Quantitative detection method and system for internal defects of concrete based on reflected wave

Abstract

The application discloses a quantitative detection method and system for internal defects of concrete based on reflected waves, and belongs to the technical field of nondestructive testing. A reflected wave data matrix is obtained through multi-angle excitation and synchronous reception; energy features of each channel are calculated to construct an energy response residual field; a disturbance comparison algorithm is used to extract waveform offset, spectrum jitter and phase change caused by defects to form a disturbance feature vector set; a defect-response function surface is further constructed, and defect topological structures are inverted based on gradient and curvature analysis to output defect geometric parameters; finally, defect parameters at multiple time points are input into a recurrent neural network to predict defect evolution paths and failure risks; the application has high resolution and trend prediction capability, and is suitable for concrete structure detection and early warning in bridges, tunnels and nuclear power engineering.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to a method and system for quantitative detection of internal defects in concrete based on reflected waves. Background Technology

[0002] Traditional methods for detecting internal defects in concrete mainly rely on manual tapping, ultrasonic penetration, and infrared imaging. However, these technologies suffer from limitations when dealing with complex conditions such as deeply buried defects, micro-cracks, honeycomb-like surfaces, and interfacial debonding. These limitations include weak penetration, strong waveform interference, low quantitative accuracy, and poor repeatability. Especially in high-safety-level projects such as urban rail transit, nuclear power plant foundations, and bridge bearings, the presence of millimeter-level voids or delamination defects within the concrete can easily evolve into structural disasters under long-term loads and wet-dry cycles, posing a significant challenge to early quantitative detection.

[0003] Existing methods struggle to achieve high-resolution extraction and quantitative assessment of low-reflectivity defects such as millimeter-scale bubbles and debonding layers. The main reasons are: first, the defect scale is much smaller than the wavelength, resulting in weak echo signal intensity that is easily masked by the texture scattering of the concrete matrix; second, the defect distribution is irregular and random, making it difficult for traditional grid scanning methods to match reflection paths, which can easily lead to misjudgment or omission; and third, there is a lack of feature fusion and evolution modeling for multi-scale reflection signals, making it impossible to predict the trend of early-stage hidden defects. Summary of the Invention

[0004] The purpose of this invention is to provide a quantitative detection method and system for internal defects in concrete based on reflected waves, so as to overcome the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a quantitative detection method for internal defects in concrete based on reflected waves, comprising:

[0006] Multi-angle excitation and synchronous reception operations are performed on the concrete component under test to acquire reflected wave signals at each measuring point under different incident angles and frequencies, and a reflected wave data matrix is ​​constructed.

[0007] For the reflected wave data matrix, the time-domain energy integral, frequency-domain power spectral density and propagation path response value of the reflected wave in each channel are calculated. Based on the energy difference between the reflected wave and the reference defect-free body model, the energy response residual field matrix is ​​constructed.

[0008] The perturbation comparison algorithm is used to dynamically align the energy response residual field with the standard defect-free waveform feature set, and extract the local perturbation response caused by the defect region, including waveform offset, spectral jitter amplitude and phase nonlinear change, to form a perturbation comparison feature vector set.

[0009] The perturbation contrast feature vector set is mapped to the spatial response coordinate system, and a ternary function model is constructed based on the perturbation amplitude, path residual and propagation direction cosine to generate the defect-response function surface.

[0010] Based on the gradient change and boundary curvature distribution of the defect-response function surface in space, the topological structure of internal defects in concrete is inverted, and a set of defect geometric parameters is output, including the defect center coordinates, outer contour shape, thickness distribution and volume estimate.

[0011] The defect geometric parameter set acquired at continuous time points is input into a time series prediction model based on a recurrent neural network to construct a defect evolution vector field and predict the expansion path, rate, and possible failure area of ​​the defect under load conditions or environmental disturbances.

[0012] Preferably, constructing the reflected wave data matrix includes:

[0013] All waveform data under the excitation-reception combination are indexed in three dimensions according to the incident angle, the receiving channel number, and the time dimension.

[0014] Denoising preprocessing is performed on each reflected wave signal;

[0015] Unify the waveform amplitude of each channel to achieve amplitude normalization.

[0016] The processed waveform signals are arranged in chronological order to form a three-dimensional reflected wave data matrix A(i,j,t), where i is the excitation angle index, j is the receiving channel index, and t is the time sampling point.

[0017] Preferably, the construction of the energy response residual field matrix includes:

[0018] The standardized energy differences of each channel are arranged in three dimensions according to the excitation source number, the receiving channel number, and the frequency index to generate an initial energy difference matrix;

[0019] Interpolation algorithms are used to spatially compensate for missing channel data;

[0020] A Gaussian smoothing kernel function is used to perform local weighted filtering on the matrix to form a continuous and smooth energy response distribution field.

[0021] The smoothed three-dimensional energy difference field is normalized and converted into an energy response residual field matrix.

[0022] Preferably, the perturbation comparison algorithm includes:

[0023] The reflection waveform corresponding to each channel in the energy response residual field matrix is ​​resampled and normalized to make it consistent with the standard defect-free waveform in terms of amplitude range and time axis length.

[0024] The waveform curve corresponding to the excitation-receive channel number in the standard defect-free waveform feature set is selected as the reference curve.

[0025] The sliding window correlation matching method is used to compare the similarity of the measured waveform and the reference waveform point by point and extract the local maximum similarity region.

[0026] Calculate the perturbation deviation value within the similar window as a local perturbation comparison response index.

[0027] Preferably, the dynamic alignment includes:

[0028] Based on the time window difference and the displacement of the main peak of the waveform, a local nonlinear time alignment function is constructed.

[0029] The measured waveform and the standard waveform are flexibly registered using a dynamic time warping algorithm to ensure that the main energy distribution and the peak structure coincide after alignment.

[0030] Set a maximum alignment error threshold; channels exceeding the threshold are considered abnormal and excluded.

