A multi-frequency ultrasonic signal fusion defect quantitative detection method and system

Abstract

This invention discloses a method and system for quantitative detection of defects using multi-frequency ultrasonic signal fusion, relating to the field of intelligent signal processing technology. The method includes the following steps: acquiring multi-frequency aligned echoes of the structure under test, and determining the propagation delay function and path weight function based on a multi-dielectric layer propagation model; extracting defect residual signals based on the multi-frequency aligned echoes and defect-free reference echoes; determining the principal depth position of the defect through cross-frequency in-situ focusing intensity; performing cross-frequency alignment and weighted superposition on the defect residual signals to obtain a coherent reconstruction echo of the defect and constructing a quantitative feature vector of the defect; and combining a classification network, a physical mapping model, and a regression network to output quantitative detection results of the defect type, defect location, and defect size. This invention can improve the accuracy, stability, and quantification level of deep defect detection in multi-dielectric layer structures.

Description

Technical Field

[0001] This invention relates to the field of intelligent signal processing technology, and in particular to a method and system for quantitative detection of defects in multi-frequency ultrasonic signal fusion. Background Technology

[0002] As building facades increasingly adopt multi-media composite structures composed of metal panels, insulation layers, and concrete substrates, the need for detecting internal defects such as interlayer debonding, hollow areas, and cracks is becoming increasingly prominent. In these structures, the acoustic parameters of different media layers vary significantly. Ultrasonic waves are prone to inconsistent propagation delays, significant differences in energy attenuation, and overlapping echoes from multiple interfaces during propagation. Particularly in deeper regions, defect echoes are often masked by normal interlayer echoes and noise components, making it difficult to accurately pinpoint defect locations, reliably distinguish defect types, and reliably quantify defect sizes. Therefore, there is an urgent need for a method capable of accurately locating and quantitatively detecting defects in multi-media layer structures using multi-frequency ultrasonic signals.

[0003] Existing technologies, when addressing the aforementioned problems, typically focus on single-frequency ultrasonic detection or simple processing of multi-frequency echoes, making it difficult to simultaneously consider propagation adaptability and detection stability across different depth regions. When ultrasonic waves traverse multiple layers of media and propagate back and forth between multiple interfaces, normal interlayer responses, abnormal defect responses, and noise responses easily become mixed in the time and frequency domains, making it difficult to effectively separate deep defect echoes. Furthermore, existing solutions do not adequately consider the differences in propagation delay and energy loss of different frequency signals, resulting in the inability to effectively enhance cross-frequency information at the same depth location, thus affecting the accuracy of defect location determination. In addition, relying solely on a single amplitude index or empirical judgment for analysis can easily lead to unstable defect type identification and significant deviations in size assessment, failing to meet the accuracy, stability, and quantification requirements for defect detection in multi-media layered building exterior walls. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies, such as the difficulty in accurately locating defects, stably identifying defect types, and reliably quantifying defect sizes due to the influence of multi-interface echo superposition and propagation attenuation on deep defect echoes in multi-dielectric layer structures. Therefore, this invention proposes a multi-frequency ultrasonic signal fusion method and system for quantitative defect detection.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution: A method for quantitative detection of defects by multi-frequency ultrasonic signal fusion includes: S1. Obtain the multi-frequency aligned echo of the structure under test, and determine the propagation delay function and path weight function of the multi-frequency signal at different depths based on the multi-dielectric layer propagation model; S2. Extract the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo; S3. Based on the propagation delay function and the path weight function, calculate the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth, and determine the candidate depth corresponding to the maximum cross-frequency in-situ focusing intensity as the main depth position of the defect. S4. Based on the defect main depth position, the propagation delay function and the path weight function, the defect residual signal is cross-frequency aligned and weighted to obtain the defect coherent reconstruction echo, and a defect quantitative feature vector is constructed based on the maximum cross-frequency in-situ focusing intensity, the defect main depth position and the defect coherent reconstruction echo. S5. Input the quantitative feature vector of the defect into the classification network to determine the defect type. Based on the defect type and the depth information and multi-frequency response information in the quantitative feature vector of the defect, calculate the defect size based on the physical mapping model and regression network, and output the quantitative detection result of the defect including the defect type, the main depth position of the defect and the defect size.

[0006] Preferably, acquiring the multi-frequency aligned echo of the structure under test includes: The raw echo signals of the structure under test at multiple frequencies are acquired, and the raw echo signals are denoised to obtain the denoised signals corresponding to each frequency. A denoised signal corresponding to a reference frequency is selected as the benchmark, and the cross-correlation method is used to calculate the cross-frequency time offset of the denoised signals of other frequencies relative to the benchmark. The denoised signals of each frequency are time-shifted and calibrated according to the cross-frequency time offset to obtain the multi-frequency aligned echo.

[0007] Preferably, the propagation delay function and path weight function of the multi-frequency signal at different depths are determined based on the multi-medium layer propagation model, including: Based on the multi-medium layer propagation model, the propagation velocity, attenuation coefficient, and interfacial transmission coefficient of ultrasonic waves at each frequency in each medium layer are obtained. Based on the thickness of each medium layer and the propagation wave velocity, the two-way propagation delay of ultrasonic waves at each frequency at each candidate depth is calculated and used as the propagation delay function. Based on the thickness of each medium layer, the attenuation coefficient, and the interface transmission coefficient, the combined weight of attenuation and transmission of ultrasonic waves at each candidate depth at each frequency is calculated and used as the path weight function.

[0008] Preferably, extracting the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo includes: Constructing a defect-free multi-medium layer reference transfer function based on the transfer matrix method; The reference transfer function is combined with the excitation spectrum corresponding to the ultrasonic excitation to generate a defect-free reference echo corresponding to each frequency. The defect residual signal is extracted based on the difference between the multi-frequency aligned echo and the defect-free reference echo.

[0009] Preferably, calculating the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth includes: The defect residual signals at each frequency are subjected to analytical signal processing; For each candidate depth, based on the propagation delay function and the path weight function corresponding to the candidate depth, cross-frequency aggregation is performed on the analytical signals of each frequency to obtain the aggregated response corresponding to the candidate depth; Based on the aggregation response and the response intensity of each frequency analytical signal at the candidate depth, the cross-frequency in-situ focusing intensity corresponding to the candidate depth is calculated; The candidate depth with the largest cross-frequency in-situ focusing intensity is determined as the main depth position of the defect, and the corresponding cross-frequency in-situ focusing intensity is determined as the maximum cross-frequency in-situ focusing intensity.

[0010] Preferably, the defect residual signal is cross-frequency aligned and weighted superimposed to obtain the defect coherent reconstruction echo, including: Based on the path weight function corresponding to each frequency at the main depth position of the defect, the fusion weight of each frequency is determined; Based on the propagation delay function corresponding to the main depth position of the defect, cross-frequency time alignment is performed on the defect residual signals at each frequency; The time-aligned defect residual signals of each frequency are weighted and superimposed according to the fusion weight to obtain the defect coherent reconstruction echo.

[0011] Preferably, constructing a quantitative feature vector for defects includes: Extract the peak amplitude, main response phase, spectral center, and full width at half maximum (FWHM) from the defective coherent reconstructed echo; The maximum cross-frequency in-situ focusing intensity, the defect principal depth location, the peak amplitude, the principal response phase, the spectral center, and the half-width at half-maximum (WHM) are combined to form the defect quantitative feature vector.

