An intelligent extraction method and system for urban road collapse hidden danger sensitive elements

Abstract

The application is particularly a kind of intelligent extraction method and system for urban road collapse hidden danger sensitive elements, relating to the technical field of urban road underground hidden danger detection, comprising: obtaining the template signal corresponding to each sensitive element category in the sensitive element feature library, respectively calculating the waveform similarity and the frequency spectrum matching degree of each independent signal source component and each category template signal.In the application, the blind source separation processing mode of wavelet packet decomposition combined with FastICA algorithm is adopted to decompose the multi-source aliasing echo signal into statistically independent signal source components, solving the technical problem that the feature extraction on the aliasing signal directly leads to feature ambiguity due to the mutual interference of signal sources, making the feature extraction process of sensitive elements have traceability and physical interpretability.

Description

Technical Field

[0001] This invention relates to the field of urban road underground hazard detection technology, and in particular to an intelligent extraction method and system for sensitive elements of urban road collapse hazards. Background Technology

[0002] The potential for underground road subsidence poses a serious threat to the safety of urban infrastructure. Ground-penetrating radar (GPR), a commonly used non-destructive testing method, is widely applied to identify hidden and sensitive elements such as underground cavities, loose areas, and water-rich bodies. However, the echo signals acquired by GPR are characterized by non-stationarity, non-linearity, and multi-source aliasing. The original signal contains superimposed components from different geological targets and background clutter, with each signal source overlapping with the others.

[0003] Existing deep learning-based end-to-end identification methods process aliased signals by inputting the entire signal into the network, lacking the ability to physically separate the signal source, resulting in insufficient interpretability of the extracted results. In terms of hazard identification, existing methods typically use fixed thresholds for judgment.

[0004] However, the background statistical characteristics of ground-penetrating radar signals vary significantly under different geological conditions and detection scenarios. Fixed thresholds cannot adapt to these changes, making it difficult to effectively separate signal components related to sensitive elements under multi-source signal aliasing conditions. Furthermore, the fixed threshold method is prone to technical problems such as increased missed detections or false alarms when the scene is shifted. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent extraction method and system for sensitive elements of potential urban road collapse hazards in order to solve the above-mentioned problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A smart extraction method for sensitive elements of urban road collapse hazards, comprising: The raw echo data acquired by the ground penetrating radar is sequentially subjected to zero-bias correction, bandpass filtering, time-varying gain compensation, and signal standardization to obtain standardized signal data. Blind source separation processing is performed on the standardized signal data to obtain multiple independent signal source components. The blind source separation processing includes: performing wavelet packet decomposition on each channel data to obtain a wavelet packet node coefficient matrix; inputting the wavelet packet node coefficient matrix as multi-channel observation data into an independent component analysis algorithm for demixing to obtain multiple independent components; and sorting and sign-correcting each independent component according to the correlation coefficient between each independent component and a preset reference signal. The template signals corresponding to each sensitive element category in the sensitive element feature library are obtained. The waveform similarity and spectrum matching degree between each independent signal source component and the template signal of each category are calculated respectively. The waveform similarity and spectrum matching degree are weighted and fused to obtain the comprehensive sensitivity similarity of each independent signal source component relative to each sensitive element category. Based on the comprehensive sensitivity similarity and prior information of the current detection scenario, Bayesian inference is used to calculate the posterior probability of each independent signal source component belonging to each sensitive element category. An adaptive judgment threshold is calculated based on the cost of missed detection and the cost of false alarm. The posterior probability is compared with the adaptive judgment threshold to determine whether each independent signal source component is a sensitive element. The sensitivity level of the independent signal source components determined to be sensitive elements is classified, and the sensitive element extraction result is output.

[0007] Preferably, the bandpass filtering process includes a direct wave removal operation, wherein the direct wave removal method is as follows: calculate the average background signal of the multi-channel data, and subtract the average background signal from each channel data to weaken the direct wave component; the time-varying gain compensation uses an exponential gain function to perform time-varying amplitude compensation on the signal, so that the reflected signals at different depths have comparable amplitude levels; the signal standardization uses Z-score standardization processing.

[0008] Preferably, the independent component analysis algorithm is the FastICA algorithm. The FastICA algorithm takes the multi-channel observation data as input, uses negative entropy to approximate the non-Gaussianity of the signal, and iteratively updates each row vector of the separation matrix by maximizing the negative entropy objective function. In each iteration, the row vector is updated using the first and second derivatives of the nonlinear function. After the update, the row vector is normalized and orthogonalized to ensure the statistical independence between the components. The iteration stops when the change in the row vector between two adjacent iterations is lower than the convergence criterion, and the separated signal matrix is ​​output as multiple independent components.