[0031] Preferably, the generation of the defect-response function surface includes:

[0032] Extract three variables for each response point: the disturbance amplitude δr, the propagation path residual ΔR, and the propagation direction cosine μ.

[0033] A ternary fitting function F(δr,ΔR,μ) is established, and a weighted polynomial fitting model is adopted, wherein the weight coefficients are distributed proportionally according to the channel signal-to-noise ratio, and the fitting coefficients are solved by the least squares method.

[0034] The response value F(δr,ΔR,μ) output by the ternary function model is mapped to the corresponding node of the three-dimensional spatial mesh;

[0035] A cubic spline interpolation algorithm is used to reconstruct the surface of discrete response nodes, forming a continuous and smooth defect-response function surface;

[0036] The local gradient change rate and principal curvature distribution on the surface are calculated to identify regions of concentrated energy disturbance.

[0037] When the local gradient exceeds twice the average value, it is automatically identified as the center region of a potential defect, and the defect coordinates are output.

[0038] Preferably, the topological structure of the inverted internal defects of concrete includes:

[0039] The first-order spatial gradient values ​​are calculated for all nodes on the defect-response function surface, and the rate of change of response intensity is calculated in three directions using the finite difference method.

[0040] Set response gradient threshold It is twice the global average gradient of the function surface, and when the local gradient value exceeds... When this occurs, mark it as a potential defect boundary point;

[0041] Continuity constraint rules are used to cluster adjacent high gradient points to form a preliminary defect response region;

[0042] Morphological closing operations are performed on the clustering results to repair boundary gaps and generate an initial defect region outline that is spatially closed.

[0043] The second derivative is calculated on the boundary of the initially identified defect response region to determine the principal curvature value of each boundary point.

[0044] Identify the turning points of the defect morphology based on the trend of curvature value changes, and extract the outer contour structure of the defect;

[0045] The accuracy of defect identification is verified by the consistency between the degree of contour closure and the internal gradient distribution. If the condition is not met, the gradient threshold is automatically adjusted.

[0046] A three-dimensional α-shape reconstruction algorithm is used to encapsulate the defect response point set into a volume domain, forming a three-dimensional closed topology.

[0047] Preferably, the construction of the defect evolution vector field includes:

[0048] Defect detection is performed on the same concrete component at multiple consecutive time points to obtain the time series of defect geometric parameters, including defect center coordinates, volume, thickness distribution and boundary morphology characteristics;

[0049] The parameter set is arranged in chronological order into a multi-dimensional time series data matrix, and each feature dimension is normalized.

[0050] The preprocessed data is input into a recurrent neural network structure for training. The recurrent neural network is a two-layer long short-term memory network used to model the temporal evolution of defects.

[0051] The network output is the predicted value of the defect features at several future moments. Combined with the changes in their spatial coordinates, a three-dimensional defect evolution vector field is constructed.

[0052] Preferably, the predicted expansion path, rate, and potential failure region of the defect under load conditions or environmental disturbances include:

[0053] The first time-domain derivative of the defect center coordinate change sequence is calculated based on the predicted defect center coordinate change sequence to obtain the defect space expansion velocity vector.

[0054] The morphological differences of the defect contour region at each prediction time are transformed into a spatial expansion rate index, and its propagation trend in different directions is output.

[0055] The predicted defect center migration trajectory is fitted into a three-dimensional curve, and its potential evolution path deviation is analyzed in combination with the load direction and boundary constraints of the concrete structure.

[0056] Structural stress field distribution data are superimposed on the defect propagation path to identify areas where the defect path overlaps with high-stress areas and mark them as potential failure zones.

[0057] This invention also provides a quantitative detection system for internal defects in concrete based on reflected waves, comprising:

[0058] The reflected wave data acquisition module performs multi-angle excitation and synchronous reception operations on the concrete component under test, acquires reflected wave signals at each measuring point under different incident angles and frequencies, and constructs a reflected wave data matrix.

[0059] The residual modeling module calculates the time-domain energy integral, frequency-domain power spectral density, and propagation path response of the reflected waves for each channel in the reflected wave data matrix. Based on the energy difference between the reflected wave data matrix and the reference defect-free body model, it constructs the energy response residual field matrix.

[0060] The feature extraction module uses a perturbation comparison algorithm to dynamically align the energy response residual field with the standard defect-free waveform feature set, extracting the local perturbation response caused by the defect region, including waveform offset, spectral jitter amplitude and phase nonlinear change, to form a perturbation comparison feature vector set.

[0061] The defect response modeling module maps the disturbance comparison feature vector set to the spatial response coordinate system, constructs a ternary function model based on the disturbance amplitude, path residual and propagation direction cosine, and generates the defect-response function surface.

[0062] The geometric parameter extraction module, based on the gradient change and boundary curvature distribution of the defect-response function surface in space, inverts the topological structure of internal defects in concrete and outputs a set of defect geometric parameters, including defect center coordinates, outer contour shape, thickness distribution and volume estimate.

[0063] The defect evolution prediction module inputs the defect geometric parameter set acquired at continuous time points into a time series prediction model based on a recurrent neural network to construct a defect evolution vector field and predict the expansion path, rate, and possible failure area of ​​the defect under load conditions or environmental disturbances.

[0064] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0065] 1. This invention establishes a complete detection and analysis system suitable for identifying and predicting the evolution of minute defects inside concrete structures by integrating multi-angle reflected wave data acquisition, disturbance contrast feature extraction, response function modeling, and recurrent neural network prediction techniques. Compared with existing traditional methods that rely solely on single detection, two-dimensional image reconstruction, or empirical threshold judgment, this invention can construct a continuous energy response residual field and defect-response function surface, achieving high-precision positioning, morphological reconstruction, and three-dimensional topological modeling of millimeter-level defects under complex disturbance conditions, significantly improving the spatial resolution, quantitative capability, and reliability of detection.

