Pavement and underground disease AI space-time digital intelligence platform

Abstract

The invention relates to the technical field of image processing, and discloses a pavement and underground disease AI space-time digital intelligence platform, which is characterized in that a multi-scale wave field unified model is constructed by fusing micro-vibration sensor and underground radar data, then two types of signals are mapped to the same space, unified representation is generated, mutual interference signals are analyzed to form a conversion table, and the conversion table is used for processing the pavement and underground disease AI. And obtaining a cross-physical field enhanced feature set, generating an associated feature body through three-dimensional registration, obtaining a super-resolution three-dimensional feature body by using a super-resolution algorithm, and further generating a disease diagnosis report. Through a physical constraint deep fusion imaging system, the defects of a traditional technology are overcome, unified modeling and imaging of earth surface and underground diseases are achieved, the detection precision is improved, and early-stage accurate diagnosis and treatment of the diseases are assisted.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more specifically, to an AI spatio-temporal digital intelligence platform for pavement and underground diseases. Background Art

[0002] In the field of road surface and underground structure disease detection and imaging, there are many key problems in traditional technologies: different technical means are used for surface and underground disease detection, resulting in data fragmentation and information silos, and it is impossible to present an integrated disease panorama; surface micro-vibration sensing and underground radar imaging technologies are independent of each other, and the mutual interference in signal processing is regarded as an interference factor; the heterogeneity of different physical field signals makes information fusion only stay at the post-processing level, and it is difficult to achieve deep fusion at the physical level; in harsh environments, the reliability of a single sensing modality drops significantly, seriously affecting the detection accuracy; existing technologies cannot effectively utilize the internal relationship between different physical fields, resulting in small diseases and early lesions being difficult to be reliably identified. In view of these problems, the present invention proposes to construct a theory and application system for integrated surface and underground disease imaging with physical constraint deep fusion, realize unified physical modeling and collaborative imaging of surface and underground diseases, break through the limitations of existing technologies, and provide a new technical solution for the early accurate identification and treatment of road diseases. Summary of the Invention

[0003] The present invention provides an AI spatio-temporal digital intelligence platform for pavement and underground diseases, which solves the key problems in traditional road disease detection technologies in related technologies, such as data fragmentation, technology independence, limited information fusion, insufficient reliability, and failure to utilize internal relationships, and it is difficult to achieve integrated and accurate identification and collaborative treatment of surface and underground diseases.

[0004] The present invention provides an AI spatio-temporal digital intelligence method for pavement and underground diseases, including: Receiving surface data collected by a micro-vibration sensor and imaging data obtained by an underground radar, mapping the micro-vibration waveform and the radar electromagnetic wave signal into the same physical space, and generating a multi-scale wave field unified model representing the common characteristics of the two physical fields; Using an interference signal analysis method to extract valuable information from the interference signal, forming an interference information conversion table, and realizing cross-physical field information complementary enhancement based on the interference information conversion table to obtain a cross-physical field enhanced feature set; Precisely registering the cross-physical field enhanced feature set in three-dimensional space to generate a three-dimensional surface and interior associated feature body, applying a physically constrained super-resolution enhancement algorithm to the three-dimensional surface and interior associated feature body to form a super-resolution three-dimensional feature body, analyzing the super-resolution three-dimensional feature body, and generating an integrated surface and underground disease diagnosis report.

[0005] Further, the steps of generating the multi-scale wave field unified model representing the common characteristics of two physical fields include: extracting and normalizing the parameters of various physical field signals to obtain a multi-physical field unified parameter table; creating a set of non-linear coupled wave field dynamic equations based on the multi-physical field unified parameter table; constructing a deep neural network architecture based on the cross-physical field coupling equations, and training through optimizing the objective function to obtain the multi-scale wave field unified model.

[0006] Further, the deep neural network architecture is trained through the following objective function: ; Information retention loss function: ; Physical constraint loss function: ; Among them, and are neural network models, is the physical model function, is the trade-off parameter, is the input data, is the output data, represents the L2 norm.

[0007] Further, the steps of extracting valuable information from the interference signal include: using an adaptive iterative separation algorithm to extract the interference components from the mixed signal, generating an interference signal feature map, establishing a conversion model from the interference signal to useful information, and forming an interference information conversion table; the adaptive iterative separation algorithm is implemented through the iterative formula: ; ; Among them, is the original mixed signal, is the interference component obtained in the th iteration, is the estimated pure signal of the th class in the th iteration, is the estimated pure signal of the th class in the th iteration, is the signal extraction function, is the algorithm parameter set, where is the number of physical fields, is the number of iterations.

[0008] Further, the conversion model from the interference signal to useful information is realized through the following mathematical transformation: ; Among them, is the cross-interference signal feature, is the intermediate information representation, and are the transformation matrix and the bias vector, is the enhanced information after transformation, is the non-linear activation function, is the feature transformation function, is the element-wise multiplication.

[0009] Furthermore, the steps of implementing cross-physical-field information complementary enhancement are realized by the following enhancement function: ; wherein, is the original feature of the th type of physical field, is the enhanced feature, is the corresponding cross-interference information, is the feature of the th type of physical field, is the feature transformation function from the th type of physical field to the th type of physical field: wherein, is the feature adaptive transformation network.