[0012] Preferably, the defect size is calculated based on a physical mapping model and a regression network, including: Based on the defect type and the depth and multi-frequency response information in the defect quantitative feature vector, the initial value of the defect size is obtained by solving the physical mapping model that reflects the correspondence between defect parameters and multi-frequency echo parameters. The quantitative feature vector of the defect is input into the regression network, and the size correction amount is output. The initial defect size is combined with the size correction amount to obtain the defect size.

[0013] To address the above problems, the present invention also provides a multi-frequency ultrasonic signal fusion defect quantitative detection system, comprising: The acquisition and calibration module is used to acquire the multi-frequency aligned echo of the structure under test, and to determine the propagation delay function and path weight function of the multi-frequency signal at different depths based on the multi-dielectric layer propagation model. The residual extraction module is used to extract the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo; The focusing and positioning module is used to calculate the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth based on the propagation delay function and the path weight function, and to determine the candidate depth corresponding to the maximum cross-frequency in-situ focusing intensity as the main depth position of the defect. The coherent reconstruction module is used to perform cross-frequency alignment and weighted superposition of the defect residual signal based on the defect main depth position, the propagation delay function and the path weight function to obtain the defect coherent reconstruction echo, and to construct a defect quantitative feature vector based on the maximum cross-frequency in-situ focusing intensity, the defect main depth position and the defect coherent reconstruction echo. The quantitative assessment module is used to input the defect quantitative feature vector into the classification network, determine the defect type, and calculate the defect size based on the defect type and the depth information and multi-frequency response information in the defect quantitative feature vector, using a physical mapping model and a regression network, and output a defect quantitative detection result containing the defect type, the main depth position of the defect, and the defect size.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention establishes propagation delay and path weight relationships at different depths by denoising, cross-frequency time calibration, and multi-medium layer propagation modeling of echo signals from the structure under test at multiple frequencies. Furthermore, it utilizes defect-free reference echoes to perform interpolation processing on the measured echoes, extracting defect residual signals to effectively reduce the interference of normal interlayer responses on the detection results. Based on this, cross-frequency aggregation of defect residual signals at each candidate depth is performed, and the maximum cross-frequency in-situ focusing intensity is used to determine the main depth location of the defect. This allows multi-frequency signals to form a consistent enhancement around the same defect location, thereby improving the accuracy and stability of deep defect localization. After determining the main depth location of the defect, this invention further performs cross-frequency alignment and weighted superposition on the defect residual signal to obtain the coherent reconstruction echo of the defect. Parameters such as peak amplitude, main response phase, spectral center, and full width at half maximum (FWHM) are extracted from this echo and, together with the maximum cross-frequency in-situ focusing intensity and the main depth location of the defect, constitute a quantitative feature vector of the defect. By inputting this quantitative feature vector into a classification network to determine the defect type, and combining it with a physical mapping model and a regression network to calculate the defect size, this invention achieves joint quantitative detection of defect type, location, and size, enhancing the completeness and precision of the detection results. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to an embodiment of the present invention. Figure 2 This is a functional block diagram of a multi-frequency ultrasonic signal fusion defect quantitative detection system provided in an embodiment of the present invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example: This example provides a method for quantitative detection of defects through multi-frequency ultrasonic signal fusion. See [link to example]. Figure 1 Specifically, it includes the following steps: S1. Obtain the multi-frequency aligned echo of the structure under test, and determine the propagation delay function and path weight function of the multi-frequency signal at different depths based on the multi-dielectric layer propagation model; S2. Extract the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo; S3. Based on the propagation delay function and the path weight function, calculate the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth, and determine the candidate depth corresponding to the maximum cross-frequency in-situ focusing intensity as the main depth position of the defect. S4. Based on the defect main depth position, the propagation delay function and the path weight function, the defect residual signal is cross-frequency aligned and weighted to obtain the defect coherent reconstruction echo, and a defect quantitative feature vector is constructed based on the maximum cross-frequency in-situ focusing intensity, the defect main depth position and the defect coherent reconstruction echo. S5. Input the quantitative feature vector of the defect into the classification network to determine the defect type. Based on the defect type and the depth information and multi-frequency response information in the quantitative feature vector of the defect, calculate the defect size based on the physical mapping model and regression network, and output the quantitative detection result of the defect including the defect type, the main depth position of the defect and the defect size.

[0018] In embodiments of the present invention, multi-frequency aligned echoes of the structure under test are acquired, and the propagation delay function and path weight function of the multi-frequency signals at different depths are determined based on a multi-dielectric layer propagation model, including: This embodiment focuses on defect detection in multi-media layered structures of building exterior walls. The structure under test sequentially comprises a 3mm thick 6061 aluminum alloy metal panel layer, a 50mm thick extruded polystyrene board insulation layer, and a 100mm thick C30 ordinary silicate concrete substrate layer. The detection process utilizes a transceiver-type ultrasonic transducer with a working bandwidth covering 200kHz to 800kHz. Vaseline is applied between the transducer and the surface of the structure under test as an acoustic coupling medium to ensure effective ultrasonic wave penetration into the structure. The transducer is connected to an ultrasonic pulse-echo detection system, and the system sampling trigger mode is set to an internal trigger mode synchronized with the excitation pulse. The emission moment of the excitation pulse is used as the unified zero point of the time axis for all echo signals. The sampling frequency was set to 10MHz, the sampling duration to 200μs, and the corresponding number of discrete sampling points to 2000. The detection used ultrasonic narrowband excitation pulses with three center frequencies: 250kHz, 500kHz, and 750kHz. The excitation pulses were 5-cycle sine waves modulated by the Hanning window. The excitation pulses at each frequency were repeated 64 times, with each excitation corresponding to a single echo acquisition. After 64 acquisitions, the 64 echo sequences were arithmetically averaged to obtain the original echo signals corresponding to each frequency. 500kHz was selected as the reference frequency, which has both good structural penetration and defect resolution capabilities, ensuring that the reference signal has sufficient signal-to-noise ratio and feature identification.

[0019] Wavelet threshold denoising is performed on the original echo signals corresponding to each frequency to suppress system white noise and structural scattering noise in the original echoes. This reduces the interference of noise on subsequent time offset calculations while preserving the effective peak envelope characteristics of the echoes. The denoising calculation formula is as follows: ; in This is the denoised signal corresponding to the k-th frequency. For discrete wavelet positive transform operators, It is the discrete wavelet inverse transform operator. This is a soft threshold shrinkage operator. The wavelet shrinkage threshold corresponding to the k-th frequency. Let t be the original echo signal corresponding to the k-th frequency, and t be the discrete-time sampling point. The specific expression of the soft-threshold shrinkage operator is: ; in The input wavelet decomposition coefficients, For the wavelet shrinkage threshold corresponding to the corresponding frequency, soft thresholding ensures the continuity of wavelet coefficients, avoids signal oscillation problems caused by hard thresholding, and improves the smoothness and feature fidelity of the denoised signal. The wavelet shrinkage threshold adopts an adaptive threshold based on unbiased likelihood estimation. The threshold calculation process involves calculating the unbiased risk estimate for each high-frequency wavelet coefficient at each scale, and taking the threshold corresponding to the minimum risk estimate as the shrinkage threshold for that scale. In specific implementation, the db8 wavelet is selected as the wavelet basis function, and the number of decomposition levels is set to 6. After performing discrete wavelet transform on the original echo signal, the wavelet decomposition coefficients corresponding to each scale are obtained. The high-frequency wavelet coefficients at each scale are subjected to threshold shrinkage using the aforementioned soft thresholding function. After thresholding, the processed wavelet coefficients are subjected to inverse discrete wavelet transform to obtain the denoised signals corresponding to each frequency. Using the denoised signal corresponding to the reference frequency as a benchmark, the cross-frequency time offset of the denoised signals at other frequencies relative to the benchmark is calculated using the cross-correlation method. The cross-correlation method determines the time offset by calculating the similarity between two time-domain signals, which can accurately eliminate the systematic time deviation caused by the transducer trigger delay and the initial propagation delay of multi-frequency signals, ensuring the consistency of subsequent cross-frequency focusing processing. The formula for calculating the cross-correlation function is as follows: ; in Let be the cross-correlation function between the denoised signal at frequency k and the denoised signal at reference frequency r, and let τ be the time offset. The denoised signal corresponding to the reference frequency. Let t be the denoised signal corresponding to the k-th frequency, t be the discrete-time sampling point, and N be the total length of the discrete-time sequence. In this embodiment, N is taken as 2000. During the calculation process, when When the value exceeds the range of 1 to 2000, the corresponding signal value is 0. In a discrete sampling system, The value range is from -100 to 100 sampling points, corresponding to a time offset interval of -10μs to 10μs. This interval covers the maximum possible deviation range between the trigger delay and the initial propagation delay of the multi-frequency signal.

