Nondestructive testing method for defects of composite material

Abstract

The invention discloses a nondestructive testing method for composite material defects, and relates to the technical field of nondestructive testing, and the method comprises the following steps: firstly, collecting a scanning waveform of a test piece A, determining a potential defect area based on an echo amplitude and a time difference, and outputting a sequence containing space coordinates, an original waveform and focusing parameters; an effective time window is intercepted after preprocessing, sub-bands are generated through wavelet packet decomposition, and total energy is calculated and normalized to obtain an energy vector; secondly, based on a known sample, evaluating the separability of sub-bands by using a Fisher criterion, sorting the sub-bands, determining an optimal energy dimension through cross validation, combining sub-band energy features with phase and time difference features into composite vectors, inputting the composite vectors into a dual-channel lightweight deep network, outputting defect categories and confidence coefficients, and mapping the defect categories and confidence coefficients to a C scanning frame image; and finally, backtracking the three-dimensional coordinates, calculating the defect volume and the residual wall thickness, and comparing with a material performance database to output a conclusion. The problem of misjudgment caused by echo waveform similarity is solved, and detection closed-loop optimization is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of nondestructive testing, more particularly, to a nondestructive testing method for composite material defects. BACKGROUND

[0002] In the prior art, ultrasonic testing is an important means for nondestructive testing of composite material defects, and phased array ultrasonic technology acquires a sequence of A-scan waveforms by scanning, and realizes defect detection by using echo signal characteristics, and is widely used in the industry.

[0003] However, the prior art has the following obvious deficiencies: first, different defect echo waveforms have high similarity, and only relying on single frequency band energy characteristics for analysis makes it difficult to accurately distinguish defect types, and is prone to misjudgment; second, the feature extraction dimension is limited, and multi-dimensional information such as phase centroid, phase dispersion, echo time difference is often ignored, and the feature fusion is insufficient, which affects the defect recognition accuracy; third, the scanning strategy for potential defect areas is fixed, and lacks dynamic fine scanning optimization, which is prone to imbalance between detection efficiency and accuracy; fourth, defect evaluation is mostly limited to type recognition, and does not effectively combine defect three-dimensional parameters (such as volume and residual wall thickness) with material performance database for comprehensive judgment, and it is difficult to provide clear quality control guidance.

[0004] In view of the above problems, the present application provides a solution. SUMMARY

[0005] The present application aims to provide a nondestructive testing method for composite material defects, which solves the technical problem that different echo waveforms have high similarity in the prior art ultrasonic testing, and it is difficult to accurately distinguish defect types by relying only on single frequency band energy.

[0006] The purpose of the present application can be achieved by the following technical solutions: A nondestructive testing method for composite material defects, comprising the following steps: scanning a composite material test piece by a phased array transducer, acquiring A-scan waveforms of each measurement point, and determining a potential defect area based on the echo amplitude and echo time difference between each type of echo in the A-scan waveform, and outputting an A-scan waveform sequence containing spatial coordinates, original A-scan waveform and focusing parameters; wherein the A-scan waveform is a time domain echo signal acquired by scanning with a phased array transducer; the measurement point focusing parameters include delay line parameters, element aperture and center frequency; After preprocessing the original A-scan waveform, an effective time window containing surface echo, defect echo and bottom echo is intercepted, a plurality of sub-bands are obtained by wavelet packet decomposition, the total energy of each sub-band is calculated and normalized to form a corresponding sub-band energy vector; Based on the sample of the known defect type, the Fisher criterion is used to evaluate the separability index of each sub-band, and the energy vectors of each sub-band are sorted in descending order according to the separability index to form an energy vector priority list of each sub-band; An incremental dimension cross-validation method is used to adaptively determine the optimal energy dimension, and the selected sub-band energy vector features corresponding to the optimal energy dimension are combined with the phase centroid, the phase dispersion and the time difference features to form a composite feature vector. Based on the obtained composite feature vector, the classification model is input to output the defect category and the classification confidence, and the output result is mapped to the C-scan frame image with spatial coordinates as the dimension; based on the C-scan frame image, the three-dimensional coordinates of the defect are traced back, the defect volume and the remaining wall thickness are calculated in combination with the sound velocity, and the defect detection conclusion is output by comparing the material performance database.

[0007] As a further scheme of the application, when the phased array transducer is scanned, an N-element phased array transducer is used, where N is the number of elements, the element spacing of the phased array transducer is P, and the center frequency is f0.