[0066] 2. This invention introduces for the first time a time series prediction model based on long short-term memory networks to model the evolution of defect geometric parameters, constructing a defect evolution vector field to support the prediction of the expansion trend, speed, and possible failure paths of defects under load conditions and environmental disturbances. This mechanism breaks through the limitations of traditional static evaluation models, providing forward-looking and intelligent structural health monitoring for key projects such as bridges, tunnels, hydraulic structures, and nuclear power plant foundations, and has good engineering adaptability. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0068] Figure 1 This is a flowchart of the method of the present invention.

[0069] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Example 1, please refer to Figure 1 As shown in this embodiment, the quantitative detection method for internal defects in concrete based on reflected waves includes:

[0072] Multi-angle excitation and synchronous reception operations are performed on the concrete component under test to acquire reflected wave signals at each measuring point under different incident angles and frequencies, and a reflected wave data matrix is ​​constructed.

[0073] For the reflected wave data matrix, the time-domain energy integral, frequency-domain power spectral density and propagation path response value of the reflected wave in each channel are calculated. Based on the energy difference between the reflected wave and the reference defect-free body model, the energy response residual field matrix is ​​constructed.

[0074] The perturbation comparison algorithm is used to dynamically align the energy response residual field with the standard defect-free waveform feature set, and extract the local perturbation response caused by the defect region, including waveform offset, spectral jitter amplitude and phase nonlinear change, to form a perturbation comparison feature vector set.

[0075] The perturbation contrast feature vector set is mapped to the spatial response coordinate system, and a ternary function model is constructed based on the perturbation amplitude, path residual and propagation direction cosine to generate the defect-response function surface.

[0076] Based on the gradient change and boundary curvature distribution of the defect-response function surface in space, the topological structure of internal defects in concrete is inverted, and a set of defect geometric parameters is output, including the defect center coordinates, outer contour shape, thickness distribution and volume estimate.

[0077] The defect geometric parameter set acquired at continuous time points is input into a time series prediction model based on a recurrent neural network to construct a defect evolution vector field and predict the expansion path, rate, and possible failure area of ​​the defect under load conditions or environmental disturbances.

[0078] In this invention, multiple excitation sources and receiving channels are first arranged on the surface of the concrete component to be tested. The excitation source is an ultra-wideband pulse wave transmitter with an adjustable center frequency and a frequency band covering 20kHz to 500kHz, designed to meet both deep penetration and surface resolution requirements. The receiver is a high-sensitivity piezoelectric sensor with wideband response characteristics.

[0079] To achieve multi-angle incidence, the exciter array is arranged at equal angular intervals around the target area, forming at least eight excitation azimuth angles. In each detection, only one excitation channel is activated, while the other channels act as passive receiving nodes to construct the spatial distribution response of the reflected wave under different excitation-receiver paths.

[0080] This invention employs a unified clock triggering mechanism. A trigger pulse is issued by the central control unit to activate a single excitation source, simultaneously activating the high-frequency data acquisition modules of all receivers, ensuring the time synchronization and waveform integrity of the received signals. The sampling frequency is set to 2MHz to meet the requirements for fine resolution of high-frequency reflected waves on the time axis.

[0081] To avoid the impact of time delay differences in multi-channel signal acquisition on waveform phase, this invention performs a time offset correction operation after signal acquisition is completed. Specifically:

[0082] Set up a synchronization reference channel;

[0083] All received waveforms are aligned using a zero-time calibration pulse;

[0084] If the offset exceeds 1 / 20 of the sampling period, i.e. 0.5 microseconds, then interpolation adjustment is performed.

[0085] This invention employs a programmable variable frequency signal source, applying narrowband pulse waves of different frequencies to each excitation channel to obtain material response behavior at multiple frequencies. The excitation frequencies are divided into three groups: a low-frequency band of 20–50 kHz for deep probing; a mid-frequency band of 80–150 kHz for penetrating medium-thickness concrete components; and a high-frequency band of 200–500 kHz to enhance the resolution of surface defect boundaries.

[0086] After each excitation process, the control system records the current excitation frequency parameters and excitation angle code for subsequent dimension calibration of reflected wave data.

[0087] All time-domain waveform data under excitation-receiver combinations will be organized according to the following ternary parameter indexing method: the azimuth angle of the excitation source is denoted as i; the receiver number is denoted as j; and the time sampling point is denoted as t. A three-dimensional data matrix A(i,j,t) is constructed in this way, where each element represents the amplitude of the reflected wave received at time point t under the combination of the i-th excitation direction and the j-th receiver.

[0088] Before constructing the matrix, all received raw signals must undergo the following preprocessing steps:

[0089] Bandpass filtering: Apply a bandpass filter to each group of waveforms to retain the spectrum within ±10% of the center frequency and remove environmental electromagnetic noise and mechanical interference;

[0090] Amplitude normalization: The maximum amplitude of each waveform is normalized to 1 to eliminate amplitude deviation caused by energy inconsistency;

[0091] Wavelet denoising: The Daubechies 6th order wavelet is used for three-level decomposition to remove high-frequency noise components and retain the main energy of the reflected wave.

[0092] In this invention, for the reflected wave data matrix, the time-domain energy integral, frequency-domain power spectral density and propagation path response value of the reflected wave of each channel are calculated, and the energy response residual field matrix is ​​constructed based on the energy difference between the reflected wave and the reference defect-free body model.

[0093] First, for each pair of excitation and receiving channels, the effective time period of the main echo signal is extracted from its corresponding reflected waveform. This time period is from the initial rise point of the waveform to the point where the main reflected energy decays to less than 5% of its initial value. This time window is used to limit the integration interval and eliminate subsequent multiple reflections or background noise interference. Based on this, the squares of the amplitudes of the reflected waveforms within this time period are summed to obtain the total time-domain energy value of the channel. This value reflects the total energy retained by the reflected wave during its propagation along this path and can directly characterize the degree of wave energy loss within the structure.