[0020] For each non-reference frequency, the maximum value of the cross-correlation function is taken. The τ value corresponding to the maximum value is the cross-frequency time offset of the denoised signal relative to the reference signal at that frequency, denoised as τ. Based on the cross-frequency time offset corresponding to each frequency, the denoised signal at the corresponding frequency is time-domain shifted and calibrated. The calibration calculation formula is as follows: ; in This is the multi-frequency aligned echo corresponding to the k-th frequency. This represents the cross-frequency time offset corresponding to the k-th frequency. Let be the denoised signal corresponding to the k-th frequency, and t be the discrete-time sampling point. During the time-domain shift, when When the value exceeds the range of 1 to 2000, the corresponding signal value is 0. The denoised signal corresponding to the reference frequency is not shifted and is used as the multi-frequency aligned echo of the corresponding frequency. Finally, the multi-frequency aligned echo corresponding to the full frequency sequence is obtained, which provides input for the subsequent extraction of defect residual signals.

[0021] A multi-medium layer propagation model based on the theory of vertically incident plane wave propagation is constructed. This model conforms to the actual testing conditions of a transceiver-integrated ultrasonic transducer with vertical incidence on the structure under test. It can accurately describe the propagation delay, energy attenuation, and interface transmission characteristics of ultrasonic waves in multi-layer media, avoiding propagation path deviations caused by non-vertical incidence and ensuring the accuracy of subsequent cross-frequency in-situ focusing calculations. Based on this multi-medium layer propagation model, the propagation velocity, attenuation coefficient, and interface transmission coefficient of ultrasonic waves at each frequency in each medium layer are obtained. For the three-layer medium structure in this embodiment, the propagation velocity, attenuation coefficient, and interface transmission coefficient of each medium layer are obtained in advance through a combination of standard sample testing and on-site calibration. The basic acoustic parameters of the media were determined using the pulse-echo transmission method at three detection frequencies. The longitudinal wave propagation velocities of ultrasonic waves at each frequency in the corresponding media layer were measured. Specifically, the propagation velocities of the aluminum alloy panel at the three frequencies were 6300 m / s, 6320 m / s, and 6330 m / s; the propagation velocities of the extruded polystyrene board insulation layer at the three frequencies were 2350 m / s, 2380 m / s, and 2400 m / s; and the propagation velocities of the C30 concrete substrate at the three frequencies were 3800 m / s, 3820 m / s, and 3840 m / s. These frequency-specific measured propagation velocities can be used to match the weak frequency dispersion of ultrasonic waves in the media. Features include eliminating time delay calculation errors caused by differences in propagation wave speeds at different frequencies, ensuring the accuracy of two-way propagation time delay calculations; for the attenuation coefficient of each frequency in the corresponding medium layer, the penetration method is used for actual measurement. By testing the amplitude attenuation of ultrasonic waves before and after passing through a standard medium sample of known thickness, the attenuation coefficient per unit thickness is calculated. Specifically, the attenuation coefficients of the aluminum alloy panel at the three frequencies are 0.02 Np / mm, 0.03 Np / mm, and 0.04 Np / mm, respectively; and the attenuation coefficients of the extruded polystyrene board insulation layer at the three frequencies are 0.15 Np / mm, 0.32 Np / mm, and 0.55 Np / mm, respectively. C3 The attenuation coefficients of the concrete substrate at three frequencies were 0.08 Np / mm, 0.16 Np / mm, and 0.28 Np / mm, respectively. These frequency-specific measured attenuation coefficients accurately characterize the energy attenuation of ultrasonic waves at different frequencies in the medium, especially the strong attenuation characteristics of high-frequency signals in the insulation layer and concrete, ensuring the accuracy of the path weighting function for energy compensation of signals at different frequencies. The interface transmission coefficient between layers was calculated based on the acoustic impedance of each medium layer and the two-way transmission law of the vertically incident plane wave. The acoustic impedance is the product of the medium density and the propagation wave velocity at the corresponding frequency. The density of the aluminum alloy panel is 2700 kg / m³. 3 The density of the extruded polystyrene board insulation layer is 35 kg / m³. 3 The density of C30 concrete substrate is 2400 kg / m³. 3In a two-way propagation process, ultrasound waves need to pass through the same interlayer interface twice. The outgoing path is from the upper medium to the lower medium, and the return path is from the lower medium to the upper medium. Therefore, the interface transmission coefficient of a single interlayer interface is the product of the outgoing transmission coefficient and the return transmission coefficient. The formula for calculating the sound pressure transmission coefficient under perpendicular incidence conditions is: ; in Let be the sound pressure transmission coefficient of the ultrasonic wave incident from medium m to medium n. Let m be the acoustic impedance of the medium. Let n be the acoustic impedance of the medium. Based on this formula, the interface transfer coefficients between the metal panel layer and the insulation layer at three frequencies are calculated to be 0.0201, 0.0203, and 0.0205, respectively. The interface transfer coefficients between the insulation layer and the concrete substrate layer at three frequencies are calculated to be 0.1182, 0.1190, and 0.1198, respectively. Based on the interface transfer coefficients, the energy loss of ultrasonic waves passing through the interlayer interface can be accurately quantified, ensuring that the path weight function can fully reflect all energy losses during the propagation of ultrasonic waves. After acquiring all acoustic parameters, the two-way propagation delay of ultrasonic waves at each frequency at each candidate depth is calculated based on the thickness of each medium layer and the propagation velocity at the corresponding frequency. This two-way propagation delay is used as the propagation delay function. The two-way propagation delay is the total propagation time of the ultrasonic wave from the transducer emitting surface to the candidate depth and back to the transducer receiving surface. It can accurately characterize the theoretical arrival time of reflected echoes at different depths, providing a unified depth-time mapping benchmark for subsequent cross-frequency in-situ focusing. The formula for calculating the propagation delay function is as follows: ; in Let be the two-way propagation time delay of the ultrasound at the k-th frequency at the candidate depth z, and 2 be the multiplication factor corresponding to the two-way propagation. The number of complete dielectric layers that the ultrasonic wave passes through from the transducer incident surface to the candidate depth z. Let m be the thickness of the m-th layer of medium. Let z be the propagation velocity of the k-th frequency ultrasonic wave in the m-th medium. In this embodiment, the candidate depth z is set to a range of 0 mm to 150 mm, with a depth step of 0.1 mm, covering the entire effective detection depth of the structure under test. When the candidate depth z is located inside the (M+1)-th medium layer, The value is M, and the thickness of the (M+1)th layer in the summation term is replaced by z minus the total thickness of the first M layers of medium, so as to realize the continuous calculation of the two-way propagation delay at any candidate depth. Based on the thickness of each medium layer, the attenuation coefficient at the corresponding frequency, and the interfacial transmission coefficient between layers, the combined weight of attenuation and transmission of ultrasonic waves at each candidate depth at each frequency is calculated as the path weight function. This path weight function can completely quantify all energy loss during the propagation of ultrasonic waves to the candidate depth and back, providing energy compensation weights for different frequency signals in subsequent cross-frequency in-situ focusing. It eliminates the energy imbalance caused by differences in propagation paths between different frequency signals, ensuring the accuracy of the focusing intensity calculation. The formula for calculating the path weight function is as follows: ; in Let z be the path weight of the ultrasound wave of the kth frequency at the candidate depth z. The number of complete dielectric layers that the ultrasonic wave passes through from the transducer incident surface to the candidate depth z. Let be the interfacial transmission coefficient of the ultrasound at the k-th frequency at the interlayer interface between the m-th and (m+1)-th media layers. Let be the attenuation coefficient of the ultrasound wave at frequency k in the m-th layer of the medium. Let z be the thickness of the m-th medium layer, and exp be the natural exponential function. The natural exponential term is used to quantify the amplitude attenuation of ultrasound propagation in the m-th medium layer, and the product term is used to accumulate the transmission energy loss of ultrasound through the interfaces between layers. When the candidate depth z is located inside the (M+1)-th medium layer... The value is M, and the product term is supplemented with the attenuation term corresponding to the (M+1)th layer of medium. This allows for continuous calculation of path weights at any candidate depth, ultimately yielding the propagation delay function and path weight function for each frequency within the entire candidate depth range, providing a calculation basis for subsequent cross-frequency focusing localization and defect coherent reconstruction.