[0010] Furthermore, the steps of accurately registering the cross-physical-field enhanced feature set in the three-dimensional space adopt the following spatial mapping equation: ; wherein, is the feature mapping value of the spatial point , is the value of the th feature field, is the position-related dynamic weight, and its calculation method is: ; wherein, is the reliability measure of the feature field at the position , is the temperature parameter, is the exponential function.

[0011] Furthermore, the steps of applying the physical constraint super-resolution enhancement algorithm to the three-dimensional table-related feature volume are realized by solving the following optimization problem: ; wherein, is the reconstructed high-resolution feature volume, is the downsampling matrix, is the observed data, and is the regularization parameter, is the prediction function based on the physical model, is the total variation regularization term, is the physical constraint regularization term, represents the L2 norm.

[0012] Furthermore, the step of analyzing the super-resolution three-dimensional feature volume applies a multi-scale anomaly detection function: ; ; wherein, is the outlier at position , is the eigenvalue of the super-resolution three-dimensional feature volume at position , is the feature distance function, is the normal feature template, is the feature reliability function, is the Mahalanobis distance calculation, is the feature covariance matrix.

[0013] The present invention provides a pavement and underground disease AI spatio-temporal digital intelligence system, including: A data acquisition module for receiving surface data collected by a micro-vibration sensor and imaging data obtained by a ground penetrating radar; A wave field unified modeling module for constructing a multi-scale wave field unified model and mapping the micro-vibration waveform and the radar electromagnetic wave signal into the same physical space; An interference signal analysis module for extracting valuable information in the interference signal and forming an interference information conversion table; A feature enhancement module for realizing cross-physical field information complementary enhancement and obtaining a cross-physical field enhanced feature set; A spatial registration module for accurately registering the cross-physical field enhanced feature set in a three-dimensional space to generate a three-dimensional table-related feature volume; A super-resolution enhancement module for applying a super-resolution enhancement algorithm with physical constraints to the three-dimensional table-related feature volume to form a super-resolution three-dimensional feature volume; A disease diagnosis module for analyzing the super-resolution three-dimensional feature volume and generating an integrated table and interior disease diagnosis report.

[0014] The beneficial effects of the present invention are as follows: By constructing a surface-integrated disease imaging system with deep integration of physical constraints, the present invention solves problems in traditional technologies such as data fragmentation, improper processing of signal interference, insufficient information fusion, low reliability in harsh environments, and difficulty in identifying minor diseases, realizes unified modeling and collaborative imaging of surface and underground diseases, improves detection accuracy and reliability, and provides an innovative solution for early and accurate identification and treatment of road diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a schematic flow diagram of an AI spatio-temporal digital method for road surface and underground diseases of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] Now, the subject matter described herein will be discussed with reference to example embodiments. It should be understood that discussing these embodiments is only to enable those skilled in the art to better understand and thus implement the subject matter described herein, and changes can be made to the functions and arrangements of the elements discussed without departing from the protection scope of the content of this specification. Each example can omit, substitute, or add various processes or components as needed. Additionally, the features described in some examples can also be combined in other examples.

[0017] In at least one embodiment of the present invention, an AI spatio-temporal digital method for road surface and underground diseases is disclosed, as Figure 1 shown, including the following steps: Step 100, constructing a unified multi-scale wavefield model; using the surface data collected by micro-vibration sensors and the imaging data obtained by ground penetrating radar, a unified representation model characterizing the common features of the two physical fields is generated through a unified multi-scale wavefield modeling method. This step maps the surface micro-vibration waveform and the ground penetrating radar electromagnetic wave signal to the same mathematical description framework to achieve the unified expression of the two heterogeneous physical fields. Specifically, it includes the following steps: Sub-step 101, parameterization of physical field characteristics: Input: Surface micro-vibration raw signal data and ground penetrating radar echo raw data; Execution process: Parameter extraction and normalized representation are performed on various physical field signals. Time-frequency domain features such as frequency, amplitude, and phase of the micro-vibration signal are extracted, parameters such as the electromagnetic wave reflection characteristics, dielectric constant, and propagation speed of the radar signal are obtained, and a unified parameter mapping function is applied, and its calculation formula is:

[0018] where is the parameterization mapping function of the th type of physical field, is the raw signal, are the extracted th characteristic parameters respectively; Output: Obtain a unified multi-physical field parameter table, which contains the set of characteristic parameters and physical meaning explanations of the signals in each physical field, and is used for the subsequent construction of the wave field dynamics equation.

[0019] Sub-step 102, Construction of the wave field dynamics equation: Input: Unified multi-physical field parameter table; Execution process: Create a set of non-linear coupled wave field dynamics equations according to the unified multi-physical field parameter table, which is used to establish the mutual influence relationship between different physical fields. Its calculation formula is:

[0020] where, represents the second-order partial derivative of the th type of wave field (micro-vibration mechanical wave or radar electromagnetic wave) with respect to time and reflects the change rate of the wave field with time; is the propagation speed corresponding to the th type of wave field; is the Laplace operator (second-order spatial derivative), which is used to describe the change of the wave field in space; represents the total influence of all other wave fields on the th type of wave field except the th type of wave field itself. Among them, represents the gradient of the th type of wave field, is the coupling function between wave fields, which represents the influence of the th type of wave field on the th type of wave field; The specific form of the coupling function is:

[0021] where, , and are the linear coupling, gradient coupling and non-linear coupling coefficients respectively; Output: Form a cross-physical field coupling equation set, which records the mutual interaction mathematical relationship and corresponding coefficients between each physical field, and provides a theoretical basis for the physical constraint co-representation space.