[0008] As a further scheme of the application, the determination of the potential defect area includes: calculating the average amplitude of the surface echo, the average amplitude of the bottom echo and the noise standard deviation of the defect-free area in the A-scan waveform. When the phased array transducer completes a scanning line along the X direction at an initial scanning interval, the A-scan waveform of each measurement point is analyzed in real time; if any of the following conditions is met, it is marked as a potential defect area and triggers the phased array transducer fine scanning. When the absolute deviation of the surface echo amplitude or the bottom echo amplitude from the average surface echo amplitude exceeds the amplitude anomaly threshold, it is marked as a potential defect area and triggers the phased array transducer fine scanning. Or when the A-scan waveform peak value between the surface echo occurrence time and the bottom echo occurrence time is greater than the defect echo determination threshold, it is marked as a potential defect area and triggers the phased array transducer fine scanning.

[0009] As a further scheme of the application, the phased array transducer fine scanning includes: calculating the defect depth estimation value based on the defect depth estimation formula as follows: h=d×(tb−ta)2; where tb−ta is the difference between the bottom echo occurrence time and the surface echo occurrence time, and d is the longitudinal wave speed of the tested composite material; the center frequency is increased by a preset increase value, and the number of elements of the phased array transducer is halved.

[0010] As a further scheme of the present application: in the wavelet packet decomposition, comprising: adopting a typical wavelet to perform denoising processing on the waveform, and retaining main reflection information; only intercepting an effective time window [ta-Δtpre, c+Δtpost] containing surface echo, defect echo and bottom echo, wherein Δtpre and Δtpost are pre-reserved margin thresholds; performing wavelet packet decomposition on the original A-scan waveform in the above effective time window; setting a decomposition layer number L to obtain 2L sub-bands; for each sub-band j, wherein j=1, 2,.., 2L represents a sub-band with different center frequencies; obtaining a corresponding time domain waveform thereof through a wavelet packet reconstruction algorithm, denoted as a reconstructed sub-waveform xj[n], wherein n is a time domain measurement point index of the sub-waveform; and calculating the total energy Ej of each sub-band as the sum of the squares of the amplitudes of all measurement points of the corresponding reconstructed sub-waveform; dividing the total energy Ej of each sub-band by the sum of the total energies of all sub-bands j=12LEj to obtain a normalized relative energy value Mj. and arranging the normalized relative energy values Mj of all sub-bands in order to form a sub-band energy vector Mj=M1, M2,...M2L; wherein 2L is the number of sub-bands.

[0011] As a further scheme of the present application: in the evaluation of the sub-band separability index: based on the known defect type samples prepared in advance, the mean and standard deviation of the energy of each sub-band are calculated; the Fisher criterion is used to define the separability index of the sub-band, specifically: for each sub-band, the energy mean difference between all pairs of defect types is calculated, and then divided by the sum of the energy standard deviations of the two defect types, and finally the results of all pairs of defect types are added to obtain the separability index of the sub-band; then, the energy vector priority list of the sub-bands is formed by sorting all sub-band energy vectors in descending order of the separability index value.

[0012] As a further scheme of the present application: adaptively determining the optimal energy dimension includes: setting a maximum candidate dimension Kmax, and for each dimension k from 1 to the maximum candidate dimension Kmax, sequentially selecting the energy vectors of the first k sub-bands in the energy vector priority list of the sub-bands as the energy vector features; when the k-dimensional energy vector features are subjected to N-fold cross-validation, the validation accuracy is recorded; if the validation accuracy of the current dimension k is insufficiently improved by the previous dimension k-1 by a threshold value ε and appears twice in succession, the cycle is terminated, and the current dimension k is selected as the optimal energy dimension k*. If the above condition is not met before the current dimension k reaches the set maximum candidate dimension Kmax, the set maximum candidate dimension Kmax is taken as the optimal energy dimension k*.

[0013] As a further scheme of the present application: the classification model adopts a dual-channel lightweight deep network structure and takes a composite feature vector as input, and outputs a defect category and a classification confidence, including: The energy-time difference channel input dimension is the optimal energy dimension k plus the time difference feature, and an abstract feature vector FE is output through two layers of 1D convolution; wherein the time difference feature contains the time difference between the surface echo and the defect echo, and the time difference between the surface echo and the bottom echo; The phase channel inputs the phase centroid and the phase dispersion, and maps them to an abstract feature vector Fφ of the same dimension as the abstract feature vector FE through two layers of fully connected network; wherein the phase centroid refers to the weighted average value of the instantaneous phase in the time window containing the defect echo, and the weight is the signal amplitude at the corresponding time; the phase dispersion refers to the deviation of the instantaneous phase in the time window containing the defect echo from the phase centroid, and the result is obtained by squaring these deviations, multiplying them by the corresponding signal amplitude, taking the average, and then taking the square root; The fusion layer element-wise adds the abstract feature vector FE and the abstract feature vector Fφ, obtains an output vector through two layers of fully connected network, and finally outputs a category probability vector through Softmax activation; and each defect category corresponds to a category probability vector; the defect category corresponding to the maximum category probability vector is selected as the defect category determination result, and the category probability vector is recorded as the confidence, denoted as the classification confidence.