[0022] In an embodiment of the present invention, extracting the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo includes: A reference transfer function for a defect-free multi-medium layer is constructed using the transfer matrix method based on the theory of vertically incident plane waves. The transfer matrix method can comprehensively characterize the frequency domain propagation characteristics of ultrasound in multi-layer media through matrix cascading, accurately simulating the acoustic reflection and transmission patterns at each interface under defect-free conditions, avoiding cumulative errors caused by layered calculations, and ensuring that the generated defect-free reference echo completely matches the normal echo under actual testing conditions. First, the propagation matrix of a single-layer medium and the transfer matrix of the interlayer interfaces are constructed. For a homogeneous isotropic medium layer, the formula for calculating the propagation matrix when ultrasound is vertically incident is: ; in Let be the propagation matrix of the m-th layer of medium, and j be the imaginary unit. Let be the complex wave number of the ultrasound at frequency k in the m-th layer of the medium. Let m be the thickness of the m-th layer of medium. Let m be the density of the m-th layer of medium. Let be the propagation velocity of the ultrasound wave of frequency k in the m-th layer of medium, and be the complex wave number. The calculation formula is Where ω is the angular frequency, Let be the attenuation coefficient of the k-th frequency ultrasound in the m-th medium. Introducing the complex wavenumber can simultaneously characterize the phase propagation and amplitude attenuation characteristics of ultrasound in the medium, ensuring that the propagation matrix can completely reflect the entire effect of the medium layer on the ultrasound. For a defect-free, fully bonded interlayer interface, the boundary conditions of continuous sound pressure and continuous particle velocity are satisfied on both sides of the interface. The corresponding interface transfer matrix is ​​an identity matrix, eliminating the need for additional matrix correction due to interface reflection, thus matching the real working condition of a defect-free interface without additional abnormal reflections. Based on the propagation matrix and interface transfer matrix of each medium layer, the total transfer matrix of the multi-medium layer system is constructed. The total transfer matrix is ​​the cascade product of the matrices of each layer in the order of ultrasound propagation, and the calculation formula is: ; in Let be the total transfer matrix of the multi-dielectric layer system corresponding to the k-th frequency. The propagation matrix of the metal panel layer, The propagation matrix of the insulation layer, The propagation matrix of the concrete substrate layer is given. In this embodiment, the thickness of the concrete substrate layer is much larger than the effective detection depth. Within the effective detection depth, there is no bottom surface reflection interference. Therefore, in the calculation of the transfer matrix, the concrete substrate layer is regarded as a semi-infinite medium. Only the propagation characteristics of the ultrasonic wave incident on the concrete substrate layer are calculated, and the reflection from the bottom surface of the concrete substrate layer is not considered. Based on the total transfer matrix and the boundary conditions at the incident and exit ends, the sound pressure reflection coefficient of the system is calculated, and thus the reference transfer function is obtained. The incident end is the contact interface between the transducer and the metal panel surface. The acoustic impedance on the air side is much smaller than the acoustic impedance of the metal panel. Therefore, the boundary condition at the incident end is a total sound pressure reflection boundary. The exit end is a semi-infinite medium of the concrete substrate layer. There is no upward reflected wave. Therefore, the boundary condition at the exit end is that only downward transmitted wave exists. Based on this boundary condition, the ratio of incident sound pressure to reflected sound pressure, i.e., the sound pressure reflection coefficient of the system, is obtained from the total transfer matrix. The sequence of the sound pressure reflection coefficient with angular frequency is the reference transfer function of the defect-free multi-medium layer system. The reference transfer function characterizes the frequency domain mapping relationship between ultrasonic excitation and echo response under defect-free conditions, providing the core calculation basis for the subsequent generation of defect-free reference echoes.

[0023] After constructing the reference transfer function, the excitation spectrum corresponding to the ultrasonic excitation that is completely consistent with the actual transmitted signal is obtained. For the three detection frequencies, time-domain excitation signals that are completely consistent with the actual transmitted signals are generated. The sampling frequency and number of sampling points of the time-domain excitation signals are consistent with the original echo signals. Discrete Fourier transform is performed on the time-domain excitation signals corresponding to each frequency to obtain the excitation spectrum of the corresponding frequency. By using an excitation signal that is completely consistent with the actual detection, the excitation spectrum can be calculated, ensuring that the generated defect-free reference echo and the actual acquired multi-frequency aligned echo have a completely consistent excitation source basis, eliminating echo deviation caused by differences in excitation signals. The reference transfer function corresponding to each frequency is multiplied point by point in the frequency domain with the excitation spectrum of the same frequency, and then the frequency domain sequence after multiplication is subjected to inverse discrete Fourier transform to obtain the defect-free reference echo corresponding to each frequency in the time domain. The calculation formula is as follows: ; in This is the defect-free reference echo corresponding to the k-th frequency. For the discrete inverse Fourier transform operator, Let k be the reference transfer function corresponding to the k-th frequency. Let be the excitation spectrum corresponding to the k-th frequency, and t be the discrete-time sampling point. The time-domain length, sampling frequency, and time axis zero of the generated defect-free reference echo are completely consistent with the multi-frequency aligned echo, ensuring a one-to-one correspondence between their sampling points in the time domain. Based on the time-domain difference between the multi-frequency aligned echo and the defect-free reference echo, the defect residual signal is extracted using the following formula: ; in The defect residual signal corresponds to the k-th frequency. This is the multi-frequency aligned echo corresponding to the k-th frequency. Let t be the defect-free reference echo corresponding to the k-th frequency, and t be the discrete-time sampling point. By calculating the time-domain difference, the normal interlayer echo components such as the reflection from the metal panel surface, the interface between the metal panel and the insulation layer, and the interface between the insulation layer and the concrete substrate layer under defect-free conditions can be completely removed. The remaining residual signal only contains the abnormal echo response caused by the defect, which can effectively eliminate the interference of normal interface echoes on the subsequent deep defect focusing, location and quantitative analysis. Finally, the defect residual signals corresponding to the three detection frequencies are obtained, which provide input for subsequent defect location and reconstruction.