[0022] Sub-step 103, Formation of the physical constraint co-representation space: Input: Cross-physical field coupling equation set and original signal data; Execution process: Based on the cross-physical field coupling equation set, construct a deep neural network architecture, so that the neural network simultaneously meets the dual goals of information retention and physical law constraint. The network is trained through the objective function, and its calculation formula is:

[0023] Among them, is a parameter in the neural network model. By optimizing this parameter, the entire objective function reaches the minimum value; is the information retention loss, whose role is to ensure that the feature representation retains the key information of the original signal. Its calculation formula is:

[0024] Among them, is the input data, is the expected output, is the neural network model, represents the L2 norm; : is a trade-off parameter used to adjust the relative importance of data and physical constraints; is the physical constraint loss, which ensures that the feature representation conforms to physical laws. Its calculation formula is:

[0025] Among them, is the neural network model, is the physical model function; Output: Obtain the physical constraint co-representation space, that is, the multi-scale wave field unified model This model contains a set of neural network parameters for the unified representation of signals in different physical fields , physical constraint definitions and their weight configurations , which can be expressed as:

[0026] Among them, , respectively represent the th group of parameters of the neural network, , respectively represent the th physical constraint condition, , respectively represent the corresponding th constraint weight. This model can be used for subsequent cross-interference signal analysis.

[0027] Step 200, construct a cross-interference signal analysis and feature enhancement mechanism; receive the multi-scale wave field unified model and the original mixed signal, and use the cross-interference signal analysis method to extract the valuable information contained in the cross-interference signal. This step converts the cross-interference signal, which is usually regarded as interference, into an information source for enhancing the perception ability, and realizes mutual enhancement between signals. Specifically, it includes the following steps: Sub-step 201, cross-interference signal feature separation: Input: Mixed signal data of surface micro-vibration and ground radar, and multi-scale wavefield unified model; Execution process: Use the adaptive iterative separation algorithm to extract the cross-interference components from the mixed signal, which is implemented through the iterative formula. The calculation formula is:

[0028]

[0029] where, is the original mixed signal, is the cross-interference component obtained in the th iteration, is the th iteration estimate of the th type of pure signal, is the signal extraction function, is the algorithm parameter set, is the number of physical fields, is the number of iterations; Output: Generate a cross-interference signal feature map, which contains the feature representations of various cross-interference signals, time-frequency distribution maps, and signal correlation strength matrices for subsequent information enhancement.

[0030] Sub-step 202, cross-interference mode conversion: Input: Cross-interference signal feature map and multi-scale wavefield unified model; Execution process: Establish a conversion model from cross-interference signals to useful information, which is implemented through mathematical transformation. The calculation formula is:

[0031]

[0032] where, is the cross-interference signal feature, is the intermediate information representation, and are the conversion matrix and bias vector, is the enhanced information after conversion, is the non-linear activation function, is the feature transformation function, is the element-wise multiplication; Output: Form a cross-interference information conversion table, which contains the conversion rule set, information enhancement parameters, and feature mapping relationships, providing a basis for cross-physical field information enhancement.

[0033] Sub-step 203, cross-physical field complementary enhancement: Input: Cross-interference information conversion table and original physical field characteristics; Execution process: By complementing each physical field information, the feature expression ability of a single physical field is enhanced. Its calculation formula is:

[0034] where, is the original feature of the type of physical field, is the enhanced feature, is the feature of the type of physical field, is the corresponding cross-interference information; is from the type of physical field to the type of physical field feature conversion function, and its calculation formula is:

[0035] where, is the feature adaptive transformation network, which automatically adjusts the transformation strategy according to the specific physical field characteristics; Output: Obtain a cross-physical field enhanced feature set , which integrates the information advantages of each physical field and can be expressed as:

[0036] where, , , are the enhanced feature representations of the th physical field respectively, is the enhanced effect quantization index, is the feature quality evaluation result, is the total number of physical fields. This feature set provides a rich data basis for subsequent imaging.

[0037] Step 300, implement super-resolution integrated surface and subsurface imaging of diseases; based on receiving the cross-physical field enhanced feature set, use the integrated surface and subsurface super-resolution imaging algorithm to achieve unified high-precision imaging of road surface and subsurface diseases. This step completes the transformation from independent detection to integrated imaging, generating a three-dimensional imaging result containing surface and subsurface correlation information. Specifically, it includes the following steps: Sub-step 301, surface and subsurface correlation spatial registration: Input: Cross-physical field enhanced feature set and original road detection coordinate data - Execution process: Precisely align and map the surface and subsurface features in three-dimensional space, and apply the spatial mapping equation. Its calculation formula is:

[0038] where, is the characteristic mapping value of a spatial point , is the value of the th feature field is the position-related dynamic weight, satisfying , is the number of feature fields

[0039] The weight calculation uses the softmax function, and its calculation formula is

[0040] where is the reliability measure of the feature field at position , is the temperature parameter is the exponential function Output: Generate a three-dimensional table-associated feature volume, which contains the spatial correspondence, association strength distribution, and integrated three-dimensional feature data of surface and subsurface features, providing a basis for super-resolution reconstruction