[0014] As a further scheme of the present application: after the precision scanning of the phased array transducer is completed, it further includes: after the acquisition of a scan line is completed, the A-scan waveform amplitude of all measurement points of the scan line at the same depth gate is mapped into pixel color, and a C-scan column array is stacked along the scanning direction according to the pixel color, and a C-scan frame image is spliced; wherein the depth gate includes a gate start time and a gate width, and the gate start time is the time corresponding to the estimated defect depth.

[0015] As a further scheme of the present application: the defect detection conclusion generation includes: based on the defect category classification result, a new C-scan frame image is generated to obtain a new superimposed defect category determination result Ci, a classification confidence, and a coordinate C-scan frame image corresponding to the defect category determination result, the A-scan waveform is traced back based on the new C-scan frame image, the defect center depth is calculated, the center coordinate is found, the original A-scan waveform is traced back, and the surface echo occurrence time, the defect echo occurrence time and the bottom echo occurrence time are extracted; Combined with the sound speed and propagation time of the water medium, the propagation time is obtained by multiplying the defect echo occurrence time by the water propagation path ratio minus the surface echo occurrence time; and the calculation is split according to the sound wave propagation path: The distance of sound wave propagation in water is the product of the sound speed of water medium and the propagation time; The distance of sound wave propagation in the tested composite material specimen is the product of the longitudinal wave sound speed of the tested composite material and the propagation time; and half of the sum of the distance of the sound wave propagating in water and the distance of the sound wave propagating in the tested composite specimen is taken as the defect center depth; the defect depth estimation value and its corresponding two-dimensional coordinates are combined with the calculated defect center depth to form a three-dimensional point set; for the three-dimensional point set belonging to the same defect type, a spatial clustering algorithm is used for clustering, and then three-dimensional surface reconstruction is performed, after the reconstruction is completed, the defect volume is calculated by voxel integration summation, and the maximum radial size in the X-Y plane is calculated by the maximum Euclidean distance between any two points in the region and the remaining wall thickness obtained by subtracting the defect center depth and the maximum defect depth from the material thickness; wherein the maximum defect depth is the maximum value in the three-dimensional point set of the same defect type; the above three-dimensional parameters, i.e. the defect volume, the maximum radial size in the X-Y plane, and the remaining wall thickness, are input into the material performance database, and the defect volume and the remaining wall thickness are compared with the set upper limit of the defect volume and the lower limit of the remaining wall thickness to generate a defect detection conclusion; when the defect volume is not greater than the upper limit of the defect volume and the remaining wall thickness is not less than the lower limit of the remaining wall thickness, it is recorded as safe and acceptable; when the defect volume after rounding is the same as the upper limit of the defect volume, or the remaining wall thickness after rounding is the same as the lower limit of the remaining wall thickness, it is recorded as needing repair; when the defect volume is greater than the upper limit of the defect volume, or the remaining wall thickness is less than the lower limit of the remaining wall thickness, it is recorded as unqualified.

[0016] The beneficial effects of the present application are: (1) In the present application, the A-scan waveform sequence is obtained by the phased array transducer, the potential defect area is determined and the fine scanning is triggered by combining the echo amplitude and the time difference, and the scanning efficiency and accuracy are optimized; the multi-sub-band is obtained by wavelet packet decomposition, the Fisher criterion is used to evaluate the separability, and the optimal energy dimension is determined by incremental dimension cross-validation, which breaks through the limitation of single frequency band and accurately distinguishes the defect types corresponding to similar echoes; the optimal energy feature, phase centroid, dispersion and time difference are combined into a composite feature vector, which is input into a double-channel lightweight deep network to improve the classification accuracy and confidence; based on the C-scan frame image, the three-dimensional coordinates are traced back, the defect volume, the remaining wall thickness and other parameters are calculated based on the sound velocity, the safe acceptable, the repair required or the unqualified conclusion is output by comparing the material performance database, the closed loop from detection to evaluation is realized, the reliable basis is provided for the quality control of the composite material, and the defect misjudgment problem caused by the similarity of echo waveform is solved, and the detection closed loop optimization is realized. BRIEF DESCRIPTION OF DRAWINGS

[0017] The present application will be further described below with reference to the accompanying drawings.

