosis and weighted fusion, the accuracy and robustness of diagnostic results can be improved. Different types of classifiers have different focuses in decision boundaries and feature sensitivities; their combination can achieve complementary advantages, correct potential misjudgments by individual models, and thus obtain more stable and reliable diagnostic performance than any single classifier. This strategy reduces the model's dependence on specific training data distributions and enhances its generalization ability when faced with new and unseen data. By weighted fusion of the outputs of multiple models, the final diagnostic conclusion is not a single judgment, which makes the preliminary diagnostic results more reliable.

[0058] Optionally, the method includes: Extract the location and risk level information of the defects from the roadbed health status diagnosis report; The location and risk level information of the disease are rendered onto a three-dimensional digital model of the roadbed; Based on the rendering results, a three-dimensional visual risk map is generated.

[0059] Specifically, the system first needs to automatically extract key information from the previously generated roadbed health status diagnosis report through a data parsing program. This mainly includes two aspects: first, the location information of the defect, which is usually expressed in three-dimensional coordinates (X, Y, Z) in the roadbed coordinate system to accurately represent the center point of the defect or its area boundary identified in radar images; second, the risk level information corresponding to the location, such as qualitative classifications like "Level I (high risk)" or "Level II (medium risk)" or a quantitative risk index.

[0060] After extracting the information, a pre-built 3D digital model of the roadbed needs to be invoked. This model is an accurate virtual copy of the roadbed structure, typically constructed based on Building Information Modeling (BIM) technology, design drawings, or measurement technologies such as laser scanning. It contains precise geometric dimensions of the roadbed, layered structures such as pavement layers, base courses, subbase courses, and roadbed fill. This 3D digital model has a coordinate system consistent with the physical roadbed and serves as the basic canvas for spatial information rendering.

[0061] Next, the extracted disease location and risk level information are rendered onto this 3D digital model. The system first locates a unique spatial point or region in the 3D digital model based on the disease location coordinates. Then, centered on this location, the system generates a visual 3D geometry, such as a sphere, cube, or an irregular shape matching the shape of the diseased area, to represent the presence of the disease. The visual attributes of this geometry, such as color, transparency, or size, are set according to its corresponding risk level information. For example, a color mapping scheme from green to red can be established, where green represents low risk, yellow represents medium risk, and red represents high risk. Thus, a high-risk internal cavity would be rendered as a striking red semi-transparent body in the 3D model.

[0062] Finally, once all diagnosed defects are rendered onto the 3D digital model according to their location and risk level, a 3D visualized risk map of the roadbed is generated. This is not just a static image, but an interactive 3D scene. Users can freely rotate, scale, translate, and section the map to examine the internal health of the roadbed from any angle. Clicking on the rendered defect geometry can also link back to the original diagnostic report to view more detailed diagnostic evidence and data.

[0063] Based on the same inventive concept, such as Figure 3 As shown, the present invention also provides an intelligent diagnostic system for the multi-dimensional health status of roadbeds, the system comprising: The data acquisition module is used to acquire radar images of the internal structure of the roadbed, time-series strain data from multiple measuring points, and real-time location information of moving loads. The region identification module is used to identify areas with abnormal structural reflection and areas with normal structural reflection based on the radar image of the internal structure. The dynamic response extraction module is used to extract the load strain dynamic response waveform and the reference dynamic response waveform from the multi-point time-series strain data based on the real-time location information of the moving load and corresponding to the abnormal reflection area and the normal reflection area of ​​the structure. The differential feature generation module is used to quantitatively compare the load strain dynamic response waveform with the reference dynamic response waveform and generate a response differential feature vector. The intelligent diagnostic module is used to diagnose roadbed diseases using the response differential feature vector, and to obtain a roadbed health status diagnostic report that includes disease type and risk level.

[0064] To verify the feasibility of this invention in practice, it was applied to the routine health monitoring of a section of a highway from K55+200 to K55+500. Due to complex geological conditions, this section had previously experienced localized roadbed settlement and other problems. The maintenance unit hoped to use the method of this invention for high-precision preventative diagnosis of this section.