[0024] In an embodiment of the present invention, based on the propagation delay function and the path weight function, the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth is calculated, and the candidate depth corresponding to the maximum cross-frequency in-situ focusing intensity is determined as the main depth position of the defect, including: Analytical signal processing is performed on the defect residual signals at each frequency. This process converts the real-valued defect residual signals into complex-valued signals while preserving the complete amplitude and phase information of the original signals. This avoids the phase cancellation problem caused by the alternation of positive and negative amplitudes during cross-frequency aggregation of real-valued signals, ensuring that subsequent focusing intensity calculations accurately reflect the phase consistency and energy concentration characteristics of the multi-frequency signals. The formula for calculating the analytical signal is as follows: ; in The analytic signal corresponding to the k-th frequency is... Let j be the defect residual signal corresponding to the k-th frequency, where j is the imaginary unit. The Discrete Hilbert Transform (DHFT) operator is used, where t represents the discrete-time sampling point. The DHFT is achieved by performing frequency domain processing on the discrete Fourier transform result of the defect residual signal. Specifically, a phase shift of -π / 2 is applied to the positive frequency components, and a phase shift of π / 2 is applied to the negative frequency components. After processing, an inverse discrete Fourier transform is performed on the frequency domain sequence to obtain the orthogonal components of the defect residual signal. For each candidate depth within the effective detection depth range, cross-frequency aggregation and focusing intensity calculation are performed sequentially. The step size for the candidate depth is 0.1 mm to ensure that the depth positioning accuracy meets engineering inspection requirements. The two-way propagation delay at the k-th frequency at candidate depth z is also considered. The time delay is converted into the corresponding discrete-time sampling point number by rounding the product of the time delay value and the sampling frequency to obtain the discrete sampling point position corresponding to the time delay. This achieves a precise correspondence between the continuous propagation time delay and the discrete echo sequence, eliminating the mapping deviation between continuous time and discrete sampling. For each candidate depth, the values ​​of each frequency analytical signal at the corresponding discrete sampling point are obtained based on the propagation time delay function corresponding to the candidate depth. Then, the values ​​of each frequency analytical signal are weighted based on the path weight function corresponding to the candidate depth. The weighted values ​​are then summed to obtain the aggregated response corresponding to the candidate depth. The purpose of the weighting process is to compensate for the energy attenuation and interface transmission loss of different frequency ultrasonic waves during propagation, so that the energy contribution of each frequency signal at the same candidate depth is at a balanced level, avoiding the imbalance of focusing weights caused by excessive attenuation of high-frequency signals, and ensuring the accuracy of cross-frequency aggregation. The formula for calculating the aggregated response is: ; in Let be the aggregated response corresponding to the candidate depth z, and K be the total number of detection frequencies. In this embodiment, K is set to 3. Let z be the path weight function for the k-th frequency at the candidate depth z. Let be the value of the analytic signal at the k-th frequency at the propagation delay corresponding to the candidate depth z, where z is the candidate depth.

[0025] Based on the convergent response and the response intensity of each frequency analytical signal at the candidate depth, the cross-frequency in-situ focusing intensity corresponding to that candidate depth is calculated. The cross-frequency in-situ focusing intensity is calculated using the normalized coherence coefficient method, which can quantify the phase consistency and energy concentration of multi-frequency signals at the same candidate depth. The calculation formula is as follows: ; in Let be the cross-frequency in-situ focusing intensity corresponding to candidate depth z. The meanings of the other parameters remain consistent with the aforementioned formula. The output value of this calculation formula ranges from 0 to 1. When the multi-frequency signals originate from the real defect interface at the same candidate depth, the phases of the analytical signals of each frequency tend to converge, and the square of the summed magnitude is close to the sum of the squares of the magnitudes of each frequency, resulting in a focusing intensity close to 1. When the multi-frequency signals are random noise or pseudo-reflection signals from different interfaces, the phases of the analytical signals of each frequency are disordered, and the square of the summed magnitude is much smaller than the sum of the squares of the magnitudes of each frequency, resulting in a focusing intensity close to 0. This allows for the differentiation between the real defect interface and pseudo-reflection signals, achieving precise locking of the defect depth. After calculating the cross-frequency in-situ focusing intensity for all candidate depths within the full depth range, a focusing intensity sequence for the entire depth range is obtained. The maximum value of this focusing intensity sequence is taken, and the candidate depth corresponding to this maximum value is the main depth position of the defect, denoted as . This maximum value is the maximum cross-frequency in-situ focusing intensity, denoted as . This provides a core benchmark for subsequent coherent reconstruction of defects and feature construction.

[0026] In an embodiment of the present invention, based on the defect principal depth location, the propagation delay function, and the path weight function, the defect residual signal is cross-frequency aligned and weighted to obtain a coherent reconstruction echo of the defect. A quantitative feature vector of the defect is constructed based on the maximum cross-frequency in-situ focusing intensity, the defect principal depth location, and the coherent reconstruction echo of the defect, including: Based on the path weight function corresponding to each frequency at the main depth location of the defect, the fusion weight of each frequency is determined. The fusion weight adopts the normalized calculation method of path weight, which can achieve energy balance compensation of each frequency signal at the main depth location of the defect, eliminate the energy imbalance problem caused by the difference in propagation attenuation of different frequencies, avoid the high-frequency defect characteristics being masked by low-frequency signals, and ensure that the sum of the fusion weights is 1, ensuring that the amplitude scale of the reconstructed echo is consistent with the original defect residual signal. The calculation formula of the fusion weight is: ; in The k-th frequency at the main depth of the defect The corresponding fusion weights, The k-th frequency at the main depth of the defect The path weight function value is assigned at the corresponding location, where K is the total number of detection frequencies (in this embodiment, K is 3), and j is the frequency traversal index. The defect master depth position is defined as follows: After determining the fusion weights, cross-frequency time alignment is performed on the defect residual signals of each frequency based on the propagation delay function corresponding to the defect master depth position. Different frequencies of ultrasonic waves have inherent differences in their two-way propagation delay at the same defect master depth position. Direct superposition would cause the echo peaks of the same defect to be misaligned in the time domain, failing to achieve coherent enhancement. Therefore, the two-way propagation delay of the reference frequency at the defect master depth position is used as a unified time reference, and time-domain shift calibration is performed on the defect residual signals of each frequency to ensure that the echo components originating from the defect master depth position in each frequency are completely aligned to the same time domain position. The time-domain shift calculation formula for cross-frequency time alignment is as follows: ; in This represents the time-aligned defect residual signal corresponding to the k-th frequency. This represents the original defect residual signal corresponding to the k-th frequency. The k-th frequency at the main depth of the defect The corresponding two-way propagation delay, The reference frequency r at the principal depth of the defect The corresponding two-way propagation delay is given at this location. In this embodiment, the reference frequency r corresponds to 500kHz, and t is the discrete-time sampling point. During the time-domain shift, when... When the sampling point range exceeds 1 to 2000, the corresponding signal value is set to 0 to ensure that the signal length after time-domain shift is completely consistent with the original signal. After cross-frequency time alignment is completed, the defect residual signals of each frequency after time alignment are weighted and superimposed according to the fusion weight to obtain the defect coherent reconstruction echo. The weighted superposition process enables multi-frequency echo signals from the same defect interface to achieve coherent enhancement due to phase consistency, while incoherent components such as random noise and pseudo-reflection signals cancel each other out due to phase disorder, which greatly improves the signal-to-noise ratio of the defect echo and completely preserves the time-domain characteristics of the defect, providing a highly reliable signal basis for subsequent quantitative feature extraction. The calculation formula for the defect coherent reconstruction echo is: ; in To obtain the final coherent reconstruction echo of the defect, The fusion weights are the values ​​corresponding to the k-th frequency. The time-aligned defect residual signal corresponds to the k-th frequency, where K is the total number of detection frequencies and t is the discrete-time sampling point. The sampling frequency, sampling length, and time axis reference of the final coherent reconstruction echo of the defect are consistent with the original defect residual signal, providing input for subsequent quantitative feature extraction.