[0009] Preferably, after obtaining the plurality of independent components, the method further includes: evaluating the non-stationarity of each independent component, wherein the non-stationarity evaluation is performed by calculating the instantaneous frequency variance of each independent component within a sliding time window; when the instantaneous frequency variance of a certain independent component is greater than a preset non-stationarity threshold, the independent component is input into a variational mode decomposition algorithm for fine decomposition, wherein the variational mode decomposition algorithm adaptively decomposes the independent component into several intrinsic mode functions, calculates the sample entropy of each intrinsic mode function, retains the intrinsic mode functions with sample entropy lower than a preset entropy threshold as effective signal components, removes the intrinsic mode functions with sample entropy higher than the preset entropy threshold as noise components, replaces the original independent components with the retained effective signal components, and updates the plurality of independent signal source components.

[0010] Preferably, the waveform similarity is calculated as follows: the waveform correlation coefficient between the independent signal source component and the template signal is calculated, and the absolute value of the waveform correlation coefficient is taken as the waveform similarity measure. The waveform correlation coefficient is the product of the covariance of the independent signal source component and the template signal divided by their standard deviations. The spectral matching degree is calculated as follows: Fourier transform is performed on the independent signal source component and the template signal respectively to extract their respective spectral feature parameters. The spectral feature parameters include the center frequency, 3dB bandwidth and the main frequency energy ratio. The spectral feature parameters are then normalized by mean based on the range to form spectral feature vectors. The spectral matching degree of the two spectral feature vectors is calculated using cosine similarity.

[0011] Preferably, the calculation of the comprehensive sensitivity similarity also incorporates time-frequency distribution similarity. The time-frequency distribution similarity is obtained by using short-time Fourier transform to obtain the time-frequency distribution matrices of the independent signal source components and the template signal, and calculating the structural similarity index between the two time-frequency distribution matrices. The structural similarity index measures the comprehensive similarity of the two time-frequency distribution matrices in three dimensions: brightness, contrast, and structure. It is calculated by calculating the mean, variance, and covariance of all corresponding elements of the two time-frequency distribution matrices, respectively. The numerator is constructed based on the product of the means and covariance of the two matrices, and the denominator is constructed based on the sum of the squares and variances of the means of the two matrices. Stability constants are added to the numerator and denominator to prevent the denominator from being zero. The ratio of the numerator to the denominator is used as the structural similarity index. The comprehensive sensitivity similarity is a weighted fusion result of the waveform similarity, the spectral matching degree, and the time-frequency distribution similarity, with the sum of the weight coefficients being one.

[0012] Preferably, the step of using Bayesian inference to calculate the posterior probability of each independent signal source component belonging to each sensitive element category includes: obtaining the prior probability of each sensitive element category in the current detection scenario, wherein the prior probability is determined based on the geological conditions of the detection area, historical collapse records, and current seasonal information; obtaining the likelihood function obtained in advance based on statistical analysis of forward modeling data and historical measured data, wherein the likelihood function includes the probability density of comprehensive sensitivity similarity under the condition that sensitive elements exist and the probability density of comprehensive sensitivity similarity under the condition that sensitive elements do not exist; and, according to Bayes' theorem, dividing the product of the likelihood function and the prior probability by the marginal probability to obtain the posterior probability of each independent signal source component belonging to each sensitive element category.

[0013] Preferably, the adaptive judgment threshold is calculated as follows: the false alarm cost is divided by the sum of the false alarm cost and the missed detection cost to obtain the adaptive judgment threshold; wherein, the missed detection cost is the cost of judging a real hidden danger as normal, the false alarm cost is the cost of judging a normal area as a hidden danger, and the values ​​of the missed detection cost and the false alarm cost are determined according to the engineering risk level of the detected object; the sensitivity level is divided as follows: when the posterior probability is not lower than the first fixed quantile threshold, it is classified as high sensitivity level; when the posterior probability is not lower than the second fixed quantile threshold and lower than the first fixed quantile threshold, it is classified as medium sensitivity level; when the posterior probability is not lower than the adaptive judgment threshold and lower than the second fixed quantile threshold, it is classified as low sensitivity level.

[0014] Preferably, before outputting the sensitive element extraction results, the method further includes: performing spatial positioning and scale estimation processing on each independent signal source component identified as a sensitive element; the spatial positioning includes depth positioning and horizontal position determination; the depth positioning method is: based on the two-way travel time of the echo signal corresponding to each independent signal source component, the average electromagnetic wave propagation speed of the medium in the area is used as half of the product of the average electromagnetic wave propagation speed and the two-way travel time as the target depth; the horizontal position determination method is: based on the occurrence position of each independent signal source component in each ground-penetrating radar data, combined with the mileage information of the radar scan, the horizontal coordinates of the sensitive element are obtained; the scale estimation method is: based on the duration and amplitude envelope width of the reflected signal, combined with the spectrum analysis results, the thickness and lateral distribution range of the sensitive element are estimated; the sensitive element category, sensitivity level, spatial position and scale information are combined as a complete sensitive element extraction result and output.