[0018] Figure 1 is a method flow diagram of a non-destructive testing method for composite material defects of the present application; Figure 2 is an ultrasonic detection schematic diagram of a non-destructive testing method for composite material defects of the present application; Figure 3 is the implementation logic diagram of step two in the composite material defect nondestructive testing method. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0020] Embodiment one In this embodiment, the carbon fiber composite laminate with three different embedded inclusion defects (polytetrafluoroethylene film, CFRP sheet and plastic paper) is taken as the detection object, the size of the plate is 200 mm x 150 mm x 2.5 mm, and the laying mode is ±45° / 0° / ±45°4. The detection system includes a constant-temperature water tank, a transceiving phased array transducer, a mechanical scanning device, etc. Please refer to Figure 1 The present application is a composite material defect nondestructive testing method, which comprises the following steps: Step one: under the condition of a constant-temperature water tank, a phased array transducer is used to scan along the X-Y path, the host computer transmits pulses and collects A-scan waveforms; potential defects are determined based on the amplitude / time difference comparison of surface echo, defect echo and bottom echo with threshold values; the focal length table is called for dynamic focusing according to the defect depth estimation value of the potential defect area, and the aperture and center frequency are adaptively adjusted; the continuous A-scan waveforms are colored to generate C-scan frames according to the depth gate, and the next row scanning interval is adaptively adjusted according to the defect length threshold value; the data set containing coordinates, A-scan waveforms and focusing / gate records is outputted; In this embodiment, a phased array transducer with a center frequency of f0 is selected, and the initial value of f0 can be 5MHz, wherein N=32 and the phased array transducer element spacing P is set to 0.5mm; the water tank temperature is maintained at 25℃±1℃ to stabilize the water medium sound speed; the automatic scanning device is used to control the movement of the phased array transducer in the X-Y plane, and the initial scanning interval is set to 1mm; Please refer to Figure 2As shown, a voltage excitation signal with a pulse width of 0.2 μs and a voltage amplitude of 200 V is transmitted by a dedicated phased array host to the phased array transducer, and is incident on the tested composite specimen through the water medium; based on the difference in acoustic impedance between water and the tested composite specimen, in a typical defect-free area, the original A-scan waveform will have two main reflection peaks: the surface echo near zero time, the occurrence time is recorded as ta, which is the signal reflected by the sound wave through the surface of the target defect detection laminate; the bottom echo far from zero time, the occurrence time is recorded as tc, which is the signal reflected by the sound wave through the bottom of the target defect detection laminate; if there is an interface (such as delamination, porosity or inclusion) in the material, a defect echo will appear between ta and tc, recorded as tb, the amplitude and occurrence time of which can preliminarily reflect the position and size of the defect; It should be noted that the A-scan waveform is the time-domain echo signal curve recorded after the phased array transducer transmits and receives once at a single measurement point; it only represents the reflection information of the point along the acoustic path; The A-scan waveform is sampled at a rate of 100 MHz, and the total sampling length is set to 4 μs; after acquisition, the signal is band-pass filtered at 1~10 MHz to improve the signal-to-noise ratio; the following calibrations need to be completed before detection: 30 groups of A-scan waveforms are collected in the defect-free area of the tested composite specimen, and the average amplitude of the surface echo, the average amplitude of the bottom echo, and the noise standard deviation of the defect-free area are calculated; When the phased array transducer completes a scanning line along the X direction at the initial scanning pitch, the A-scan waveform of each measurement point is analyzed in real time; if any of the following conditions is met, it is marked as a potential defect area and triggers the phased array transducer for fine scanning: When the absolute deviation of the surface echo amplitude from the average surface echo amplitude exceeds the amplitude anomaly threshold, or the absolute deviation of the bottom echo amplitude from the average surface echo amplitude exceeds the amplitude anomaly threshold, it is marked as a potential defect area; Or between the occurrence time of the surface echo and the occurrence time of the bottom echo, the peak value of the A-scan waveform is greater than the defect echo determination threshold, which is marked as a potential defect area; Wherein, the amplitude anomaly threshold value can be 4.0 as its initial value through the set first proportion coefficient, which can be adjusted