[0041] Sub-step 302, physical-constrained super-resolution enhancement Input: Three-dimensional table-associated feature volume and multi-scale wave field unified model Execution process: Use the physical-constrained regularization method to improve the spatial resolution of the feature volume. This process is completed by the super-resolution enhancement algorithm, and its calculation formula is

[0042] where is the reconstructed high-resolution feature volume is the downsampling matrix is the observed data is the total variation regularization term

[0043] is the physical-constrained regularization term

[0044] where is the prediction function based on the physical model and are regularization parameters , , are the spatial coordinates of the feature volume, where is the spatial gradient operator Output: A super-resolution three-dimensional feature volume is formed, which has a higher spatial resolution than the original data and contains refined road surface and underground structure features, resolution improvement quantification indicators, and quality assessment data.

[0045] Sub-step 303, integrated surface and subsurface disease location and analysis: Input: Super-resolution three-dimensional feature volume and a pre-calibrated disease feature library; Execution process: Perform disease detection, location, and classification on the super-resolution three-dimensional feature volume, and apply a multi-scale anomaly detection function, whose calculation formula is:

[0046] Where, is the outlier value at position , is the feature distance function, is the normal feature template, is the feature reliability function.

[0047] The feature distance function uses the Mahalanobis distance, and its calculation formula is:

[0048] Where, is the Mahalanobis distance calculation, is the feature covariance matrix; Combining the disease type, location, and severity information obtained above, analyze the possible disease causes, and then match the predetermined disease repair plan library through the disease diagnosis information to give targeted repair suggestions; Output: Integrate the above disease diagnosis information into an integrated diagnosis report, and output an integrated surface and subsurface disease diagnosis report.

[0049] This report includes the disease type determination result, accurate three-dimensional position coordinates, disease severity level assessment, and repair plan suggestions, providing a scientific basis for road maintenance decision-making.

[0050] The physical constraint depth fusion integrated surface and subsurface disease imaging system implemented through the above steps achieves the following technical effects: Improved disease detection sensitivity: This system realizes the high-precision detection ability for fine road diseases, improves the surface micro-crack detection sensitivity from 0.5 mm of traditional methods to 0.1 mm, and extends the underground cavity identification depth from 2.5 meters of traditional methods to 4 meters, providing technical possibilities for ultra-early disease detection; Enhanced environmental adaptability: This system maintains a detection accuracy of more than 98% under harsh environmental conditions (such as heavy rain, high temperature, severe cold, vibration noise, and electromagnetic interference), which is significantly improved compared to the 75% accuracy of traditional single-modal detection technologies under the same conditions; Improved accuracy of disease classification: This system has increased the accuracy of disease classification in the early stage from 70% of traditional methods to 95%, reducing false alarms and missed detections, and significantly reducing unnecessary maintenance costs caused by misjudgments. Function of predicting the development of diseases: By obtaining integrated surface and underground disease information, this system can predict the development trend of diseases in the next 3 - 6 months with an accuracy of 85%, providing a scientific basis for road maintenance decisions. Optimization of detection speed and energy consumption: This system has increased the road disease detection speed by 2.5 times while reducing energy consumption by 40%, breaking through the trade - off between speed and accuracy in traditional detection technologies. Ability of surface - underground correlation analysis: This system has realized surface - underground disease correlation analysis for the first time, being able to determine the influence mechanism of underground structure changes on the development of surface diseases and the reflection law of surface conditions on underground structures, forming a complete disease evolution chain and laying a foundation for the full - cycle road health management. Most significantly, for the detection of composite diseases such as "surface micro - cracks + base moisture intrusion", the accuracy of this method reaches 93.8%, while that of traditional methods is only 59.3%, an increase of 34.5 percentage points.

[0051] In this application example, two key technical effects of the physically - constrained depth - integrated surface - underground disease imaging system are mainly verified: the ability to detect fine diseases and the environmental adaptability.

[0052] To verify the system's ability to detect fine diseases, 32 disease sample points of different types and severities were selected on the example section, including surface micro - cracks (width 0.05 - 0.5mm), interlayer debonding, base segregation, etc. Real - labeled data was obtained through core - drilling sampling. The detection results of this method were compared with those of traditional methods, as shown in Table 1: Table 1: Comparison of the ability to detect fine diseases

[0053] Data shows that the average detection rate of this method reaches 93.6%, an increase of 32.5 percentage points compared with 61.1% of traditional methods. Especially in the detection of fine diseases, the minimum surface crack width that can be identified reaches 0.08mm, which is 6.25 times higher than 0.5mm of traditional methods; the maximum underground disease depth that can be detected reaches 380cm, which is 58.3% higher than 240cm of traditional methods. This effect benefits from the innovative use of cross - interference signals as information carriers and the application of physically - constrained super - resolution enhancement algorithms.