[0015] An intelligent system for extracting sensitive elements of potential urban road collapse hazards includes: The signal preprocessing module is used to sequentially perform zero-bias correction, bandpass filtering, time-varying gain compensation, and signal standardization on the raw echo data acquired by the ground penetrating radar to obtain standardized signal data. The blind source separation module is used to perform wavelet packet decomposition on the standardized signal data to obtain the wavelet packet node coefficient matrix, input the wavelet packet node coefficient matrix as multi-channel observation data into the independent component analysis algorithm for demixing to obtain multiple independent components, and sort and sign-correct each independent component according to the correlation coefficient between each independent component and the preset reference signal to obtain multiple independent signal source components. The sensitivity similarity calculation module is used to obtain the template signal corresponding to each sensitive element category in the sensitive element feature library, calculate the waveform similarity and spectrum matching degree between each independent signal source component and the template signal of each category, and perform weighted fusion on the waveform similarity and spectrum matching degree to obtain the comprehensive sensitivity similarity of each independent signal source component relative to each sensitive element category. The sensitive element determination module is used to calculate the posterior probability of each independent signal source component belonging to each sensitive element category based on the comprehensive sensitivity similarity and prior information of the current detection scene using Bayesian inference, calculate an adaptive determination threshold based on the cost of missed detection and the cost of false alarm, compare the posterior probability with the adaptive determination threshold to determine whether each independent signal source component is a sensitive element, classify the sensitivity level of the independent signal source components determined to be sensitive elements, and output the sensitive element extraction results.

[0016] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention employs a blind source separation processing method combining wavelet packet decomposition and FastICA algorithm to decompose multi-source aliased echo signals into statistically independent signal source components. This solves the technical problem of feature ambiguity caused by mutual interference between signal sources when directly extracting features from aliased signals, and makes the feature extraction process of sensitive elements traceable and physically interpretable.

[0017] 2. This invention solves the technical problem of increased missed detections or false alarms caused by the fixed threshold method due to its inability to adapt to changes in background statistical characteristics under different geological conditions and detection scenarios by adopting a Bayesian inference framework combined with a cost-sensitive adaptive decision threshold. This allows the decision threshold to be dynamically adjusted according to prior information of the scenario and the level of engineering risk. At the same time, the posterior probability is mapped to three sensitivity levels: high, medium and low, which achieves the technical effect of providing differentiated decision information for hazard disposal. Attached Figure Description

[0018] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0020] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0021] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0022] In this embodiment, it includes: Step 1: Acquire raw ground-penetrating radar echo data and preprocess it to obtain standardized signal data. Raw echo data generated by a multi-channel ground-penetrating radar system scanning underground areas of urban roads was acquired. The raw echo data was in B-scan or C-scan format, with each channel containing a time-sampled sequence reflecting the intensity of reflected signals as electromagnetic waves propagate through the underground medium. The raw echo data underwent zero-bias correction, bandpass filtering, time-varying gain compensation, and signal normalization sequentially to obtain standardized signal data. Specifically, zero-bias correction subtracted the mean of each time-series data channel to eliminate DC bias components; bandpass filtering retained the signal components within the effective frequency band while filtering out low-frequency drift and high-frequency random noise; time-varying gain compensation used an exponential gain function to compensate for the time-varying amplitude of the signal, ensuring comparable amplitude levels for reflected signals at different depths; and signal normalization applied Z-score normalization to each pre-processed channel.

[0023] It should be noted that the above bandpass filtering process also includes direct wave removal. Direct waves include the direct coupling wave between the ground reflected wave and the transmitting and receiving antennas, which manifests as an echo with extremely high intensity and extremely short delay, and can mask the reflected signal of shallow targets. The direct wave removal method is to calculate the average background signal of the multi-channel data, and subtract the average background signal from each channel data, thereby weakening the direct wave component.

[0024] Step 2: Perform blind source separation processing on the standardized signal data to obtain multiple independent signal source components. Blind source separation processing is performed on standardized signal data to decompose the aliased signal into multiple statistically independent signal source components, thus obtaining multiple independent signal source components. Blind source separation processing includes the following sub-steps: Step 201: Perform wavelet packet decomposition on each channel of the standardized signal data to obtain the coefficient matrix of each channel at multiple wavelet packet nodes. Wavelet packet decomposition performs multi-scale fine decomposition of each signal in both the time and frequency domains. The coefficients of each wavelet packet node represent the energy distribution of the signal in a specific frequency band and time period, thereby expanding a single-channel signal into data from multiple virtual channels.

[0025] Step 202: The wavelet packet node coefficient matrix of each channel's data is used as multi-channel observation data and input into the FastICA algorithm for demixing to obtain multiple independent components. In the ground-penetrating radar signal scenario, the observed echo signal can be regarded as a linear mixture of multiple signal sources, and its mathematical expression is: ; in, For the observed signal matrix, For an unknown matrix of independent signal sources, The mixture matrix is ​​unknown. The FastICA algorithm uses multi-channel observation data. Using negative entropy as input, the non-Gaussianity of the signal is approximated, and the separation matrix is ​​iteratively updated. Output the separation result until convergence. ,in To separate the signal matrix, make The components are statistically independent of each other as much as possible, thus approximating the real signal source. .