[0027] Four quantitative parameters—peak amplitude, principal response phase, spectral center, and full width at half maximum (FWHM)—are extracted from the coherent reconstructed echo of the defect. First, the peak amplitude is extracted. The peak amplitude directly characterizes the strongest reflected energy at the defect interface and reflects the degree of acoustic impedance abrupt change at the defect. It is the core time-domain amplitude feature for defect type identification and size quantification. The formula for calculating the peak amplitude is: ; in The peak amplitude of the coherent reconstructed echo from the defect is given by 'max', where 'max' is the maximum value operator. For the coherent reconstruction echo of the defect, t is the discrete-time sampling point, and the discrete-time sampling point corresponding to the peak amplitude is the peak time. Peak time The calculation formula is: ; in The peak moment is defined as the time of peak reflection. The meanings of the other parameters remain consistent with the aforementioned formula. Using the maximum absolute value eliminates the influence of the positive and negative phases of the echo signal on peak positioning, accurately pinpointing the strongest response moment and corresponding amplitude of the defect reflection. After extracting the peak amplitude, the main response phase is extracted. The main response phase characterizes the inherent phase characteristics of the reflected echo from the defect interface and can distinguish the interface acoustic impedance variation patterns of different types of defects. It is a key feature for identifying defect types such as interlayer debonding, voids, and cracks. First, the coherent reconstructed echo of the defect is subjected to analytical signal processing to obtain a complex analytical signal containing amplitude and phase information. The calculation formula for analytical signal processing is as follows: ; in The complex analytical signal corresponding to the coherent reconstruction echo of the defect. For the defect coherent reconstruction echo, j is the imaginary unit. Here, t represents the discrete Hilbert transform operator, and t is the discrete-time sampling point. The discrete Hilbert transform is achieved by performing frequency domain phase shifting on the discrete Fourier transform result of the defective coherent reconstructed echo. A phase shift of -π / 2 is applied to the positive frequency component, and a phase shift of π / 2 is applied to the negative frequency component. After processing, an inverse discrete Fourier transform is performed to obtain the orthogonal components of the echo. The complex analytic signal constructed in this way can completely retain the instantaneous amplitude and instantaneous phase information of the original echo. Based on the complex analytic signal, the principal response phase corresponding to the peak time is extracted, and the calculation formula is as follows: ; in Main response phase, For complex numbers, take the argument operator. The value of the complex analytical signal corresponding to the peak time is taken. At the peak time, the phase at the peak time is used as the main response phase, which ensures that the extracted phase information corresponds to the core reflection characteristics of the defect interface and avoids interference from noise and clutter in non-peak areas. The spectral center is extracted; the spectral center represents the centroid position of the defect echo power spectrum, reflecting the reflection and scattering characteristics of the defect to ultrasonic signals of different frequencies. Defects of different sizes show significant differences in their responses to ultrasonic signals of different frequencies, which is the core frequency domain feature for defect size quantification. First, a discrete Fourier transform is performed on the coherent reconstructed echo of the defect to obtain the frequency domain sequence of the echo. The calculation formula is as follows: ; in To coherently reconstruct the frequency domain sequence of the echo from the defect, For discrete Fourier transform operators, For the coherent reconstruction echo of the defect, f is the discrete point on the frequency axis. Based on the system sampling frequency and the number of sampling points, the resolution of the frequency axis is determined to be 5kHz, and the frequency axis coverage range is 0 to 5MHz, which conforms to the Nyquist sampling theorem. The power spectrum of the echo is calculated based on the frequency domain sequence, and then the spectral center is obtained by calculating the centroid of the power spectrum. The calculation formula is as follows: ; in As the center of the spectrum, For the power spectrum of the coherent reconstructed echo of the defect, f is the discrete point of the frequency axis. The integral operation is replaced by the summation operation within the corresponding frequency range in the discrete sampling system. The center of the spectrum is calculated by using the centroid of the power spectrum, which can eliminate the influence of random fluctuations in the spectrum amplitude, stably characterize the main frequency distribution characteristics of the defect echo, and avoid the instability caused by the extraction of a single peak frequency. After extracting the spectral center, the half-width at half-maximum (HWHM) is extracted. HWHM represents the temporal duration at half the peak value of the defect echo, reflecting the spatial extent and interface smoothness of the defect. It is a core temporal scale feature for defect size quantification. First, based on the peak amplitude, the HWHM threshold is determined to be half the peak amplitude. Within the time interval to the left of the peak time, the first discrete-time sampling point whose absolute amplitude value equals the HWHM threshold is found, denoted as . Within the time domain interval to the right of the peak time, find the first discrete-time sampling point whose absolute amplitude value equals the half-maximum amplitude threshold, denoted as . If multiple sampling points meet the conditions within the time domain interval, select the sampling point corresponding to the narrowest interval containing the peak time to avoid interference from echo sidelobes and clutter on the half-width at half-maximum (WHM) extraction. The formula for calculating the WHM is: ; in Half height and width, These are the discrete-time sampling points corresponding to the right half-height amplitude at the peak time. The discrete-time sampling point corresponding to the left half-width at the peak moment is used. The half-width at half-maximum (WHM) is used to characterize the temporal scale characteristics of the echo, which can eliminate the influence of overall amplitude scaling and stably reflect the temporal broadening of the defect reflection echo, which is directly related to the spatial size of the defect. The maximum cross-frequency in-situ focusing intensity, defect principal depth position, peak amplitude, principal response phase, spectral center, and WHM are combined in a fixed order to form the defect quantitative feature vector. The formula for calculating the defect quantitative feature vector is as follows: ; in Quantitative feature vector of defects, To achieve maximum cross-frequency in-situ focusing intensity, This indicates the main depth location of the defect. This represents the peak amplitude. Main response phase, As the center of the spectrum, With a half-width-height ratio, this feature vector simultaneously contains cross-frequency focusing consistency information, defect spatial location information, temporal amplitude information, interface phase characteristic information, frequency domain distribution information, and temporal scale information. It can provide comprehensive and stable input features for subsequent defect type identification and size quantification, and finally complete the construction of the defect quantitative feature vector.