[0054] To verify the robust performance of the system under different environmental conditions, the same detection point on the example section was tested multiple times under different weather conditions and road surface states to evaluate the detection accuracy of the system. The results are shown in Table 2: Table 2: Comparison of detection accuracy under different environmental conditions:

[0055] Improvement in anti-interference ability = (Error reduction rate of this method - Error reduction rate of traditional method) / Error reduction rate of traditional method × 100% The data shows that this method maintains a high detection accuracy under various environmental conditions, with an average of 96.5%, which is 25.1 percentage points higher than 71.4% of the traditional method. Especially under harsh environmental conditions (such as heavy rain / waterlogged road surface, vehicle vibration interference, etc.), the advantages of this method are more obvious, with the accuracy increasing by more than 30 percentage points and the anti-interference ability increasing by more than 80%. This is mainly due to the fact that this method deeply integrates different physical field information, enabling the system to maintain overall performance through complementary information of another modality when one sensing modality is interfered with.

[0056] In summary, this application example fully verifies the technical effects of the physical constraint deep fusion table and surface integrated disease imaging system, especially having significant advantages in micro disease detection and environmental adaptability, providing strong support for the early accurate identification and full life cycle management of road diseases.

[0057] In an embodiment of the present invention, an example of the aforementioned AI spatio-temporal digital method for pavement and underground diseases is provided: Application scenario description: This embodiment is applied to the pavement disease detection and diagnosis of the section from K1457 + 200 to K1457 + 400 of the Beijing-Shanghai Expressway. This section is an asphalt concrete pavement built in 2008. After more than 10 years of use, various diseases such as surface cracks, interlayer debonding, and base moisture intrusion have occurred. This section has the following characteristics: The annual average daily traffic volume (AADT) of the section is as high as 42,000 vehicle trips, and the proportion of heavy traffic volume is 28.6%; The section spans two geological conditions: from K1457 + 200 to K1457 + 300 is a rocky foundation, and from K1457 + 300 to K1457 + 400 is a soft soil foundation; The section is significantly affected by seasonal rainfall, with an annual rainfall of 1,280 mm; The regional temperature difference is large, with the highest temperature in summer reaching 40 °C and the lowest temperature in winter reaching -10 °C; The section is located in the suburban area of the city, with various electromagnetic interference sources; The historical maintenance records of road sections show that traditional single detection methods cannot effectively identify deep - layer diseases, resulting in the recurrence of diseases in the short term after maintenance.

[0058] This application scenario covers complex pavement structures, diverse environmental conditions, and severe traffic loads, representing the typical difficulties in disease detection of high - grade highways. This implementation method aims to achieve integrated internal and surface disease imaging through a physical constraint depth fusion method, providing a scientific basis for accurate diagnosis and targeted repair.

[0059] On the section of the Beijing - Shanghai Expressway from K1457 + 200 to K1457 + 400, a detection vehicle equipped with the following equipment was used for data collection: A set of laser scanning systems (scanning frequency: 2000Hz, scanning width: 3.6m); A set of micro - vibration sensor arrays (frequency response: 0.5 - 5000Hz, sampling rate: 25kHz); A set of ground - penetrating radar systems (center frequency: 1.6GHz, scanning depth: 5m); A set of infrared thermal imaging systems (temperature resolution: 0.05°C, spatial resolution: 3mm); Data collection was carried out under normal traffic conditions. The detection vehicle traveled at a constant speed (60km / h), and each sensor collected data synchronously. The collection process is shown in Table 3: Table 3: Parameters of the data collection process:

[0060] Multiple collections were carried out under different environmental conditions to verify the stability and adaptability of the system under various conditions.

[0061] Example of Step 100, constructing a multi - scale wave - field unified model: Apply the first - step processing to the collected raw data to extract the characteristic information of each physical field. The results are shown in Table 4: Table 4: Results of cross - physical - field feature extraction:

[0062] Feature extraction uses The convolutional sparse transform algorithm, combined with expert knowledge for feature selection, effectively removes redundant information, reduces the feature dimension, and at the same time retains the maximum amount of effective information.

[0063] Example of Step 200, constructing an interference signal analysis and feature enhancement mechanism: Use the adaptive iterative separation algorithm to process the interference components between each sensing signal. For the typical mixed signal at the 250th meter of the example section K1457 + 250, the iterative separation results are shown in Table 5: Table 5: Results of iterative separation of mutual interference signals (at the location of K1457+250 of the road section):

[0064] The analysis results of mutual interference signals at four typical disease locations in the road section were particularly concerned. Through the mutual interference information conversion model, the mutual interference components were converted into valuable feature information, as shown in Table 6: Table 6: Results of information enhancement of mutual interference signals (at four typical disease locations):

[0065] Analysis shows that the mutual interference signal recovery rate of the composite disease of settlement and water intrusion at the location of K1457+315 of the road section is the highest, reaching 90.4%. This is mainly because water intrusion has a significant impact on both the electromagnetic wave and the thermal field, and the mutual interference signal contains rich water distribution information.

[0066] Example of step 300, realizing integrated imaging of super-resolution disease surface and interior: Apply the physical constraint super-resolution enhancement algorithm to enhance the three-dimensional feature volume, and focus on the feature fusion situation at different depths, as shown in Table 7: Table 7: Feature fusion parameters at different depths (section from K1457+300 to K1457+320):

[0067] The finally generated integrated high-resolution three-dimensional disease image of the surface and interior is compared with the traditional method, as shown in Table 8: Table 8: Comparison of integrated imaging results of the surface and interior:

[0068] At the typical disease area of K1457+315, the integrated imaging of the surface and interior clearly shows the complete disease chain from surface settlement to subbase water intrusion, revealing the evolution mechanism that water intrusion leads to the decline of subgrade strength and then causes surface settlement, providing a scientific basis for precise maintenance.