according to the actual scene; the amplitude anomaly threshold value is obtained by multiplying the set first proportion coefficient by the noise standard deviation of the defect-free area; The noise mean value is 0 after zero offset correction of the noise signal, and the noise standard deviation of the defect-free area is obtained by repeated statistics; based on the set second proportion coefficient, its initial value can be 3.0, which can be adjusted according to the actual scene; the defect echo determination threshold is obtained by multiplying the set second proportion coefficient by the noise standard deviation of the defect-free area; When the phased array transducer moves to the preset neighborhood radius of the potential defect area, which can be 2mm, the following fine scanning operation is performed: calculate the appropriate measurement point focusing parameters according to the current estimated defect depth, and use the dynamic focusing strategy to improve the resolution, specifically: The center frequency is increased by a preset increase value, which can be 2MHz, and the number of array elements is halved, that is, the array element aperture is adjusted from 32 array elements to 16 array elements, thereby reducing the sound beam diameter and increasing the lateral resolution; Calculate the defect depth estimate value h based on the defect depth estimation formula, that is, the time difference between the defect echo and the surface echo is divided by 2 and then multiplied by the longitudinal wave speed of the tested composite material, and the result is recorded as the defect depth estimate value; The formula is as follows: h=d×(tb−ta)2; In the formula, d is the longitudinal wave speed of the tested composite material, and the initial value of CFRP is set to 5500m / s; Through the time difference of the bottom echo of the defect-free area, the online self-calibration is d=2etb−ta, and e is the known material thickness; At the same time, call the experimentally calibrated focal length lookup table to rewrite the delay line parameters based on the defect depth estimate value, so that the sound beam is focused on the center of the defect area; Wherein, For the abnormal area, that is, the area that does not trigger the condition, keep the N array element configuration, that is, 32 array elements, and set the initial center frequency f0 to 5MHz to improve the scanning efficiency; When a scan line is collected, map the A-scan waveform amplitude of all measurement points of the scan line at the same depth gate to pixel color, stack the pixel color along the scanning direction to form a C-scan column array, and splice to form a C-scan frame image; Wherein, the depth gate includes gate start time and gate width, and the gate start time is the time corresponding to the defect depth estimate value, and the gate width initial value is 0.3μs, which can be automatically adjusted according to the material, sound speed and resolution requirement; And generate a C-scan frame image in the cache area every time a scan line is completed, and then obtain the defect length of the current defect by multiplying the pixel number of the defect area in the C-scan frame image by the scanning pitch; And according to the current defect length, dynamically adjust the scanning pitch of the next row: If the defect length is less than or equal to a preset first length threshold; Wherein, the preset first length threshold is initially set to 3mm, the scanning pitch of the next row remains the initial pitch, that is, 1mm; If the defect length is greater than the preset first length threshold and less than a preset second length threshold; Wherein, the preset second length threshold is initially set to 5mm; Then the scanning pitch of the next row is linearly reduced from the initial pitch to the specified minimum scanning pitch by linear interpolation; Wherein, the specified minimum scanning pitch can be set to 0.2mm according to actual requirements; If the defect length is greater than or equal to the preset second length threshold; the next line scanning interval is the initial interval multiplied by a preset scaling factor; the preset scaling factor initial value is 0.5; Before the next line scanning, the focusing depth table is updated according to the current defect depth interval to ensure that the sound beam is focused on the target depth range; At each measurement point, in addition to storing the original A-scan waveform, a gate range is set according to the phased array transducer focusing depth; the average amplitude within the gate is calculated and mapped to a color table to obtain a C-scan array color value; the change of this value is used to indicate the probability of potential defects at the next measurement point, further guiding dynamic focusing and interval adjustment; for example, set the gate width to 0.3 μs and the gate start time to the time corresponding to the estimated depth of the defect; calculate the average amplitude within the gate and map it to a color table to obtain a C-scan array color value; the change of this value is used to indicate the probability of potential defects at the next measurement point, thereby further guiding dynamic focusing and interval adjustment.