[0028] In an embodiment of the present invention, the defect quantitative feature vector is input into a classification network to determine the defect type. Based on the defect type and the depth information and multi-frequency response information in the defect quantitative feature vector, the defect size is calculated using a physical mapping model and a regression network. The result is then a defect quantitative detection result containing the defect type, the main depth position of the defect, and the defect size, including: The quantitative feature vector of defects is input into a pre-trained classification network to determine the defect type. The classification network adopts an architecture combining a one-dimensional convolutional neural network and a fully connected layer, which can automatically extract deep correlation features from the feature vector to achieve stable identification of defect types. The specific architecture of the classification network is as follows: input layer, first one-dimensional convolutional layer, first pooling layer, second one-dimensional convolutional layer, second pooling layer, first fully connected layer, Dropout layer, second fully connected layer, and softmax output layer. The dimension of the input layer is 6, which perfectly matches the dimension of the quantitative feature vector of defects. The first one-dimensional convolutional layer has 16 convolutional kernels with a kernel size of 3 and uses the ReLU activation function. The first pooling layer uses max pooling with a pooling window size of 2. The second one-dimensional convolutional layer has 32 convolutional kernels with a kernel size of 3 and uses the ReLU activation function. The second pooling layer uses max pooling with a pooling window size of 2. The first fully connected layer has 64 neurons and uses the ReLU activation function. The inactivation rate of the Dropout layer is set to 0.3 to suppress... To prevent network overfitting, the second fully connected layer has three neurons, corresponding to three preset defect types: interlayer debonding, voids, and cracks. The Softmax output layer converts the output of the fully connected layer into probability values ​​for each defect type. The training dataset for the classification network is constructed through a combination of finite element simulation and standard specimen testing. The dataset contains quantitative feature vector samples of different defect types, depths, and sizes, with a total of no less than 10,000 samples, of which 70% are simulated samples and 30% are measured samples. During training, the cross-entropy loss function is used as the loss calculation benchmark, and the Adam optimizer is used for iterative updates of network weights. The initial learning rate is set to 0.001, the batch size is set to 32, and the number of iterations is set to 200. Training is terminated early when the validation set loss does not decrease for 10 consecutive iterations. After training, the classification network can stably identify the three defect types. After inputting the quantitative feature vectors of the defects into the trained classification network, the network outputs the probability values ​​corresponding to the three defect types, calculated using the following formula: ; in Given a quantitative feature vector v for the input defect, the probability value that the defect type is c is given. For the forward propagation operator of the trained classification network, v is the quantitative feature vector of defects, and c is the defect type index, corresponding to three types of defects: interlayer debonding, voids, and cracks. The defect type with the highest probability value is taken as the final identified defect type. The calculation formula is as follows: ; in For the final determined defect type, the meanings of the remaining parameters remain consistent with the aforementioned formula. This classification network can fully explore the correlation between the parameters of each dimension in the feature vector and the defect type, exhibiting higher recognition accuracy and anti-interference capability compared to single threshold judgment. Based on the defect type and the depth and multi-frequency response information in the quantitative feature vector of the defect, the initial value of the defect size is obtained through a physical mapping model reflecting the correspondence between defect parameters and multi-frequency echo parameters. The physical mapping model is constructed based on the acoustic scattering theory of ultrasonic perpendicularly incident defect interfaces and the plane wave reflection law. It can establish a deterministic physical mapping relationship between defect type, defect depth, defect size, and multi-frequency echo response parameters, providing initial values ​​consistent with the acoustic propagation mechanism for defect size calculation, avoiding the problem of insufficient mechanistic interpretation in purely data-driven models. The expression of the physical mapping model is: ; in The measured multi-frequency response is the value of the defect residual signal at the k-th frequency. In this embodiment, the measured multi-frequency response is the peak amplitude of the defect residual signal at the corresponding frequency. For the physical mapping operator corresponding to the k-th frequency, For the identified defect type, Let s be the principal depth of the defect and s be the defect size. The physical mapping operator is constructed through prior theoretical derivation and standard sample calibration. For three defect types—interlayer debonding, voids, and cracks—the mapping relationship between multi-frequency response quantities and defect sizes at different depths and sizes is calibrated, forming a mapping operator library categorized by defect type and frequency. When solving for the initial value of the defect size, the optimization objective is to minimize the sum of squared residuals between the measured multi-frequency response quantities and the output values ​​of the physical mapping model. The least squares method is used for the solution, and the calculation formula is as follows: ; in To obtain the initial value of the defect size, K is the total number of detection frequencies. In this embodiment, K is set to 3. The meanings of the other parameters are consistent with the aforementioned formula. Using the least squares method to solve the problem can ensure that the initial value of the defect size best matches the physical laws of the measured multi-frequency response, providing a reliable reference value for subsequent size correction.

[0029] After solving for the initial defect size, the quantitative defect feature vector is input into a pre-trained regression network, which outputs a size correction. The regression network corrects the system bias caused by the discreteness of medium parameters and fluctuations in detection conditions in the physical mapping model, further improving the accuracy of defect size quantification. The regression network adopts a one-dimensional convolutional neural network architecture derived from the classification network. The specific architecture consists of an input layer, a first one-dimensional convolutional layer, a first pooling layer, a second one-dimensional convolutional layer, a second pooling layer, a first fully connected layer, a Dropout layer, a second fully connected layer, and a linear output layer. The input layer has a dimension of 6, which perfectly matches the dimension of the quantitative defect feature vector. The first one-dimensional convolutional layer has 16 convolutional kernels with a kernel size of 3, and the ReLU activation function is used. The first pooling layer uses max pooling with a pooling window size of 2. The second one-dimensional convolutional layer has 32 convolutional kernels. The convolutional kernel size is 3, the activation function is ReLU, the second pooling layer uses max pooling with a pooling window size of 2, the first fully connected layer has 64 neurons, the activation function is ReLU, the dropout layer has a deactivation rate of 0.3, the second fully connected layer has 1 neuron, and the linear output layer directly outputs the size correction. The training dataset for the regression network is from the same source as the training dataset for the classification network. During training, the mean squared error loss function is used as the loss calculation benchmark, and the Adam optimizer is used for iterative updates of network weights. The initial learning rate is set to 0.001, the training batch size is set to 32, and the number of iterations is set to 200. Training is terminated early when the validation set loss does not decrease for 10 consecutive iterations. After training, the regression network can stably output the size correction that conforms to the actual test conditions. The formula for calculating the size correction is: ; in This is the size correction amount. The forward propagation operator of the trained regression network is given by v, which is the quantitative feature vector of the defect. The initial defect size is fused with the size correction value to obtain the final defect size. The fusion calculation formula is as follows: ; in For the final determined defect size, This is the initial value for the defect size. For the size correction, a fusion approach of physical mapping initial value plus data-driven correction is adopted. This retains the interpretability of the physical mechanism of defect size quantification while eliminating systematic errors caused by fluctuations in medium parameters under actual inspection conditions, thus improving the stability and accuracy of size quantification. After determining the defect type and defect size, the output is a quantitative defect detection result containing the defect type, main depth location, and defect size. The expression for the quantitative defect detection result is: ; in The final output is the quantitative detection result of defects. Defect type For defect depth, The value of is related to the location of the main depth of the defect. exactly the same, The defect size is used to complete the full-process quantitative detection of defects in multi-medium layer structures.

[0030] like Figure 2 The diagram shown is a functional block diagram of a multi-frequency ultrasonic signal fusion defect quantitative detection system provided in an embodiment of the present invention.