[0069] In an embodiment of the present invention, for the scenario of dynamic evolution prediction and early warning of road surface and underground structure diseases, an AI spatio-temporal digital method for road surface and underground diseases further includes the following steps: On the basis of retaining the capabilities of Embodiment 1, this embodiment realizes the accurate prediction of the future development of diseases by introducing the time dimension and the multi-scale dynamic prediction model. The "Multi-scale Spatio-temporal Evolution Prediction Theory of Diseases" (MSEPT) is proposed and implemented. The main innovation points include: The multi-scale dynamic equations of disease evolution establish the association between surface and interior disease evolution through the diffusion coefficient tensor, environmental response function, and structural stress function; An adaptive multi-time scale prediction mechanism that dynamically adjusts the prediction time window and model parameters; a method for quantifying prediction uncertainty based on physical constraints to achieve reliability assessment of prediction results; An algorithm for identifying key inflection points in disease evolution to predict the acceleration points and stable points of disease development; A dynamic response model for the coupled action of environment - material - structure to enhance the adaptability to external condition changes.

[0070] This embodiment extends static disease detection to dynamic prediction, achieving a technical leap from "diagnosing what" to "predicting what will happen".

[0071] Steps 100 to 300 are exactly the same as those in Embodiment 1. The unified physical field representation model, enhanced feature information, and high-precision disease diagnosis report obtained in these three steps will be used as the input data for Step 400.

[0072] Step 400, disease spatio-temporal evolution prediction and early warning; through multiple sub-steps, the function of realizing the transition from static disease diagnosis to dynamic spatio-temporal evolution prediction and early warning is achieved: Sub-step 401, construction of disease spatio-temporal database: Receiving the integrated table - form disease diagnosis report data and historical monitoring records as input information; Structuring the received data according to spatio-temporal dimensions to obtain three types of time - series data: Disease state time - series data: including time - varying parameters such as disease type, location, size, and severity; Environmental condition time - series data: including environmental monitoring parameters such as temperature, humidity, and precipitation; Load condition time - series data: including load parameters such as traffic flow, vehicle type, and axle load The data is organized into a structured data set after spatio - temporal coordinate alignment, and its calculation formula is:

[0073] where, represents the state parameter vector of the th disease sample at time , is the corresponding environmental condition vector, is the load condition vector, is the total number of disease samples, is the th sample's number of observed time points; Generate a disease spatio - temporal data set, which contains complete records of spatio - temporally aligned disease states, environmental conditions, and load conditions.

[0074] Sub-step 402, Construction of the disease evolution mathematical model: Receive the disease spatio-temporal dataset and the multi-scale wave field unified model as input information; Based on the received data, construct a differential equation for the evolution of the disease state, and its calculation formula is:

[0075] Where, represents the spatial position at time of the disease state variable, is the gradient operator, represents the disease diffusion coefficient tensor, represents the environmental response function, represents the structural stress function.

[0076] Establish parametric models for different types of diseases respectively: Surface crack disease model: Adopt an anisotropic diffusion coefficient, and this coefficient can be expressed by the formula , where is the spatial position coordinate. This coefficient is mainly used to characterize the influence of temperature change and stress action on surface crack diseases. Assume the temperature is , and the stress is , a relationship function can be established:

[0077] The specific function form needs to be determined according to actual physical laws and experimental data; Interlayer debonding disease model: Adopt interface characteristic parameters, and let the interface characteristic parameters be , and these parameters mainly characterize the interface stress and the influence of freeze-thaw action on interlayer debonding diseases, and a relationship function can be established:

[0078] Where represents the freeze-thaw related parameters, and the specific form of the function needs to be determined according to the actual situation; Underground disease model: Adopt material characteristic parameters, and let the material characteristic parameters be , which mainly characterize the influence of moisture change and load transfer on underground diseases, and a relationship function can be established:

[0079] The function Its specific form needs to be determined according to the actual physical process and experimental data; When determining the model parameters, it is achieved by minimizing the mean square error between the predicted error and the measured value. Let the predicted value be , and the measured value be , the mean square error The calculation formula is:

[0080] where, is the disease type, is the number of samples. By adjusting the model parameters, make reach the minimum, so as to determine each model parameter.

[0081] Finally, a mathematical model of disease evolution is generated. This model takes the current disease state and environmental conditions as inputs, and through the parameterized model established above and the determined parameters, outputs the predicted value of the disease state at the future time point .

[0082] Sub-step 403, time-scale adaptive prediction execution; receive the mathematical model of disease evolution and the current disease state data as input information; Based on the received data, perform multi-time-scale prediction calculations: Calculate the time interval sequence according to the disease type characteristics, and its calculation formula is:

[0083] where, represents the th prediction time interval, represents the basic time interval parameter, represents the disease state change rate function, represents the disease state change acceleration function, and represent the adjustment parameters; Fuse the prediction results of multiple time-scale sub-models, and its calculation formula is:

[0084] where, represents the fused prediction result, represents the prediction result of the th sub-model, represents the time-related weight function, represents the total number of sub-models; Apply physical constraint conditions to correct the prediction results to ensure compliance with the laws of material mechanics and structural deformation; Generate disease status prediction data, which includes the predicted values of the disease status and their confidence intervals in the short term (within 3 months), medium term (6 months), and long term (1 year).