[0065] In this embodiment, technicians first deployed multi-point fiber optic strain gauges and temperature and humidity sensors along the road section, and then used a vehicle-mounted ground-penetrating radar system to acquire radar images of the roadbed's internal structure. A standard test truck with a total weight of 45 tons served as the mobile load, equipped with an RTK-GPS device to record its real-time location information. Data acquisition was conducted in September 2023, covering various environmental and traffic conditions.

[0066] During the area identification phase, the system analyzes the acquired radar images of the internal structure. For example, at location K55+310, the system uses a deep learning network based on a self-attention mechanism to perform semantic segmentation on the radar image, identifying an anomalous structural reflection area of ​​approximately 4.5 square meters at a depth of 2.5 meters below the left shoulder of the roadbed. This area exhibits discontinuous and rapidly attenuating reflection signals. Simultaneously, at location K55+450, a region with homogeneous structure and continuous, stable reflection signals was selected as the normal structural reflection area.

[0067] During the dynamic response extraction phase, when the test truck passed through the monitored section at a constant speed of 70 km / h, the system accurately extracted the time-series strain data during the period when the truck load acted on the two areas mentioned above, based on the RTK-GPS location information. For example, the system determined that the time window for the truck acting on the abnormal area at K55+310 was 1.35 seconds, and the time window for acting on the normal area at K55+450 was 1.32 seconds. After the extracted initial dynamic response waveform was denoised by Butterworth low-pass filtering, a smooth load-strain dynamic response waveform (originating from the abnormal area) and a baseline dynamic response waveform (originating from the normal area) were generated.

[0068] During the differential feature generation stage, the system quantifies and compares the two waveforms. The calculated dynamic time warp distance between them is 18.2. Further extraction of morphological features and calculation of relative differences reveal the following: a relative difference in peak value of +41% (indicating greater strain in the anomalous area under the same load), a relative difference in response duration of +25% (indicating slower stress recovery in this region), and a relative difference in waveform energy of +52%. These values ​​collectively constitute an original response differential feature vector. At that time, the average temperature measured by the sensor deployed in the K55+310 area was 22℃, and the average moisture content was 23%. The vector was corrected using a preset temperature and humidity correction model to eliminate the influence of environmental factors on material stiffness, ultimately yielding the corrected response differential feature vector.

[0069] During the intelligent diagnosis phase, the corrected response differential feature vector is input into a pre-built subgrade health status diagnosis model. This model integrates three heterogeneous base classifiers: Support Vector Machine (SVM), Gradient Boosting Machine (GBM), and Random Forest. After parallel inference by the three classifiers, they are weighted and fused based on their performance on the historical validation set. Finally, the model outputs the preliminary diagnosis result: the disease type is "loose subgrade," with a confidence probability of 95%; the risk level is "Level II (Medium Risk)." The system then generates a subgrade health status diagnosis report containing the above information.

[0070] To verify the accuracy of the diagnosis, the maintenance unit subsequently conducted core sampling at location K55+310 and found that the roadbed fill material at that location did indeed have insufficient compaction and high porosity, verifying the accuracy of the model's diagnosis of "loose roadbed".

[0071] Finally, the system renders the diagnosed defect location (K55+310), defect type (loose subgrade), and risk level (Level II - Medium Risk) onto a 3D digital model of the highway section. In the model, the defect location is marked as a prominent yellow semi-transparent area, and technicians can interactively view its detailed diagnostic report. This provides intuitive and accurate spatial information support for maintenance decisions.

[0072] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0073] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. An intelligent diagnostic method for the multi-dimensional health status of roadbeds, characterized in that, The method includes: Acquire radar images of the roadbed's internal structure, time-series strain data from multiple measuring points, and real-time location information of moving loads; Based on the radar image of the internal structure, areas with abnormal structural reflection and areas with normal structural reflection are identified. Based on the real-time location information of the moving load, and corresponding to the abnormal reflection area and the normal reflection area of ​​the structure, the load strain dynamic response waveform and the reference dynamic response waveform are extracted from the multi-point time-series strain data, respectively. The load strain dynamic response waveform is quantitatively compared with the reference dynamic response waveform to generate a response difference feature vector; The roadbed disease diagnosis is performed using the aforementioned response differential feature vector, resulting in a roadbed health status diagnosis report that includes disease type and risk level.

2. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 1, characterized in that, The identification of anomalous and normal structural reflection areas based on the radar image of the internal structure includes: Image analysis is performed on the radar image of the internal structure to generate an image feature map; The weights of regions related to structural defects in the image feature map are enhanced by using a self-attention mechanism, and semantic segmentation is performed to output a region mask labeled with location information. Based on the region mask, the structural reflection anomalous region and the structural reflection normal region are determined.

3. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 1, characterized in that, The real-time location information based on the moving load, corresponding to the structural reflection anomaly zone and the structural reflection normal zone, includes the extraction of load strain dynamic response waveforms and reference dynamic response waveforms from the multi-point time-series strain data, respectively: The real-time location information of the moving load is mapped to a preset roadbed coordinate system to determine the start and end time windows of the moving load's action in the abnormal reflection zone and the normal reflection zone of the structure. The time-series strain data of the multiple measurement points are extracted according to the start and end time windows to form the initial dynamic response waveform; The initial dynamic response waveform is filtered and denoised to generate a smooth response waveform, and the load strain dynamic response waveform and the reference dynamic response waveform are extracted based on the smooth response waveform.

4. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 1, characterized in that, The quantitative comparison of the load-strain dynamic response waveform with the reference dynamic response waveform to generate a response difference feature vector includes: Calculate the dynamic time warp distance between the load strain dynamic response waveform and the reference dynamic response waveform; Morphological features are extracted from the two waveforms respectively, and the relative difference between the morphological features is calculated. The dynamic time-warped distance is combined with the relative difference to form a response-differentiated feature vector.

5. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 4, characterized in that, The method further includes: Acquire multi-point temperature and humidity sensing data of the structural reflection anomaly zone, and extract temperature and humidity parameters from them; The temperature and humidity parameters are used to correct the response differential feature vector.

6. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 5, characterized in that, The process of using the differentiated feature vectors of the responses to diagnose subgrade diseases, and obtaining a subgrade health status diagnostic report that includes disease type and risk level, includes: Construct a roadbed health status diagnostic model; The modified response differential feature vector is input into the roadbed health status diagnosis model for inference analysis, and a preliminary diagnosis result containing information on disease type and risk level is generated. Based on the preliminary diagnostic results, a roadbed health status diagnostic report is generated.

7. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 6, characterized in that, The constructed roadbed health status diagnostic model includes: Obtain response differential feature vectors with known disease type labels and construct a historical training set; The historical training set is used to conduct supervised training on the probabilistic classification model, so that the model learns the mapping relationship from feature vectors to disease types and corresponding confidence probabilities, and obtains a roadbed health status diagnosis model. The roadbed health status diagnostic model includes multiple heterogeneous base classifiers.

8. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 7, characterized in that, The step of inputting the response differential feature vector into the roadbed health status diagnosis model for inference analysis to generate preliminary diagnostic results containing information on disease type and risk level includes: The corrected response differential feature vectors are input into the multiple heterogeneous base classifiers to obtain multiple parallel diagnostic sub-results; Based on preset decision fusion rules, multiple parallel diagnostic sub-results are weighted and fused to generate preliminary diagnostic results.

9. The intelligent diagnostic method for multi-dimensional health status of roadbed according to claim 1, characterized in that, The method includes: Extract the location and risk level information of the defects from the roadbed health status diagnosis report; The location and risk level information of the disease are rendered onto a three-dimensional digital model of the roadbed; Based on the rendering results, a three-dimensional visual risk map is generated.

10. An intelligent diagnostic system for the multi-dimensional health status of a roadbed, characterized in that, The system includes: The data acquisition module is used to acquire radar images of the internal structure of the roadbed, time-series strain data from multiple measuring points, and real-time location information of moving loads. The region identification module is used to identify areas with abnormal structural reflection and areas with normal structural reflection based on the radar image of the internal structure. The dynamic response extraction module is used to extract the load strain dynamic response waveform and the reference dynamic response waveform from the multi-point time-series strain data based on the real-time location information of the moving load and corresponding to the abnormal reflection area and the normal reflection area of ​​the structure. The differential feature generation module is used to quantitatively compare the load strain dynamic response waveform with the reference dynamic response waveform and generate a response differential feature vector. The intelligent diagnostic module is used to diagnose roadbed diseases using the response differential feature vector, and to obtain a roadbed health status diagnostic report that includes disease type and risk level.

Images