[0094] Secondly, to further enhance the analytical capability of the frequency dimension for defect response, a Fast Fourier Transform (FFT) is performed on the waveform signal of each channel to transform it into the frequency domain, obtaining the frequency amplitude spectrum. This spectrum is then squared to obtain the power spectral density curve. Subsequently, the frequency range is divided into three typical intervals: low frequency (20 to 80 kHz), mid frequency (80 to 200 kHz), and high frequency (200 to 500 kHz). The average power spectral density value within each interval is calculated to characterize the energy distribution characteristics of the signal in different frequency regions. Specifically, this invention introduces a "spectral density contrast factor" as a defect response criterion, which is obtained by comparing the high-frequency power spectral density with the low-frequency power spectral density. A higher value indicates a stronger local disturbance, typically associated with small-scale defects such as cracks and cavities.

[0095] In addition to the energy dimension, to reflect the impact of defects on the wave propagation path, this invention also introduces the calculation of the propagation path response value. The method is as follows: First, based on the spatial coordinates of the exciter and receiver, the theoretical propagation distance between the two points is calculated, and the concrete wave velocity is set as a standard value, for example, 4000 meters per second. Based on this path length and sound velocity, the arrival time of the reflected wave under ideal conditions is calculated. Then, the moment when the main echo first exceeds the background noise threshold is extracted from the measured waveform and defined as the measured arrival time. The two are compared, and their relative difference is calculated, i.e., the propagation response residual. If the propagation response value of a certain channel deviates significantly from the standard time, it usually indicates that there is wave velocity slowdown or multipath scattering in that path, suggesting a possible structural anomaly in that area.

[0096] After extracting the above three types of energy parameters, a one-to-one difference calculation is performed between them and the "reference defect-free body model". The reference model is established as follows: a standard specimen with the same material as the concrete component to be tested and no internal defects is prepared. A complete reflected wave test is performed under the same excitation and receiving setup, frequency range, and parameter settings as the formal test. The total energy in the time domain, the power spectral density in the frequency domain, and the propagation response value are calculated from the waveform data of the model sample. Through normalization, the reference parameters are formed into a standard template matrix for subsequent residual comparison.

[0097] The residuals are calculated as follows: for each channel, the differences between the measured time-domain energy and the reference energy, the measured frequency-domain spectral ratio factor and the reference spectral ratio factor, and the measured propagation response value and the reference propagation response value are calculated separately. To integrate the three types of residual information, the three residual components are weighted and combined. The weights can be set according to the actual detection sensitivity requirements. For example, the weight of the time-domain energy residual is set to 0.5, the frequency-domain spectral ratio factor residual is set to 0.3, and the path response residual is set to 0.2 to ensure that energy is the dominant feature for resolution.

[0098] The fused residual values ​​are reassembled into a two-dimensional matrix on a channel-by-channel basis, with coordinate axes constructed using the excitation source number and receiver number. Each element in the matrix represents the combined energy residual response of the corresponding channel. This matrix is ​​the initial form of the energy response residual field. To improve the spatial continuity and visualization quality of this field, bilinear interpolation is used to compensate for invalid or missing channel data. Subsequently, a two-dimensional Gaussian kernel smoothing filter is applied to make the gradient transition of outliers in the matrix smoother. The standard deviation σ of the Gaussian kernel can be set according to the geometric spacing between channels; a recommended value is 1.5 times the minimum distance between channels.

[0099] In this invention, the core idea of ​​the perturbation comparison algorithm is to compare the reflected wave energy response of each channel in the concrete component under test with the waveform sample of a standard defect-free component, channel by channel, one-to-one. Through dynamic time alignment and micro-perturbation signal feature extraction, the fine-grained waveform deviation characteristics induced by defects are quantified. Unlike traditional full waveform amplitude difference determination methods, this algorithm analyzes local perturbation behavior based on dynamic alignment, thus possessing stronger anti-interference capabilities and defect-specific identification capabilities.

[0100] The standard waveform feature set is a collection of reflected wave data obtained on standard defect-free concrete specimens using the same probe array, excitation frequency settings, and data sampling parameters. This dataset is constructed as follows: exciters and receivers are deployed on standard concrete blocks with the same material and dimensions as the test object, with excitation frequencies covering 20 kHz to 500 kHz; under defect-free conditions, reflected waveforms from each excitation-receiver channel combination are acquired and uniformly normalized in amplitude and time; median filtering and wavelet denoising are used to preprocess the waveforms, retaining the main reflected wave and adjacent disturbance segments; the processed waveforms are then stored in a multi-dimensional index table based on channel number, frequency parameters, and sampling index, serving as a standard defect-free waveform feature set for comparison.

[0101] Because there may be time-scale drift or phase misalignment between the measured reflected wave and the reference waveform, dynamic alignment is required before perturbation feature extraction. The specific steps are as follows: Normalize the measured reflected waveform and the standard waveform, ensuring their amplitude range is between 0 and 1, and uniformly set the waveform length to a fixed number of sampling points; select the time period containing the main reflected wave as the alignment window, with the window length set to 20% of the total waveform length, typically 100 to 200 microseconds; use the Dynamic Time Warping (DTW) algorithm to match the measured waveform and the standard waveform. DTW is an alignment algorithm that allows nonlinear time stretching / compression and can identify the time offset path in the main wave packet region; set the maximum alignment error threshold to 10%, meaning that if the required stretching ratio exceeds 10% of the original time axis, the data for that channel is marked as abnormal and excluded from modeling. After DTW alignment, ensure that the measured waveforms of all channels highly overlap with the reference waveform within the main wave peak and its neighborhood, thus providing a standard coordinate system for subsequent perturbation feature extraction.

[0102] The waveform offset primarily reflects the time delay deviation of the main peak of the reflected wave caused by the defect. The specific calculation method is as follows: First, calculate the cross-correlation function between the measured waveform and the reference waveform; find the number of hysteresis sampling points corresponding to the maximum value of the cross-correlation function, and use this as the main peak time offset; divide this offset by the window width where the main peak is located to obtain the standardized offset value; the resulting value is denoted as... The unit is a dimensionless relative time offset. This feature is particularly sensitive in defect regions such as bubbles and cavities that cause a local decrease in wave velocity.