[0021] At this point, step one finally outputs: a sequence of A-scan waveforms corrected by dynamic focusing, which contains the spatial coordinates of the A-scan waveforms, the original A-scan waveforms, and the focusing parameters of each measurement point including delay line parameters, element aperture, and center frequency, providing complete context information for subsequent analysis.

[0022] Please refer to Figure 3 As shown, step two: zero offset correction and normalization of the A-scan waveform sequence from step one, and extraction of the effective time window containing the surface echo, defect echo, and bottom echo; wavelet packet decomposition and calculation of sub-band normalized energy; evaluation of separability using Fisher criterion and sorting; adaptive determination of optimal energy dimension using incremental dimension cross-validation; combination of selected optimal energy dimension features with phase centroid, phase dispersion, and time difference features of corresponding waveforms in the A-scan waveform sequence to form a composite feature vector set; The original A-scan waveforms obtained from the A-scan waveform sequence output in step one have problems such as noise and baseline drift; first, zero offset correction and amplitude normalization are performed; then, typical wavelets such as Daubechies db8 or bior5.5 are used for denoising processing to preserve the main reflection information; in order to reduce the interference of incident waves and reverberation, only the effective time window [ta−Δtpre, c+Δtpost] containing the surface echo, defect echo, and bottom echo is extracted, where Δtpre and Δtpost are the initial values of the reserved margin threshold, both of which are 0.2 μs and can be automatically adjusted according to the material thickness and sound speed. Through the above processing, the input data quality is guaranteed, making the subsequent feature extraction more stable; Further, the original A-scan waveform within the effective time window is decomposed by wavelet packet to represent the original A-scan waveform on different frequency subbands and analyze the energy distribution of each frequency band in more detail. The initial value of the decomposition layer L is set to 5 layers to obtain 2L subbands. For each subband j, where j = 1, 2,..., 2L represents a subband with a different center frequency, the corresponding time-domain waveform is obtained by a wavelet packet reconstruction algorithm, denoted as reconstructed subwaveform xj[n], where n is the time-domain measurement point index of the subwaveform. The total energy Ej of each subband is calculated as the sum of the squares of the amplitudes of all measurement points of the corresponding reconstructed subwaveform. To eliminate the difference in absolute energy size between subbands, the total energy Ej of each subband is divided by the sum of the total energies of all subbands, i.e., j = 1, 2,..., 2L, to obtain the normalized relative energy value Mj. The subband energy vector formed by the normalized relative energy values Mj reflects the relative distribution of energy in different frequency bands. The normalized relative energy values Mj of all subbands are arranged in order to form the subband energy vector Mj = M1, M2,..., M2L, where 2L is the number of subbands. On this basis, to objectively evaluate the contribution of different subbands to the classification of defect types, based on known defect type samples such as inclusions, delamination, and holes, which are constructed from standard composite material test specimens or historical detection data that have been verified by destructive testing, a training data set is prepared. The mean and standard deviation of the energy of each subband are calculated for each type of defect type sample in the training data set. The Fisher criterion is used to define the separability index of the subband. Specifically, for each subband, the energy mean difference between all pairs of defect types is calculated, and then divided by the sum of the energy standard deviations of the two defect types. Finally, the results of all pairs of defect types are added to obtain the separability index of the subband. The larger the separability index, the better the frequency band can distinguish different defect types. Subsequently, all subband energy vectors are sorted in descending order of the separability index value to form a priority list of subband energy vectors. Based on the above sub-band energy vector priority list, further through the incremental dimension cross validation adaptive selection of optimal energy dimension k*, set the maximum candidate dimension Kmax initial value can be 10, for each dimension k from 1 to the maximum candidate dimension Kmax, sequentially select the energy vector of the first k sub-band in the energy vector priority list of the sub-band as the energy vector feature; after combining these energy vector features with the phase centroid (phase unwrapping after Hilbert transform, weight is signal amplitude), surface echo-defect echo time difference feature, surface echo-bottom echo time difference feature, etc. Fixed feature combination, use classification model (such as support vector machine or simple neural network) to perform N-fold cross validation on the training set, record the verification accuracy; if the verification accuracy of the current dimension k is less than the threshold value ε initial value can be 0.5% than the previous dimension k-1 and appears twice in succession, then terminate the loop, select the current dimension k as the optimal energy dimension k*; if the current dimension k reaches the set maximum candidate dimension Kmax before not meeting the above conditions, then take the set maximum candidate dimension Kmax as the optimal energy dimension k*; This process ensures that the selection of energy dimension comes from data, and the best frequency band combination can be adaptively output for different detection tasks; After that, the processed A-scan waveform is subjected to Hilbert transform to obtain its analytic signal, and then the instantaneous phase is obtained; define the time window containing the defect echo as W, and calculate the phase centroid and phase dispersion in the time window W: The phase centroid refers to the weighted average value of the instantaneous phase in the time window W, and the weight is the signal amplitude at the corresponding time; the phase dispersion refers to the deviation of the instantaneous phase from the phase centroid in the time window W, which is obtained by multiplying the square of these deviations by the corresponding signal amplitude, taking the average, and then taking the square root; wherein the phase centroid reflects the average position of the echo phase, and the phase dispersion reflects the dispersion degree of the phase; phase information is very effective for distinguishing defects of the same energy but different phase types; At the same time, the peak occurrence time of the surface echo, defect echo and bottom echo in the A-scan waveform is automatically detected, and the time difference between the surface echo and the defect echo, and the time difference between the surface echo and the bottom echo are calculated; for the defect-free measurement points, the time difference between the surface echo and the defect echo is filled with 0 or distinguished from the time difference between the surface echo and the bottom echo to avoid information loss; these time differences reflect the relative position of the echo on the time axis, which can assist the energy and phase features to more accurately estimate the defect depth and type; Next, the selected k* optimal energy dimension features are combined with the phase centroid, phase dispersion, and the two time difference features mentioned above to form a composite feature vector set; the vector dimension is k*+4, k*+3 if only one time difference is used, and the vector dimension is determined by the algorithm and written into the system parameter table to maintain the consistency of subsequent model input; after training, if it is found that some energy bands or phase features have limited contribution to classification, they can be further removed or replaced according to the verification results.

[0023] Finally, the outputs of this step include: the optimal energy dimension k*, the composite feature vector set, and the sub-band separability indicators for evaluation; these outputs are used for the classification model training and online recognition in step three; through the above process, the adaptive selection of energy dimension and the composite features formed by the fusion of phase and time difference can effectively amplify the subtle differences between different defects, laying a solid foundation for solving the echo waveform similarity problem.