[0026] Furthermore, the calculation method for the negative entropy approximation in the FastICA algorithm described above is as follows: For the separation matrix... Each row vector By maximizing the objective function To measure the non-Gaussianity of the corresponding components, where Represents the expectation operator. It is a nonlinear function. Let be a standard normally distributed random variable. Representing vectors transpose, This is the projection signal along the current separation direction. In each iteration, according to... renew ,in for The first derivative, for The second derivative, after updating, is used for... The components are normalized and orthogonalized to ensure statistical independence, continuing until two adjacent iterations. When the change in the value is lower than the convergence criterion, the iteration stops and the corresponding independent component is output.

[0027] Step 203: Calculate the correlation coefficient between each independent component and the preset reference signal. Based on the correlation coefficient, sort and sign-correct the independent components to obtain multiple independent signal source components. The reference signal mentioned above is a typical sensitive element reflection simulation signal generated based on forward modeling. Sign correction refers to multiplying the independent component by a negative value when the correlation coefficient between a certain independent component and the reference signal is negative. This ensures that the sign of the independent components is consistent with the reference signal.

[0028] It should be noted that the number of decomposition layers in the above wavelet packet decomposition is determined based on the ground-penetrating radar antenna frequency and the target detection depth range. When the antenna frequency is 400MHz and the target detection depth range is 0 to 5m, the number of decomposition layers can be 3 or 4, corresponding to expanding a single-channel signal into 8 or 16 virtual channels.

[0029] In this embodiment of the application, in order to improve the separation accuracy of strong non-stationary signal components, based on step 202, the following steps are also included: non-stationarity evaluation is performed on each independent component output by FastICA. The non-stationarity evaluation method is to calculate the instantaneous frequency variance of each independent component within a sliding time window. When the instantaneous frequency variance of a certain independent component is greater than a preset non-stationarity threshold, the independent component is further input into a variational mode decomposition algorithm for fine decomposition. The variational mode decomposition algorithm takes the independent component as input, adaptively decomposes the independent component into several intrinsic mode functions (IMFs), calculates the sample entropy of each IMF, retains IMFs with sample entropy lower than a preset entropy threshold as valid signal components, removes IMFs with sample entropy higher than the preset entropy threshold as noise components, replaces the original independent components with the retained valid signal components, and updates multiple independent signal source components.

[0030] Furthermore, the aforementioned non-stationarity threshold and entropy threshold are empirical thresholds pre-calibrated based on statistical analysis of forward modeling data and historical measured data. The non-stationarity threshold is used to distinguish between stationary components and strongly non-stationary components, while the entropy threshold is used to distinguish between effective signal components with regular structures and noise components with strong randomness. Both the non-stationarity threshold and entropy threshold are stored in the system parameter library in the form of dimensionless statistics. When processing different detection scenarios, the corresponding pre-calibrated values ​​can be called according to the antenna frequency and geological condition category.

[0031] Furthermore, the processing procedure of the above variational mode decomposition algorithm is as follows: taking the independent components to be decomposed as input signals, and adaptively decomposing the independent components to be decomposed into... by solving the constrained variational problem. eigenmode functions ,in The total number of intrinsic mode functions. The first The, the The first to the second The input signal is composed of eigenmode functions, each with its own center frequency and finite bandwidth. The sum of these eigenmode functions reconstructs the original input signal. The constrained variational problem is solved iteratively in the frequency domain using the alternating direction multiplier method. In each iteration, the frequency domain representation of each eigenmode function and its corresponding center frequency are updated sequentially. The process continues until the change in the frequency domain representation of each eigenmode function between two consecutive iterations falls below the convergence criterion, at which point the output signal is obtained. There are 1 eigenmode functions. Among them, The value of is preset based on the spectral complexity of the input signal and can be determined by observing the number of major peaks in the spectrum.

[0032] Step 3: Calculate the overall sensitivity similarity between each independent signal source component and the template signal in the sensitive element feature library. Obtain a pre-generated sensitive element feature library, which contains template signals corresponding to various categories of sensitive elements for potential collapse, including cavities, loose areas, and water-rich bodies. Calculate the similarity between each independent signal source component and the template signals of each category in the sensitive element feature library to obtain the comprehensive sensitivity similarity of each independent signal source component relative to each sensitive element category. The similarity measurement calculation includes the following sub-steps: Step 301: Calculate the independent signal source components With template signal waveform correlation coefficient : ; in, for and covariance, and They are respectively and standard deviation The numbering of the independent signal source components. The number for the sensitive feature category, This is a time variable. Since step 1 already performed Z-score standardization on each data point, and All are dimensionless standardized signals. The correlation coefficient is dimensionless, and we take... absolute value As a measure of waveform similarity.

[0033] Step 302: Process the independent signal source components separately. and template signal Fourier transforms are performed to extract the respective spectral feature parameters, including the center frequency, 3dB bandwidth, and dominant frequency energy percentage. The center frequency and 3dB bandwidth are both in Hz, and the dominant frequency energy percentage is a dimensionless percentage. To eliminate the influence of dimensional differences on vector operations, the above spectral feature parameters are subjected to mean normalization based on the range, scaling each parameter to a value that is within the range of the standard range. The interval is then used to construct a spectral feature vector from the normalized parameters. and The cosine similarity is used to calculate the spectral matching degree of two spectral feature vectors. : ; in, For vectors with vector The inner product, and They are vectors sum vector The model.