[0103] Spectral jitter reflects the frequency instability of the reflected wave signal on the time axis. Its implementation includes: applying a short-time Fourier transform to the waveform with a window length of 128 sampling points, using a Hanning window function with a 50% overlap; calculating the spectral centroid (frequency centroid) within each time window to obtain the frequency trajectory changing over time; and calculating the sample standard deviation of this frequency trajectory as the spectral jitter amplitude, denoted as . Further normalization: using actual measurements Dividing by the standard deviation of the reference trajectory yields the perturbation factor. This characteristic is highly sensitive to frequency instability caused by multipath interference, microcrack scattering, and other factors.

[0104] Phase change reflects the nonlinear distortion of the phase structure during signal propagation, and is particularly suitable for identifying adhesion detachment and micro-delamination defects. The calculation steps are as follows: apply a Hilbert transform to the measured waveform to obtain the instantaneous phase; use a phase expansion algorithm to process jumps, forming a continuous phase sequence; calculate the second-order difference of this phase sequence to reflect its degree of nonlinear change; calculate the root mean square (RMS) value of the second-order difference sequence, denoted as . The ratio of this ratio to the RMS value of the defect-free reference channel is used to obtain the standardized phase perturbation index.

[0105] The above three features: standardized waveform offset Standardized Spectrum Jitter Amplitude and standardized phase nonlinear variation They are spliced ​​together in a fixed order to form channel-level perturbation contrast feature vectors. .

[0106] This invention defines a spatial response coordinate system for associating perturbation contrast feature vectors with their corresponding geometric and propagation properties. This coordinate system comprises three coordinate axes, as follows:

[0107] First axis X: Represents the azimuth angle of the excitation source. The unit is angle (°), which is determined by the position of the exciter on the concrete surface.

[0108] Second axis Y: Represents the azimuth angle of the receiver. The unit is angle (°), which is also set by the array arrangement;

[0109] The third axis Z: represents the normalized propagation depth along the wave propagation direction. This refers to the ratio of the vertical propagation path length of the reflected wave in the channel to the thickness of the component, ranging from 0 to 1.

[0110] The perturbation contrast feature vector extracted for each excitation-receiver channel combination Its corresponding spatial location parameters By binding, a point cloud set with attributes can be formed in the spatial response coordinate system. Indicates waveform offset. Indicates the amplitude of spectrum jitter. This indicates a nonlinear phase change, and all three values ​​are normalized values.

[0111] To simplify subsequent fitting and modeling, the perturbation response of each channel is weighted and fused to define a comprehensive perturbation amplitude. for: ;in, , , These are the weighting factors for the three types of perturbation features, respectively, whose sum is 1. A recommended value is... =0.5, =0.3, =0.2.

[0112] Before constructing the defect-response function surface, the perturbation response values ​​must first be established. The functional relationship between spatial propagation characteristics and the model is described. This invention employs a weighted polynomial fitting method to construct the following ternary function model: ;in: The amplitude of the combined disturbance is defined in the previous section; ΔR is the residual of the corresponding channel propagation path, defined as the difference between the actual propagation time and the theoretical propagation time, specifically... μ is the direction cosine, defined as the cosine of the angle between the wave propagation direction vector in space and the reference perpendicular direction; Let be the response intensity value at this point on the defect-response function surface, representing the degree of defect perturbation. The function model is a second-order polynomial fitting structure, constructed as follows:

[0113] ;in, to These are the fitting coefficients to be solved. The coefficient fitting method is weighted least squares, with the weighting coefficients set according to the signal-to-noise ratio of each channel signal, and higher signal-to-noise ratio channels are given higher fitting weights.

[0114] The fitting error is evaluated using the mean squared error (MSE) index. If the mean squared value of the fitting residual exceeds 0.05 (normalized energy unit), the local weighted regression algorithm (LOWESS) is used to refit the local region to ensure that the overall surface has smoothness and continuity.

[0115] After completing the function fitting model, the next step is to map the function values ​​back to the spatial response coordinate system to generate a continuous defect-response function surface. The main steps include:

[0116] For all sampling points in the spatial response coordinate system , and its corresponding ternary variables Inputting ΔR and μ into the above function model, calculate value;

[0117] A cubic spline interpolation algorithm is used to fit a spatial surface to all discrete points, generating a continuous three-dimensional function surface;

[0118] Calculate the local gradient value in the function surface, that is, the rate of change of the response intensity of adjacent grid points, as an indicator of the concentration of energy disturbance;

[0119] A gradient threshold Tg is set to be twice the average gradient of the surface. When the gradient value of a local area exceeds Tg, the area is automatically identified as a potential defect response concentration area, and its spatial coordinates are recorded. With response value This serves as the output result for defect location.

[0120] The final generated defect-response function surface is stored in three-dimensional matrix form, with the following data structure: X-axis : Excitation angle distribution, with a total of N exciters; Y-axis : Receiver angle distribution, a total of M receivers; Z-axis ( ): Propagation depth distribution, stratified by fixed sampling (e.g., 20 layers); Value range ( ): The defect response value of each node, in units of normalized disturbance amplitude. This surface can be visualized through contour plots, gradient color heatmaps, or 3D volume rendering, facilitating structural engineers to judge, locate, and classify the distribution of defects.

[0121] In this invention, the defect-response function surface is defined in a three-dimensional spatial response coordinate system, with the three axes being the excitation source angle, receiver angle, and normalized propagation depth, respectively, and the corresponding axes numbered x, y, and z. Function value This indicates the intensity of the disturbance response corresponding to that spatial point.

[0122] To identify potential defect regions, the first-order gradient of the function surface in space must first be calculated. The partial derivatives of the response value at each node are then calculated using the three-dimensional finite difference method.