[0031] In this embodiment, the functions of each module / unit are as follows: The acquisition and calibration module is used to acquire the multi-frequency aligned echo of the structure under test, and to determine the propagation delay function and path weight function of the multi-frequency signal at different depths based on the multi-dielectric layer propagation model. The residual extraction module is used to extract the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo; The focusing and positioning module is used to calculate the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth based on the propagation delay function and the path weight function, and to determine the candidate depth corresponding to the maximum cross-frequency in-situ focusing intensity as the main depth position of the defect. The coherent reconstruction module is used to perform cross-frequency alignment and weighted superposition of the defect residual signal based on the defect main depth position, the propagation delay function and the path weight function to obtain the defect coherent reconstruction echo, and to construct a defect quantitative feature vector based on the maximum cross-frequency in-situ focusing intensity, the defect main depth position and the defect coherent reconstruction echo. The quantitative assessment module is used to input the defect quantitative feature vector into the classification network, determine the defect type, and calculate the defect size based on the defect type and the depth information and multi-frequency response information in the defect quantitative feature vector, using a physical mapping model and a regression network, and output a defect quantitative detection result containing the defect type, the main depth position of the defect, and the defect size.

[0032] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for quantitative detection of defects by fusing multi-frequency ultrasonic signals, characterized in that, Includes the following steps: S1. Obtain the multi-frequency aligned echo of the structure under test, and determine the propagation delay function and path weight function of the multi-frequency signal at different depths based on the multi-dielectric layer propagation model; S2. Extract the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo; S3. Based on the propagation delay function and the path weight function, calculate the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth, and determine the candidate depth corresponding to the maximum cross-frequency in-situ focusing intensity as the main depth position of the defect. S4. Based on the defect main depth position, the propagation delay function and the path weight function, the defect residual signal is cross-frequency aligned and weighted to obtain the defect coherent reconstruction echo, and a defect quantitative feature vector is constructed based on the maximum cross-frequency in-situ focusing intensity, the defect main depth position and the defect coherent reconstruction echo. S5. Input the quantitative feature vector of the defect into the classification network to determine the defect type. Based on the defect type and the depth information and multi-frequency response information in the quantitative feature vector of the defect, calculate the defect size based on the physical mapping model and regression network, and output the quantitative detection result of the defect including the defect type, the main depth position of the defect and the defect size.

2. The method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to claim 1, characterized in that, Obtaining multi-frequency aligned echoes of the structure under test includes: The raw echo signals of the structure under test at multiple frequencies are acquired, and the raw echo signals are denoised to obtain the denoised signals corresponding to each frequency. A denoised signal corresponding to a reference frequency is selected as the benchmark, and the cross-correlation method is used to calculate the cross-frequency time offset of the denoised signals of other frequencies relative to the benchmark. The denoised signals of each frequency are time-shifted and calibrated according to the cross-frequency time offset to obtain the multi-frequency aligned echo.

3. The method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to claim 1, characterized in that, Based on the multi-medium layer propagation model, the propagation delay function and path weight function of multi-frequency signals at different depths are determined, including: Based on the multi-medium layer propagation model, the propagation velocity, attenuation coefficient, and interfacial transmission coefficient of ultrasonic waves at each frequency in each medium layer are obtained. Based on the thickness of each medium layer and the propagation wave velocity, the two-way propagation delay of ultrasonic waves at each frequency at each candidate depth is calculated and used as the propagation delay function. Based on the thickness of each medium layer, the attenuation coefficient, and the interface transmission coefficient, the combined weight of attenuation and transmission of ultrasonic waves at each candidate depth at each frequency is calculated and used as the path weight function.

4. The method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to claim 1, characterized in that, Extracting the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo includes: Constructing a defect-free multi-medium layer reference transfer function based on the transfer matrix method; The reference transfer function is combined with the excitation spectrum corresponding to the ultrasonic excitation to generate a defect-free reference echo corresponding to each frequency. The defect residual signal is extracted based on the difference between the multi-frequency aligned echo and the defect-free reference echo.

5. The method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to claim 1, characterized in that, Calculating the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth includes: The defect residual signals at each frequency are subjected to analytical signal processing; For each candidate depth, based on the propagation delay function and the path weight function corresponding to the candidate depth, cross-frequency aggregation is performed on the analytical signals of each frequency to obtain the aggregated response corresponding to the candidate depth; Based on the aggregation response and the response intensity of each frequency analytical signal at the candidate depth, the cross-frequency in-situ focusing intensity corresponding to the candidate depth is calculated; The candidate depth with the largest cross-frequency in-situ focusing intensity is determined as the main depth position of the defect, and the corresponding cross-frequency in-situ focusing intensity is determined as the maximum cross-frequency in-situ focusing intensity.

6. The method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to claim 1, characterized in that, The defect residual signal is cross-frequency aligned and weighted to obtain the defect coherent reconstruction echo, including: Based on the path weight function corresponding to each frequency at the main depth position of the defect, the fusion weight of each frequency is determined; Based on the propagation delay function corresponding to the main depth position of the defect, cross-frequency time alignment is performed on the defect residual signals at each frequency; The time-aligned defect residual signals of each frequency are weighted and superimposed according to the fusion weight to obtain the defect coherent reconstruction echo.

7. The method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to claim 1, characterized in that, Constructing a quantitative feature vector for defects includes: Extract the peak amplitude, main response phase, spectral center, and full width at half maximum (FWHM) from the defective coherent reconstructed echo; The maximum cross-frequency in-situ focusing intensity, the defect principal depth location, the peak amplitude, the principal response phase, the spectral center, and the half-width at half-maximum (WHM) are combined to form the defect quantitative feature vector.

8. The method for quantitative detection of defects in multi-frequency ultrasonic signal fusion according to claim 1, characterized in that, Defect size is calculated based on a physical mapping model and a regression network, including: Based on the defect type and the depth and multi-frequency response information in the defect quantitative feature vector, the initial value of the defect size is obtained by solving the physical mapping model that reflects the correspondence between defect parameters and multi-frequency echo parameters. The quantitative feature vector of the defect is input into the regression network, and the size correction amount is output. The initial defect size is combined with the size correction amount to obtain the defect size.

9. A quantitative detection system for multi-frequency ultrasonic signal fusion defects, used to execute the quantitative detection method for multi-frequency ultrasonic signal fusion defects according to any one of claims 1-8, characterized in that, include: The acquisition and calibration module is used to acquire the multi-frequency aligned echo of the structure under test, and to determine the propagation delay function and path weight function of the multi-frequency signal at different depths based on the multi-dielectric layer propagation model. The residual extraction module is used to extract the defect residual signal based on the multi-frequency aligned echo and the defect-free reference echo; The focusing and positioning module is used to calculate the cross-frequency in-situ focusing intensity of the defect residual signal at each candidate depth based on the propagation delay function and the path weight function, and to determine the candidate depth corresponding to the maximum cross-frequency in-situ focusing intensity as the main depth position of the defect. The coherent reconstruction module is used to perform cross-frequency alignment and weighted superposition of the defect residual signal based on the defect main depth position, the propagation delay function and the path weight function to obtain the defect coherent reconstruction echo, and to construct a defect quantitative feature vector based on the maximum cross-frequency in-situ focusing intensity, the defect main depth position and the defect coherent reconstruction echo. The quantitative assessment module is used to input the defect quantitative feature vector into the classification network, determine the defect type, and calculate the defect size based on the defect type and the depth information and multi-frequency response information in the defect quantitative feature vector, using a physical mapping model and a regression network, and output a defect quantitative detection result containing the defect type, the main depth position of the defect, and the defect size.

Images