[0034] Step 303: Perform weighted fusion of waveform similarity and spectral matching to obtain independent signal source components. Relative to sensitive element categories Comprehensive sensitivity similarity : ; in, and Let be the weight coefficient, and satisfy... The weighting coefficients were obtained by calibration based on forward simulation data and measured verification data. Because and Both are dimensionless similarity measures, and their values ​​range from 1 to 10. Within the interval, and Weighted fusion calculations can be performed directly.

[0035] It should be noted that the aforementioned sensitive element feature database was generated as follows: Ground-penetrating radar responses of various sensitive elements were simulated using the finite-difference time-domain method. Simultaneously, statistical analysis of a large amount of measured data was combined to extract typical reflection waveform templates and spectral characteristic parameters for each sensitive element. Specifically, the signal characteristics of cavities are characterized by strong amplitude reflected waves with a phase opposite to the incident wave; the signal characteristics of loose regions are characterized by a chaotic group of reflected waves with enhanced amplitude but poor continuity; and the signal characteristics of water-rich bodies are characterized by reflected waves with extremely strong amplitude and a phase identical to the incident wave, with significant attenuation of high-frequency components.

[0036] In this embodiment of the application, in order to obtain a more comprehensive joint time and frequency feature representation, the calculation of comprehensive sensitivity similarity in step 303 also incorporates time-frequency distribution similarity. Specifically, the independent signal source components are obtained separately using short-time Fourier transform. and template signal Given two time-frequency distribution matrices, calculate the structural similarity index between them. By incorporating time-frequency distribution similarity into the weighted fusion, the formula for calculating comprehensive sensitivity similarity is updated as follows: ; in, , and Let be the weight coefficient, and satisfy... Structural similarity index The structural similarity index measures the overall similarity of two time-frequency distribution matrices across three dimensions: brightness, contrast, and structure. The calculation method is as follows: assuming independent signal source components The time-frequency distribution matrix is Template signal The time-frequency distribution matrix is ,but: ; in, and They are respectively and The mean of all elements, and They are respectively and The variance of all elements, for and Covariance of all corresponding elements and To prevent the stability constant from having a denominator of zero.

[0037] Furthermore, the aforementioned stability constant and The value is determined as follows: Let the maximum value of the dynamic range of the elements of the time-frequency distribution matrix be... ,but , ,in and The proportionality constant is dimensionless, and its values ​​are as follows: and , Consistent with the dimensions of the elements of the time-frequency distribution matrix, and The dimensions are The square of the value is kept consistent with the dimensions of the product of the means and variance terms in the numerator and denominator of the formula to ensure the structural similarity index. It is a dimensionless similarity measure, with values ​​ranging from... Within the range.

[0038] because , and All three are dimensionless similarity measures, and their values ​​range from [value range missing]. Within the interval, , and The three items can be directly weighted and fused. Time-frequency distribution similarity can capture the joint time and frequency variation characteristics of non-stationary signals, and can be used to help identify features such as high-frequency attenuation and multi-layer reflection of cavities in water-rich areas.

[0039] Step 4: Calculate the posterior probability of each independent signal source component based on Bayesian inference, determine the sensitivity level according to the adaptive threshold, and output the sensitive element extraction results. For each independent signal source component, based on the comprehensive sensitivity similarity of each independent signal source component and the prior information of the current detection scene, Bayesian inference is used to calculate the posterior probability of each independent signal source component belonging to each sensitive element category, and a cost-sensitive adaptive threshold is used for determination, outputting the sensitive element extraction result. This step includes the following sub-steps: Step 401: Obtain the prior probability of each sensitive element category in the current detection scenario. Prior probability The determination is based on geological conditions, historical collapse records, and current season information of the detection area. Simultaneously, the likelihood function, pre-obtained through statistical analysis of forward simulation data and historical measured data, is acquired. and ,in In order to be sensitive elements Given the existence of this condition, the overall sensitivity similarity value is taken as: The probability density, In order to be sensitive elements In the absence of a condition, the comprehensive sensitivity similarity value is taken as: The probability density.

[0040] Furthermore, the above likelihood function and All of these are pre-stored in the system parameter library in the form of probability density functions. Describe when sensitive elements are indeed present in the detection area. At that time, comprehensive sensitivity similarity exist Distribution pattern over intervals High probability density regions are concentrated in A larger range; Describes the situation when there are no sensitive elements in the detection area. At that time, comprehensive sensitivity similarity The distribution pattern, High probability density regions are concentrated in A smaller interval. Likelihood function. and All of these are obtained through statistical fitting of forward simulation data and historical measured data, and are directly called as input parameters for Bayesian inference in step 402.