[0024] Step three: Based on the composite feature vector in step two, a lightweight deep network is constructed for energy-time difference channels and phase channels, and after training / validation / testing, the real-time output of defect categories and classification confidence for each measurement point is obtained; according to the high / low threshold, the reliability, low confidence, or unknown can be determined, and the unknown samples are put into the small sample pool; the pool is full to trigger active learning, merge samples, recalculate the optimal energy dimension, and fine-tune the network; the results are superimposed on the C-scan frame image for two-dimensional visualization, and the confidence distribution is fed back to step one to adjust the interval and frequency adaptively; Using the composite feature vector constructed in step two, real-time defect classification is performed for each scanning measurement point, and relying on the deep network structure and online learning strategy, the recognition accuracy is continuously improved, and the classification results are combined with the C-scan frame image to realize two-dimensional plane visualization and provide accurate plane positioning for subsequent depth calculation; The training data set constructed in step two is called, which contains pre-prepared known defect type samples and has been processed by step two to generate corresponding composite feature vectors; these sample composite feature vectors and their known defect types are bound to form a (composite feature vector, defect type) sample pair, which is divided into training set, validation set, and test set in a ratio of, for example, 7:1:2, wherein the training set is used for model training, the validation set is used for monitoring the training process, and the test set is used for evaluating the final performance; the defect types include, for example, no defect, polytetrafluoroethylene inclusion, CFRP sheet inclusion, plastic paper inclusion, etc. Based on the composite feature vector, a classification model is constructed, which adopts a dual-channel lightweight deep network structure and takes the composite feature vector as input, and outputs the defect category and classification confidence, including: The energy-time difference channel input dimension is the optimal energy dimension k* plus the time difference feature quantity, which depends on the actual feature quantity, and is obtained by two layers of 1D convolution; wherein the convolution kernel number Nc1, Nc2 and kernel length Lc1, Lc2 are parameters, ReLU activation and a pooling layer extracts local patterns, and outputs an abstract feature vector FE; The phase channel input phase centroid and dispersion and other low-dimensional features are mapped to an abstract feature vector Fφ with the same dimension as the abstract feature vector FE by one to two layers of fully connected network; The fusion layer element-wise adds the abstract feature vector FE and the abstract feature vector Fφ, obtains an output vector by two layers of fully connected network, and finally outputs a class probability vector Pi through Softmax activation; wherein i represents the defect class index in the defect type classification task, which is used to distinguish different defect classes; each defect class corresponds to a class probability vector; The defect class corresponding to the maximum class probability vector is selected as the defect class determination result Ci, and the class probability vector Pi is recorded as the classification confidence; And compare the classification confidence with the preset high confidence threshold and the preset low confidence threshold; wherein the preset high confidence threshold can be 0.9, and the preset low confidence threshold can be 0.5; When the classification confidence is greater than the preset high confidence threshold, the classification result is considered reliable; When the classification confidence is between the preset low confidence threshold and the preset high confidence threshold, it is marked as low classification confidence; When the classification confidence is less than the preset low confidence threshold, it is marked as an unknown sample and stored in the few-sample pool for subsequent learning; It should be noted that the number of layers, the size of the convolution kernel and the number of fully connected neurons of the network are set as adjustable initial values in the parameter table. This structure can simultaneously process multi-dimensional energy / time difference information and a small amount of phase information, and balance model complexity and real-time performance; After classification, according to the gate setting of step one, the amplitudes within the gate corresponding to the A-scan waveform are normalized, mapped to pixel values according to the color table, and these pixel values are arranged in the order of the measurement points to form a C-scan column array of the current scan line. The new C-scan frame image is generated by vertically splicing the C-scan column array with the previous frame C-scan column array, and the defect class determination result Ci, the classification confidence and the coordinates corresponding to the defect class determination result are superimposed on the image, and different colors or symbols are used to distinguish different defects, so that the operator can directly see the position, range and type of the defects on the plane; When low classification confidence or unknown samples are detected, their corresponding abstract feature vectors, A-scan waveforms and tentative defect class determination results are put into the few-sample pool, and the administrator or expert can periodically review the few-sample pool and manually annotate the unknown samples; Whenever the few-sample pool accumulates to a certain number, for example, 50 samples, the active learning process is automatically triggered: the few-sample pool is combined with the original training set in step two, the corresponding composite feature vector is generated for the new sample, that is, the composite feature vector is updated and the artificially labeled label is bound, that is, the defect category judgment result is updated, the energy dimension analysis of step two is re-executed to update the optimal energy dimension k* and the abstract feature vector; if the new sample introduces a new frequency mode, the network input layer size is adjusted according to the new optimal energy dimension, the combined (composite feature vector, defect type) sample pair is retrained or fine-tuned, the new model is tested on the validation set, and if the performance is improved, the old model is replaced, and if the performance is decreased, it can be rolled back. Active learning ensures that the model continuously adapts to new defect types or material changes, improving the long-term stability and generalization ability of the system. The classification result and the confidence are also used to feed back the scanning strategy in step one. For example, when a certain area with unknown samples appears in a plurality of consecutive scanning lines, the scanning pitch can be automatically reduced, the excitation frequency can be increased, and more detailed scanning can be performed in the area. For areas without defects or with high recognition confidence, the pitch can be appropriately increased to improve the detection efficiency. In addition, the classification result can also be used as early warning information. If a certain type of defect (such as delamination) frequently occurs, the system can prompt maintenance personnel to check the manufacturing process or material batch. It should be noted that the network is trained using the cross-entropy loss function and optimization algorithms such as Adam. Batch normalization, dropout, and other techniques are used during the training process to prevent overfitting. The model performance is monitored using a validation set, and an early stopping strategy is used to stop training when the validation loss no longer decreases. For imbalanced samples, class weights or data augmentation strategies can be used to balance the influence of each type of sample during training. After training, the recognition accuracy, recall rate, and other indicators are evaluated on the test set, and the best weights are saved. The above training process is a mature technology, and the specific implementation process will not be described here.