[0085] Sub-step 404, identifying the inflection points of disease evolution; receiving the disease status prediction data as input information, and based on the received data, performing key inflection point identification analysis: Let the prediction trajectory be a function , whose time derivative is , by calculating to obtain the rate of change, and then taking the derivative of to obtain the acceleration ; set the rate of change threshold and the acceleration threshold , when and , the corresponding time point is the time point at which the rate of change and acceleration change significantly; For the disease status vector , assuming its dimension is , map it to the -dimensional phase space , and construct the phase space trajectory . By analyzing the geometric characteristics of the phase space trajectory, such as the curvature of the trajectory , the distance between trajectories , etc., when these geometric characteristics change significantly, it is identified as the structural change of the disease status vector in the high-dimensional space; Let the disease status variable be , and the environmental factor be , the calculation formula for the mutual information value is:

[0086] where, and are the value sets of and respectively, is the joint probability that and , and are the marginal probabilities of and respectively. When is greater than a certain set threshold , the corresponding environmental factor is the key environmental trigger factor.

[0087] Let the rate of change, acceleration, changes in the geometric characteristics of the phase space trajectory, and mutual information values obtained from the above steps be , and assign weights to each factor . Then the importance score of the inflection point is calculated as follows:

[0088] Based on the calculated value, sort the inflection points.

[0089] Generate inflection point data for disease evolution, which includes key turning time points within the predicted time range. Each inflection point contains the time position, predicted disease state value, and importance score.

[0090] Sub-step 405, disease risk level calculation; receive the predicted disease state data and inflection point data for disease evolution as input information; Based on the received data, perform risk quantification calculations: Apply a time decay function to calculate the uncertainty index at different prediction time points; Calculate the risk value of the road section, and its calculation formula is:

[0091] Among them, represents the risk value at position at time , represents the disease severity function, represents the prediction uncertainty, represents the infrastructure importance index; Divide the risk into four levels according to the risk value threshold: normal, attention, warning, and emergency; Generate a disease risk assessment report, which includes a spatio-temporal distribution map of disease risk, a classification warning threshold division, and a risk reminder at key time points.

[0092] Sub-step 406, maintenance plan generation: Data reception: Receive the disease risk assessment report and the maintenance technology parameter library as input information; Maintenance plan calculation: Calculation of the optimal intervention time window: Let the cost function be , and the risk function be . Calculate the optimal intervention time window through the formula, and its calculation formula is:

[0093] Among them, is a functional relationship determined according to the actual situation. This function needs to comprehensively consider cost and risk factors to determine the most appropriate intervention time; Selection of maintenance technology plan: Based on the type of disease , development trend and road condition environment , select an appropriate maintenance technology plan from the maintenance technology parameter library , and its calculation formula is:

[0094] where is a selection function, which screens out the most suitable plan from numerous maintenance technology plans according to different input conditions; Calculation of long-term effect and economic evaluation indicators: For different maintenance plans , calculate their long-term effect indicators and economic evaluation indicators . The long-term effect indicators can be obtained by comprehensively considering various factors such as road performance improvement and disease recurrence probability; the economic evaluation indicators can involve calculations such as maintenance cost and expected revenue; Report generation: Generate an auxiliary report for maintenance decision-making, which includes the recommended maintenance time window , suggestions for hierarchical intervention plans and prediction of intervention effects (including long-term effect indicators and economic evaluation indicators and other related contents).

[0095] The technical effects obtained through the above steps include: Accuracy of multi-time scale prediction: Achieve accurate predictions of different time spans through disease state prediction data. The prediction accuracy for the short term (within 3 months) is 92%, for the medium term (6 months) is 85%, and for the long term (1 year) is 75%, providing reliable data for the whole life cycle management; Early identification of the inflection point of disease development: Achieve early capture of the key transition time point through the disease evolution inflection point data, and on average, discover the transition of the disease from slow to rapid deterioration 15 - 30 days in advance; Comprehensive consideration of environmental factors: Incorporate more than 15 environmental parameters into consideration through the environmental response function in the disease evolution mathematical model, significantly improving the adaptability of the prediction in a changing environment; Quantification of the reliability of prediction results: The confidence interval estimation included in the disease risk assessment report provides a 95% confidence level, enabling maintenance decisions to be made on the premise of considering risks; Optimization of maintenance cost and service life: By means of the best intervention time window and plan recommended in the maintenance decision-making assistance report, the road maintenance cost is reduced by 30%, and the service life is extended by 15% - 20%. Precision of maintenance decision-making: The accuracy rate of the intervention plan recommendation in the maintenance decision-making assistance report for different disease types reaches 87%, effectively supporting the maintenance management decision-making.

[0096] In summary, based on the static disease detection, by introducing the time dimension and constructing a disease spatio-temporal evolution prediction model, this embodiment has achieved a technical leap from "disease identification" to "development prediction", providing a complete technical solution for the whole life cycle maintenance management of roads.

[0097] The embodiments of the present invention have been described above, but these embodiments are not limited to the above specific implementation manners. The above specific implementation manners are merely illustrative rather than restrictive. Under the inspiration of this embodiment, those of ordinary skill in the art can also make more equivalent embodiments in various forms, all of which fall within the protection scope of this embodiment.