[0041] Furthermore, the above likelihood function and The statistical fitting process is as follows: based on samples with known labeled categories in the forward simulation data and historical measured data, statistics are statistically analyzed for sensitive elements. Under both the existence and non-existence of the condition, the comprehensive sensitivity similarity The distribution of values ​​is determined, and a probability density function is fitted to the distribution of values ​​under each condition. The fitting results are stored in the system parameter library in the form of a lookup table or a parameterized function, for use in step 402 based on the current condition. The probability density corresponding to the value can be queried.

[0042] Step 402: Calculate the independent signal source components according to Bayes' theorem. Belongs to the category of sensitive elements posterior probability : ; Step 403: Based on the cost of missed detection and the cost of false alarms Calculate the adaptive decision threshold : ; in, The price to pay for classifying real hidden dangers as normal. The price of classifying normal areas as potential hazards. and All are dimensionless relative cost weights. and All values ​​are positive, and the threshold is adaptively determined. The range of values ​​is , and posterior probability Since their dimensions are consistent, they can be directly compared and judged. When When this occurs, the independent signal source component is determined to be a sensitive element category. Corresponding sensitive elements; when When this occurs, the independent signal source component is determined to be a non-sensitive element.

[0043] Step 404: For independent signal source components identified as sensitive elements, classify their sensitivity levels based on the range of their posterior probabilities: when When, it is classified as a high-sensitivity level; when When, it is classified as medium sensitivity; when At that time, it was classified as a low sensitivity level. The sensitivity element category and sensitivity level corresponding to each independent signal source component were output as the sensitivity element extraction result.

[0044] Furthermore, in the above-mentioned sensitivity level classification, the boundary for determining the high sensitivity level... and the lower boundary of the sensitivity level All are posterior probabilities The fixed quantile threshold, in its physical sense, means that the posterior probability is not less than a certain threshold. This indicates that, given the current prior information and comprehensive sensitive similarity evidence, the confidence level of this signal source component belonging to a sensitive element is high; the posterior probability is between and The range indicates a moderate level of confidence; the posterior probability is between the adaptive decision threshold. and The range indicates a relatively low confidence level, but still exceeds the adaptive decision threshold. These should be included in the low-sensitivity level for attention. The above-mentioned fixed quantile threshold and adaptive judgment threshold... Together, they form the complete boundary for the three-level sensitivity classification, where the adaptive judgment threshold... It changes dynamically with the detection scenario, and and The fixed gradation boundary point remains unchanged, with fixed quantile threshold and adaptive judgment threshold. All are at the same level as the posterior probability. Intervals allow for direct numerical comparisons.

[0045] It should be noted that the above prior probabilities The dynamic adjustment rules are as follows: In old urban areas with dense pipelines, the prior probability of the water-bearing anomaly category is increased due to the higher probability of pipeline leakage; in newly built roads or areas that have undergone grouting reinforcement, the prior probability of the void category is decreased; and during the rainy season or seasons with large fluctuations in groundwater levels, the prior probability of the water-rich body category is increased. These adjustment rules are obtained based on an expert knowledge base or statistical learning based on historical monitoring data.

[0046] It should be noted that the aforementioned costs of missed detection and the cost of false alarms The value is determined based on the engineering risk level of the object being tested. In high-risk scenarios such as urban main roads or roads above subway stations, The value is significantly greater than This enables adaptive decision thresholding. Reduced risk, thereby improving detection sensitivity; in low-risk scenarios, and The ratio decreases relatively, making the adaptive decision threshold more suitable. Increase the level appropriately to reduce the false alarm rate.

[0047] Furthermore, to provide more complete spatial information on sensitive elements to assist subsequent hazard mitigation decisions, in addition to step 404, the following steps are also included: spatially locating and scale estimating each independent signal source component identified as a sensitive element. The depth positioning method is based on the two-way travel time of the echo signal corresponding to each independent signal source component. Combined with the average electromagnetic wave propagation speed of the medium in this region Calculate the target depth : ; in, The unit is m / s. The unit is s. The unit is meters (m), and all terms in the formula have consistent dimensions. The horizontal position is determined by: obtaining the horizontal coordinates of the sensitive element based on the location of each independent signal source component in each ground-penetrating radar (GPR) data channel, combined with the radar scan mileage information. The scale is estimated by: estimating the thickness and lateral extent of cavities, loose zones, or water-rich bodies based on the duration and amplitude envelope width of the reflected signal, combined with spectral analysis results. The sensitive element category, sensitivity level, spatial location, and scale information are combined to output the complete sensitive element extraction result.