[0123] The gradient in the x-direction is: ;

[0124] The gradient in the y-direction is: ;

[0125] The gradient in the z-direction is: .

[0126] Taking the square root of the sum of the squares of the three directional gradients yields the spatial gradient magnitude G_total at that point, i.e.: The present invention sets the gradient anomaly threshold Tg to twice the mean of G_total for the entire function surface. Any point where G_total exceeds Tg is considered an abnormal response region and is initially marked as a potential defect boundary point.

[0127] For the set of points marked as having high gradients, a spatial clustering algorithm based on connectivity constraints (such as 3D DBSCAN) is used for clustering. This algorithm divides adjacent high-gradient points into several independent regions based on spatial distance and neighborhood density. Each clustered region is a potential defect response cluster.

[0128] To extract the true geometric shape of the defect boundary, this invention further performs curvature analysis on the boundary surface of the defect response body. The specific method is as follows:

[0129] A set of three-dimensional contour points is extracted from the boundary of the closed voxel block, and a three-dimensional principal curvature calculation model is used to calculate the principal curvature for each point. and The curvature amplitude K is defined as the absolute value of the principal curvature sum, i.e. The surface curvature is measured using a curvature distribution map for all boundary points, identifying local curvature maxima as defect morphological inflection points. The reliability of curvature peaks is verified by combining gradient continuity between adjacent points, and missing points are automatically inserted for incompletely closed boundary regions. A 3D α-shape reconstruction algorithm is used to fit the envelope of all valid boundary points, forming a closed 3D defect outline. The selection of the α-shape algorithm parameter α is based on point cloud density estimation, and a setting of 1.5 times the average point distance is recommended to balance surface smoothness and morphological realism.

[0130] For example, when using the 3D α-shape reconstruction algorithm to fit the envelope of all valid boundary points, the 3D discrete point set extracted from the defect boundary must first be input into the α-shape construction module. The α parameter is set to 1.5 times the average spacing of the boundary point cloud to control the smoothness and detail preservation of the fitted envelope. Subsequently, a 3D simplex mesh of the boundary points is constructed based on Delaunay triangulation, and the outer envelope surface that exceeds the curvature constraint is filtered out according to the α parameter, thereby generating a closed 3D outer contour that conforms to the actual shape of the defect.

[0131] After completing the topology construction, the defect region is parametrically described, and a set of geometric parameters is output, including the following:

[0132] All response strengths exceeding a set threshold Within the set of spatial points, calculate its geometric center as the location of the defect center. It is usually set to 1.5 times the mean of the global response. The center coordinates (X_c, Y_c, Z_c) are calculated as follows: X_c = the average of all X coordinates; Y_c = the average of all Y coordinates; Z_c = the average of all Z coordinates; the resulting three-dimensional coordinates are the spatial location result of the defect center.

[0133] Projecting the defect boundary onto the xy plane, the lengths of the maximum extension axis (major axis) and the minimum extension axis (minor axis) are calculated using an ellipse fitting method, and denoted as […]. and If the fitting error exceeds 10%, a polygon fitting model is used for contour approximation.

[0134] The output results include: the principal axis direction of the defect in the plane; the equivalent ellipse shape parameters; and the boundary closure index (closure is defined as the ratio of the average distance between boundary points to the boundary perimeter; the closer to 1, the more regular the boundary).

[0135] Thickness distribution extraction includes: statistically analyzing the distribution of each response point level within the defect along the z-axis (i.e., the propagation depth direction). The number of points contained in each z-layer is calculated, and the defect thickness distribution curve is derived by combining the interlayer spacing. Outputs include: minimum defect thickness (the thinnest layer with a response); maximum defect thickness; and defect thickness distribution variance, used to determine whether it is a layered defect or a void-type defect.

[0136] Defect volume estimation is as follows: Divide the closed topology into equal-sized cubic elements (voxels), count the total number of elements within each voxel whose response intensity exceeds a threshold, and multiply this number by the single-unit volume to estimate the defect volume. If the side length of each voxel is Δl (unit: millimeters), then the volume V is: The output is in cubic millimeters or cubic centimeters.

[0137] In this invention, a periodic inspection method is adopted, performing multiple non-destructive tests on the same inspection area at different time points during the structural service life, and generating a complete set of defect geometric parameters at each time point, including: defect center coordinates (X_c, Y_c, Z_c); defect volume V; maximum defect thickness H_max; defect boundary contour features (principal axis length, perimeter, closure degree, etc.); and defect boundary morphology descriptors (such as average curvature, directional index, etc.). These parameters are arranged chronologically to form a multi-dimensional time series dataset in the form: D=P1,P2,...,Pn; where Pn represents the defect parameter vector at the nth time point. To ensure the consistency of numerical scales among different features, a min-max normalization method is used to linearly map each feature value to the interval [0,1], facilitating neural network processing.

[0138] This invention employs a recurrent neural network architecture based on Long Short-Term Memory (LSTM) to model the evolution trend of defects. The model structure is as follows:

[0139] Input layer: The input is the sequence of defect parameters from the past N time steps, where N is the model window length, and a value of 5 is recommended; Hidden layer: It consists of two cascaded LSTM units, with 128 nodes per layer, and the activation function is the hyperbolic tangent function (tanh), used to model the long-term dependencies between temporal features; Output layer: The output is the predicted vector of defect parameters for the next time step (or several time steps in the future), with the same dimension as the input parameter set.

[0140] This network is iteratively trained using the backpropagation algorithm and the Adam optimizer to learn the evolutionary patterns in the input data. The loss function is a weighted mean square error function, while also introducing a smoothing constraint term for spatial location changes to ensure the continuity and physical realizability of the predicted trajectory. Specifically, it takes the form: L_total = α × MSE(P_pred, P_true) + β × SmoothLoss(ΔP_pred); where MSE represents the standard mean square error, SmoothLoss represents the fluctuation penalty term for the rate of location change, and α and β are weighting coefficients, with recommended values ​​of 0.8 and 0.2, respectively.