Claims

1. An AI spatiotemporal digital intelligence method for pavement and underground diseases, characterized in that: The following steps are involved: Receive surface data collected by micro-vibration sensors and imaging data acquired by underground radars, map micro-vibration waveforms and radar electromagnetic wave signals into the same physical space, and generate a multi-scale wave field unified model that represents the common characteristics of the two physical fields; The mutual interference signal analysis method is used to extract valuable information from the mutual interference signal, and a mutual interference information conversion table is formed. Based on the mutual interference information conversion table, cross-physical field information complementary enhancement is achieved to obtain a cross-physical field enhanced feature set. The cross-physical field enhanced feature set is accurately aligned in three-dimensional space to generate a three-dimensional surface-interior associated feature body. A physical constraint super-resolution enhancement algorithm is applied to the three-dimensional surface-interior associated feature body to form a super-resolution three-dimensional feature body. The super-resolution three-dimensional feature body is analyzed to generate an integrated surface-interior and surface-interior disease diagnosis report.

2. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 1 is characterized in that: The steps of generating a unified multi-scale wave field model that characterizes the common characteristics of two physical fields include: extracting parameters and normalizing various physical field signals to obtain a unified parameter table of multiple physical fields; creating a set of nonlinear coupled wave field dynamics equations based on the unified parameter table of multiple physical fields; constructing a deep neural network architecture based on the cross-physical field coupling equation group, and obtaining a unified multi-scale wave field model by optimizing objective function training.

3. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 2 is characterized in that: The deep neural network architecture is trained with the following objective function: ； Information preservation loss function: ; Physical constraint loss function: ; in, and is the neural network model, is the physical model function, To balance the parameters, For input data, To output data, represents the L2 norm.

4. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 1 is characterized in that: The step of extracting valuable information from the mutual interference signal includes: extracting the mutual interference component from the mixed signal using an adaptive iterative separation algorithm, generating a mutual interference signal characteristic spectrum, establishing a conversion model from the mutual interference signal to useful information, and forming a mutual interference information conversion table; the adaptive iterative separation algorithm is implemented by an iterative formula: ； ； in, is the original mixed signal, For the The mutual interference component obtained by the iteration is For the The estimated iterative Pure signal, For the The estimated iterative Pure signal, is the signal extraction function, is the algorithm parameter set, where is the number of physical fields, is the number of iterations.

5. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 4 is characterized in that: The conversion model of mutual interference signal to useful information is realized by the following mathematical transformation: ； in, is the mutual interference signal characteristic, is the intermediate information representation, and is the transformation matrix and bias vector, is the enhanced information after transformation, is a nonlinear activation function, is the feature transformation function, is element-wise multiplication.

6. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 1 is characterized in that: The steps to achieve cross-physics information complementary enhancement are implemented through the following enhancement functions: ； in, For the Physics-like primitive features, To enhance the features, is the corresponding mutual interference information, For the Physics-like features, For Class physics to Physics-like characteristic transfer functions: in, It is a feature adaptive transformation network.

7. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 1 is characterized in that: The step of accurately registering cross-physics enhanced feature sets in 3D space uses the following spatial mapping equation: ； in, For space point The feature map value of For the The value of the characteristic field, is the position-related dynamic weight, which is calculated as: ； in, is the characteristic field at position The reliability measure of is the temperature parameter, is an exponential function.

8. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 1 is characterized in that: The steps of the super-resolution enhancement algorithm for applying physical constraints to the three-dimensional surface-to-surface associated feature bodies are achieved by solving the following optimization problem: ； in, is the reconstructed high-resolution feature volume, is the downsampling matrix, is the observed data, and is the regularization parameter, is the prediction function based on the physical model, is the total variation regularization term, is the physical constraint regularization term, represents the L2 norm.

9. The AI ​​spatiotemporal digital intelligence method for pavement and underground diseases according to claim 1 is characterized in that: The steps for analyzing super-resolution 3D features are to apply multi-scale anomaly detection functions: ； ； in, For location The outliers of For super-resolution 3D features at position The characteristic value of is the feature distance function, is a normal feature template, is the characteristic reliability function, is the Mahalanobis distance calculation, is the feature covariance matrix.

10. A road surface and underground disease AI spatiotemporal digital intelligence system, used to execute a road surface and underground disease AI spatiotemporal digital intelligence method as described in any one of claims 1 to 9, characterized in that: include: A data acquisition module, used to receive surface data collected by the micro-vibration sensor and imaging data acquired by the underground radar; The wave field unified modeling module is used to construct a multi-scale wave field unified model and map the micro-vibration waveform and radar electromagnetic wave signal into the same physical space; A mutual interference signal analysis module is used to extract valuable information from the mutual interference signal and form a mutual interference information conversion table; A feature enhancement module is used to realize cross-physical field information complementary enhancement and obtain a cross-physical field enhanced feature set; The spatial registration module is used to accurately register the cross-physical field enhanced feature set in three-dimensional space to generate three-dimensional surface-to-surface associated feature bodies; A super-resolution enhancement module is used to apply a super-resolution enhancement algorithm of physical constraints to the three-dimensional surface-interior associated feature body to form a super-resolution three-dimensional feature body; The disease diagnosis module is used to analyze super-resolution three-dimensional feature bodies and generate an integrated internal and external disease diagnosis report.

Images