[0048] Example 2 An intelligent extraction system for sensitive elements of urban road collapse hazards includes the following components: The signal preprocessing module is used to sequentially perform zero-bias correction, bandpass filtering, time-varying gain compensation, and signal standardization on the raw echo data acquired by the ground penetrating radar to obtain standardized signal data. The blind source separation module is used to perform wavelet packet decomposition on standardized signal data to obtain the wavelet packet node coefficient matrix. The wavelet packet node coefficient matrix is ​​then used as multi-channel observation data to input the independent component analysis algorithm for demixing to obtain multiple independent components. Based on the correlation coefficient between each independent component and the preset reference signal, the independent components are sorted and sign-corrected to obtain multiple independent signal source components. The sensitivity similarity calculation module is used to obtain the template signals corresponding to each sensitive element category in the sensitive element feature library, calculate the waveform similarity and spectrum matching degree between each independent signal source component and each category of template signal, and perform weighted fusion of waveform similarity and spectrum matching degree to obtain the comprehensive sensitivity similarity of each independent signal source component relative to each sensitive element category. The sensitive element determination module is used to calculate the posterior probability of each independent signal source component belonging to each sensitive element category based on the comprehensive sensitivity similarity and prior information of the current detection scenario using Bayesian inference. It calculates an adaptive determination threshold based on the cost of missed detection and the cost of false alarm, compares the posterior probability with the adaptive determination threshold to determine whether each independent signal source component is a sensitive element, classifies the sensitivity level of the independent signal source components determined to be sensitive elements, and outputs the sensitive element extraction results.

[0049] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0050] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0051] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent extraction of sensitive elements for potential urban road collapse hazards, characterized in that, include: The raw echo data acquired by the ground penetrating radar is sequentially subjected to zero-bias correction, bandpass filtering, time-varying gain compensation, and signal standardization to obtain standardized signal data. Blind source separation processing is performed on the standardized signal data to obtain multiple independent signal source components; The blind source separation process includes: performing wavelet packet decomposition on each channel of data to obtain a wavelet packet node coefficient matrix; inputting the wavelet packet node coefficient matrix as multi-channel observation data into an independent component analysis algorithm for unmixing to obtain multiple independent components; and sorting and sign-correcting each independent component according to the correlation coefficient between each independent component and a preset reference signal. The template signals corresponding to each sensitive element category in the sensitive element feature library are obtained. The waveform similarity and spectrum matching degree between each independent signal source component and the template signal of each category are calculated respectively. The waveform similarity and spectrum matching degree are weighted and fused to obtain the comprehensive sensitivity similarity of each independent signal source component relative to each sensitive element category. Based on the comprehensive sensitivity similarity and prior information of the current detection scenario, Bayesian inference is used to calculate the posterior probability of each independent signal source component belonging to each sensitive element category. An adaptive judgment threshold is calculated based on the cost of missed detection and the cost of false alarm. The posterior probability is compared with the adaptive judgment threshold to determine whether each independent signal source component is a sensitive element. The sensitivity level of the independent signal source components determined to be sensitive elements is classified, and the sensitive element extraction result is output.

2. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, The bandpass filtering process includes a direct wave removal operation, which is performed by calculating the average background signal of multiple channels and subtracting the average background signal from each channel to weaken the direct wave component; the time-varying gain compensation uses an exponential gain function to perform time-varying amplitude compensation on the signal, so that the reflected signals at different depths have comparable amplitude levels; the signal standardization uses Z-score standardization.

3. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, The independent component analysis algorithm is the FastICA algorithm. The FastICA algorithm takes the multi-channel observation data as input, uses negative entropy to approximate the non-Gaussianity of the signal, and iteratively updates each row vector of the separation matrix by maximizing the negative entropy objective function. In each iteration, the row vector is updated using the first and second derivatives of the nonlinear function. After the update, the row vector is normalized and orthogonalized to ensure the statistical independence between the components. The iteration stops when the change in the row vector between two adjacent iterations is lower than the convergence criterion, and the separated signal matrix is ​​output as multiple independent components.

4. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, After obtaining the multiple independent components, the method further includes: evaluating the non-stationarity of each independent component by calculating the instantaneous frequency variance of each independent component within a sliding time window; when the instantaneous frequency variance of an independent component is greater than a preset non-stationarity threshold, the independent component is input into a variational mode decomposition algorithm for fine decomposition. The variational mode decomposition algorithm adaptively decomposes the independent component into several intrinsic mode functions, calculates the sample entropy of each intrinsic mode function, retains the intrinsic mode functions with sample entropy lower than a preset entropy threshold as effective signal components, removes the intrinsic mode functions with sample entropy higher than the preset entropy threshold as noise components, replaces the original independent components with the retained effective signal components, and updates the multiple independent signal source components.

5. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, The waveform similarity is calculated as follows: the waveform correlation coefficient between the independent signal source component and the template signal is calculated, and the absolute value of the waveform correlation coefficient is taken as the waveform similarity measure. The waveform correlation coefficient is the product of the covariance of the independent signal source component and the template signal divided by their standard deviations. The spectral matching degree is calculated as follows: Fourier transform is performed on the independent signal source component and the template signal respectively to extract their respective spectral feature parameters. The spectral feature parameters include the center frequency, 3dB bandwidth and the main frequency energy ratio. The spectral feature parameters are then normalized by mean based on the range to form spectral feature vectors. The spectral matching degree of the two spectral feature vectors is calculated using cosine similarity.

6. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, The calculation of comprehensive sensitivity similarity also incorporates time-frequency distribution similarity. The time-frequency distribution similarity is obtained by using short-time Fourier transform to obtain the time-frequency distribution matrices of the independent signal source components and the template signal, respectively, and calculating the structural similarity index between the two time-frequency distribution matrices. The structural similarity index measures the comprehensive similarity of the two time-frequency distribution matrices in three dimensions: brightness, contrast, and structure. It is calculated by calculating the mean, variance, and covariance of all corresponding elements in the two time-frequency distribution matrices, respectively. A numerator is constructed based on the product of the means and covariance of the two matrices, and a denominator is constructed based on the sum of the squares and variances of the means of the two matrices. Stability constants are added to both the numerator and denominator to prevent the denominator from being zero. The ratio of the numerator to the denominator is used as the structural similarity index. The comprehensive sensitivity similarity is a weighted fusion result of the waveform similarity, the spectral matching degree, and the time-frequency distribution similarity, with the sum of the weighting coefficients being one.

7. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, The calculation of the posterior probability of each independent signal source component belonging to each sensitive element category using Bayesian inference includes: obtaining the prior probability of each sensitive element category in the current detection scenario, wherein the prior probability is determined based on the geological conditions, historical collapse records, and current seasonal information of the detection area; obtaining the likelihood function obtained in advance based on statistical analysis of forward modeling data and historical measured data, wherein the likelihood function includes the probability density of comprehensive sensitivity similarity under the condition that sensitive elements exist and the probability density of comprehensive sensitivity similarity under the condition that sensitive elements do not exist; and, according to Bayes' theorem, dividing the product of the likelihood function and the prior probability by the marginal probability to obtain the posterior probability of each independent signal source component belonging to each sensitive element category.

8. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, The adaptive judgment threshold is calculated as follows: the false alarm cost is divided by the sum of the false alarm cost and the missed detection cost to obtain the adaptive judgment threshold; wherein, the missed detection cost is the cost of judging a real hidden danger as normal, the false alarm cost is the cost of judging a normal area as a hidden danger, and the values ​​of the missed detection cost and the false alarm cost are determined according to the engineering risk level of the detected object; the sensitivity level is divided as follows: when the posterior probability is not lower than the first fixed quantile threshold, it is classified as high sensitivity level; when the posterior probability is not lower than the second fixed quantile threshold and lower than the first fixed quantile threshold, it is classified as medium sensitivity level; when the posterior probability is not lower than the adaptive judgment threshold and lower than the second fixed quantile threshold, it is classified as low sensitivity level.

9. The intelligent extraction method for sensitive elements of urban road collapse hazards according to claim 1, characterized in that, Before outputting the results of sensitive feature extraction, the following is also included: Spatial positioning and scale estimation are performed on each independent signal source component identified as a sensitive element. Spatial positioning includes depth positioning and horizontal position determination. Depth positioning is achieved by multiplying the two-way travel time of the echo signal corresponding to each independent signal source component by the average electromagnetic wave propagation speed of the regional medium, and then taking half of the product of the average electromagnetic wave propagation speed and the two-way travel time as the target depth. Horizontal position determination is achieved by obtaining the horizontal coordinates of the sensitive element based on the location of each independent signal source component in each ground-penetrating radar data channel, combined with the radar scan mileage information. Scale estimation is performed by estimating the thickness and lateral distribution range of the sensitive element based on the duration and amplitude envelope width of the reflected signal, combined with spectral analysis results. The sensitive element category, sensitivity level, spatial position, and scale information are combined and output as a complete sensitive element extraction result.

10. An intelligent extraction system for sensitive elements of urban road collapse hazards, and a method for intelligent extraction of sensitive elements of urban road collapse hazards according to any one of claims 1-9, characterized in that, include: The signal preprocessing module is used to sequentially perform zero-bias correction, bandpass filtering, time-varying gain compensation, and signal standardization on the raw echo data acquired by the ground penetrating radar to obtain standardized signal data. The blind source separation module is used to perform wavelet packet decomposition on the standardized signal data to obtain the wavelet packet node coefficient matrix, input the wavelet packet node coefficient matrix as multi-channel observation data into the independent component analysis algorithm for demixing to obtain multiple independent components, and sort and sign-correct each independent component according to the correlation coefficient between each independent component and the preset reference signal to obtain multiple independent signal source components. The sensitivity similarity calculation module is used to obtain the template signal corresponding to each sensitive element category in the sensitive element feature library, calculate the waveform similarity and spectrum matching degree between each independent signal source component and the template signal of each category, and perform weighted fusion on the waveform similarity and spectrum matching degree to obtain the comprehensive sensitivity similarity of each independent signal source component relative to each sensitive element category. The sensitive element determination module is used to calculate the posterior probability of each independent signal source component belonging to each sensitive element category based on the comprehensive sensitivity similarity and prior information of the current detection scene using Bayesian inference, calculate an adaptive determination threshold based on the cost of missed detection and the cost of false alarm, compare the posterior probability with the adaptive determination threshold to determine whether each independent signal source component is a sensitive element, classify the sensitivity level of the independent signal source components determined to be sensitive elements, and output the sensitive element extraction results.

Images