[0141] After obtaining the predicted parameters for future defect moments, the time series of the defect center coordinates are constructed into a vector to form the defect evolution vector field. The specific steps are as follows:

[0142] Calculate the difference in defect center coordinates between two consecutive prediction time points to obtain spatial variation vectors ΔX, ΔY, and ΔZ; map each vector to a three-dimensional coordinate system and connect them according to the time series to form the defect evolution trajectory; calculate the defect center displacement per unit time at each prediction point, which is defined as the evolution velocity. The units are millimeters per day or millimeters per week; based on the changes in the defect profile morphology over time, the expansion direction and intensity are extracted, and a multi-point field distribution vector is further constructed to describe the changing trend of the defect boundary. The generated defect evolution vector field not only reflects the migration path of the defect center, but also reflects the expansion speed and spatial deformation trend of its boundary in each direction.

[0143] To achieve structural risk early warning, this invention superimposes the defect evolution path with the structural load response field for analysis to identify potential failure regions, specifically including:

[0144] Obtain the finite element simulation results or measured stress field distribution of the structure, and construct the spatial load intensity spectrum S(x,y,z);

[0145] The defect evolution trajectory is mapped in space and overlapped with the stress field. The intersection of the trajectory and the high-stress region (S>S_th) is calculated. Here, S_th represents the stress field failure threshold, which is determined based on the ultimate strength of the structural material. A typical value can be taken as 80% of the compressive strength of reinforced concrete.

[0146] If the evolution path crosses or approaches a region of high stress concentration, mark the location as a potential failure region, and record the expected arrival time and evolution rate.

[0147] By combining the boundary proximity and the rate of increase of defect volume, the failure risk level of the area is assessed and divided into three levels: low, medium, and high, for reference in engineering maintenance.

[0148] Example 2, please refer to Figure 2As shown in this embodiment, the quantitative detection system for internal defects in concrete based on reflected waves includes:

[0149] The reflected wave data acquisition module performs multi-angle excitation and synchronous reception operations on the concrete component under test, acquires reflected wave signals at each measuring point under different incident angles and frequencies, and constructs a reflected wave data matrix.

[0150] The residual modeling module calculates the time-domain energy integral, frequency-domain power spectral density, and propagation path response of the reflected waves for each channel in the reflected wave data matrix. Based on the energy difference between the reflected wave data matrix and the reference defect-free body model, it constructs the energy response residual field matrix.

[0151] The feature extraction module uses a perturbation comparison algorithm to dynamically align the energy response residual field with the standard defect-free waveform feature set, extracting the local perturbation response caused by the defect region, including waveform offset, spectral jitter amplitude and phase nonlinear change, to form a perturbation comparison feature vector set.

[0152] The defect response modeling module maps the disturbance comparison feature vector set to the spatial response coordinate system, constructs a ternary function model based on the disturbance amplitude, path residual and propagation direction cosine, and generates the defect-response function surface.

[0153] The geometric parameter extraction module, based on the gradient change and boundary curvature distribution of the defect-response function surface in space, inverts the topological structure of internal defects in concrete and outputs a set of defect geometric parameters, including defect center coordinates, outer contour shape, thickness distribution and volume estimate.

[0154] The defect evolution prediction module inputs the defect geometric parameter set acquired at continuous time points into a time series prediction model based on a recurrent neural network to construct a defect evolution vector field and predict the expansion path, rate, and possible failure area of ​​the defect under load conditions or environmental disturbances.

[0155] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. A quantitative detection method for internal defects in concrete based on reflected waves, characterized in that: include: Multi-angle excitation and synchronous reception operations are performed on the concrete component under test to acquire reflected wave signals at each measuring point under different incident angles and frequencies, and a reflected wave data matrix is ​​constructed. For the reflected wave data matrix, the time-domain energy integral, frequency-domain power spectral density and propagation path response value of the reflected wave in each channel are calculated. Based on the energy difference between the reflected wave and the reference defect-free body model, the energy response residual field matrix is ​​constructed. The perturbation comparison algorithm is used to dynamically align the energy response residual field with the standard defect-free waveform feature set, and extract the local perturbation response caused by the defect region, including waveform offset, spectral jitter amplitude and phase nonlinear change, to form a perturbation comparison feature vector set. The perturbation contrast feature vector set is mapped to the spatial response coordinate system, and a ternary function model is constructed based on the perturbation amplitude, path residual and propagation direction cosine to generate the defect-response function surface. Based on the gradient change and boundary curvature distribution of the defect-response function surface in space, the topological structure of internal defects in concrete is inverted, and a set of defect geometric parameters is output, including the defect center coordinates, outer contour shape, thickness distribution and volume estimate. The defect geometric parameter set acquired at continuous time points is input into a time series prediction model based on a recurrent neural network to construct a defect evolution vector field and predict the expansion path, rate, and possible failure area of ​​the defect under load conditions or environmental disturbances.

2. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 1, characterized in that: The construction of the reflected wave data matrix includes: All waveform data under the excitation-reception combination are indexed in three dimensions according to the incident angle, the receiving channel number, and the time dimension. Denoising preprocessing is performed on each reflected wave signal; Unify the waveform amplitude of each channel to achieve amplitude normalization. The processed waveform signals are arranged in chronological order to form a three-dimensional reflected wave data matrix A(i,j,t), where i is the excitation angle index, j is the receiving channel index, and t is the time sampling point.

3. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 1, characterized in that: The construction of the energy response residual field matrix includes: The standardized energy differences of each channel are arranged in three dimensions according to the excitation source number, the receiving channel number, and the frequency index to generate an initial energy difference matrix; Interpolation algorithms are used to spatially compensate for missing channel data; A Gaussian smoothing kernel function is used to perform local weighted filtering on the matrix to form a continuous and smooth energy response distribution field. The smoothed three-dimensional energy difference field is normalized and converted into an energy response residual field matrix.

4. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 1, characterized in that: The perturbation comparison algorithm includes: The reflection waveform corresponding to each channel in the energy response residual field matrix is ​​resampled and normalized to make it consistent with the standard defect-free waveform in terms of amplitude range and time axis length. The waveform curve corresponding to the excitation-receive channel number in the standard defect-free waveform feature set is selected as the reference curve. The sliding window correlation matching method is used to compare the similarity of the measured waveform and the reference waveform point by point and extract the local maximum similarity region. Calculate the perturbation deviation value within the similar window as a local perturbation comparison response index.

5. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 1, characterized in that: in, The dynamic alignment includes: Based on the time window difference and the displacement of the main peak of the waveform, a local nonlinear time alignment function is constructed. The measured waveform and the standard waveform are flexibly registered using a dynamic time warping algorithm to ensure that the main energy distribution and the peak structure coincide after alignment. Set a maximum alignment error threshold; channels exceeding the threshold are considered abnormal and excluded.

6. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 1, characterized in that: in, The generated defect-response function surface includes: Extract three variables for each response point: the disturbance amplitude δr, the propagation path residual ΔR, and the propagation direction cosine μ. A ternary fitting function F(δr,ΔR,μ) is established, and a weighted polynomial fitting model is adopted, wherein the weight coefficients are distributed proportionally according to the channel signal-to-noise ratio, and the fitting coefficients are solved by the least squares method. The response value F(δr,ΔR,μ) output by the ternary function model is mapped to the corresponding node of the three-dimensional spatial mesh; A cubic spline interpolation algorithm is used to reconstruct the surface of discrete response nodes, forming a continuous and smooth defect-response function surface; The local gradient change rate and principal curvature distribution on the surface are calculated to identify regions of concentrated energy disturbance. When the local gradient exceeds twice the average value, it is automatically identified as the center region of a potential defect, and the defect coordinates are output.

7. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 1, characterized in that: The inverted topological structure of internal defects in concrete includes: The first-order spatial gradient values ​​are calculated for all nodes on the defect-response function surface, and the rate of change of response intensity is calculated in three directions using the finite difference method. Set response gradient threshold It is twice the global average gradient of the function surface, and when the local gradient value exceeds... When this occurs, mark it as a potential defect boundary point; Continuity constraint rules are used to cluster adjacent high gradient points to form a preliminary defect response region; Morphological closing operations are performed on the clustering results to repair boundary gaps and generate an initial defect region outline that is spatially closed. The second derivative is calculated on the boundary of the initially identified defect response region to determine the principal curvature value of each boundary point. Identify the turning points of the defect morphology based on the trend of curvature value changes, and extract the outer contour structure of the defect; The accuracy of defect identification is verified by the consistency between the degree of contour closure and the internal gradient distribution. If the condition is not met, the gradient threshold is automatically adjusted. A three-dimensional α-shape reconstruction algorithm is used to encapsulate the defect response point set into a volume domain, forming a three-dimensional closed topology.

8. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 1, characterized in that: The construction of the defect evolution vector field includes: Defect detection is performed on the same concrete component at multiple consecutive time points to obtain the time series of defect geometric parameters, including defect center coordinates, volume, thickness distribution and boundary morphology characteristics; The parameter set is arranged in chronological order into a multi-dimensional time series data matrix, and each feature dimension is normalized. The preprocessed data is input into a recurrent neural network structure for training. The recurrent neural network is a two-layer long short-term memory network used to model the temporal evolution of defects. The network output is the predicted value of the defect features at several future moments. Combined with the changes in their spatial coordinates, a three-dimensional defect evolution vector field is constructed.

9. The quantitative detection method for internal defects in concrete based on reflected waves according to claim 8, characterized in that: The predicted defect propagation path, rate, and potential failure region under load conditions or environmental disturbances include: The first time-domain derivative of the defect center coordinate change sequence is calculated based on the predicted defect center coordinate change sequence to obtain the defect space expansion velocity vector. The morphological differences of the defect contour region at each prediction time are transformed into a spatial expansion rate index, and its propagation trend in different directions is output. The predicted defect center migration trajectory is fitted into a three-dimensional curve, and its potential evolution path deviation is analyzed in combination with the load direction and boundary constraints of the concrete structure. Structural stress field distribution data are superimposed on the defect propagation path to identify areas where the defect path overlaps with high-stress areas and mark them as potential failure zones.

10. A quantitative detection system for internal defects in concrete based on reflected waves, used to implement the quantitative detection method for internal defects in concrete based on reflected waves as described in any one of claims 1-9, characterized in that: include: The reflected wave data acquisition module performs multi-angle excitation and synchronous reception operations on the concrete component under test, acquires reflected wave signals at each measuring point under different incident angles and frequencies, and constructs a reflected wave data matrix. The residual modeling module calculates the time-domain energy integral, frequency-domain power spectral density, and propagation path response of the reflected waves for each channel in the reflected wave data matrix. Based on the energy difference between the reflected wave data matrix and the reference defect-free body model, it constructs the energy response residual field matrix. The feature extraction module uses a perturbation comparison algorithm to dynamically align the energy response residual field with the standard defect-free waveform feature set, extracting the local perturbation response caused by the defect region, including waveform offset, spectral jitter amplitude and phase nonlinear change, to form a perturbation comparison feature vector set. The defect response modeling module maps the disturbance comparison feature vector set to the spatial response coordinate system, constructs a ternary function model based on the disturbance amplitude, path residual and propagation direction cosine, and generates the defect-response function surface. The geometric parameter extraction module, based on the gradient change and boundary curvature distribution of the defect-response function surface in space, inverts the topological structure of internal defects in concrete and outputs a set of defect geometric parameters, including defect center coordinates, outer contour shape, thickness distribution and volume estimate. The defect evolution prediction module inputs the defect geometric parameter set acquired at continuous time points into a time series prediction model based on a recurrent neural network to construct a defect evolution vector field and predict the expansion path, rate, and possible failure area of ​​the defect under load conditions or environmental disturbances.

Images