[0025] Step four: The center coordinates of each new C-scan frame image are taken, the surface echo occurrence time, defect echo occurrence time, and bottom echo occurrence time are extracted by backtracking the A-scan waveform, the depth of the defect center is calculated by combining the water / material sound speed and the sound path, a three-dimensional point set is formed, the defects are clustered by DBSCAN according to the defect type, the alpha shape is reconstructed and the volume is integrated by voxel, and the X-Y maximum radial size and the remaining wall thickness are calculated; the parameters are compared with the defect volume upper limit and the remaining wall thickness lower limit of the material performance database to output qualified, repair, and unqualified conclusions, and the coordinates and threshold adjustment instructions are fed back to step one to realize a closed loop; For each new C-scan frame image in step three, the center coordinates are found, and the surface echo occurrence time, defect echo occurrence time, and bottom echo occurrence time are extracted by backtracking the original A-scan waveform recorded in step one; Combined with the speed of sound in water medium, which is about 1480 m / s at 25℃, and the propagation time, which is real-time calibrated with water tank temperature and the time difference between the occurrence time of defect echo and the occurrence time of surface echo multiplied by the proportion of propagation path in water, the propagation distance of sound wave in water is calculated according to the propagation path of sound wave; The propagation distance of sound wave in water is the product of the speed of sound in water medium and the propagation time; The propagation distance of sound wave in the tested composite material sample is the product of the longitudinal wave speed of the tested composite material and the propagation time; And half of the sum of the propagation distance of sound wave in water and the propagation distance of sound wave in the tested composite material sample is taken as the center depth of the defect; Based on the estimated value of defect depth obtained in step one, the two-dimensional coordinates (X, Y) of the defect are combined with the calculated center depth of the defect to form a three-dimensional point set. For the point set belonging to the same defect type (the classification result of step three) At the same time, the depth gate setting is optimized according to the defect depth range: the gate width is set to 1.5 times the pulse width of the defect echo, ensuring that the complete echo is included and no extra noise is introduced, and the starting time of the gate corresponds to the estimated time of the defect depth, which is consistent with the gate setting logic of step one; The two-dimensional coordinates (X, Y) of each measurement point are combined with the calculated depth value to form a three-dimensional point set Pᵢ=(Xᵢ,Yᵢ,Zᵢ). For the three-dimensional point set belonging to the same defect type, a spatial clustering algorithm such as DBSCAN is used to segment independent defect regions, and then a three-dimensional surface reconstruction is performed through α shape reconstruction, where the value of α is the average distance of the point cloud multiplied by 1.5. The reconstruction parameters such as radius and α value are automatically adjusted according to the density of the point cloud. After the reconstruction is completed, the defect volume is calculated by voxel integration summation, and the maximum radial size in the X-Y plane is calculated by the maximum Euclidean distance between any two points in the region, and the remaining wall thickness is obtained by subtracting the defect center depth and the maximum defect depth from the material thickness. The maximum defect depth is the maximum value of the depth in the three-dimensional point set of the same defect type. For complex defects, additional shape features such as ellipticity and flatness are calculated as supplementary parameters for strength evaluation; The above three-dimensional parameters, i.e. defect volume, maximum radial size in X-Y plane, and remaining wall thickness, are input into the material performance database, and the defect volume and remaining wall thickness are compared with the set upper limit of defect volume and lower limit of remaining wall thickness to generate a defect detection conclusion. The upper limit of defect volume and the lower limit of remaining wall thickness are adjusted according to the material category and working condition. For example, the lower limit of remaining wall thickness for aerospace composite materials is set to 60% of the design thickness, and for automotive materials, it is set to 50%. Based on the comparison results of defect volume and remaining wall thickness with the upper limit of defect volume and the lower limit of remaining wall thickness, an evaluation conclusion is given: It should be noted that the material performance database at least includes {material brand, design thickness, working condition level, working condition level, defect volume upper limit, remaining wall thickness lower limit}; the defect volume upper limit and the remaining wall thickness lower limit are obtained by taking {the same brand, the same working condition level} as the retrieval condition; and the defect volume and the remaining wall thickness are compared; When the defect volume is not greater than the defect volume upper limit and the remaining wall thickness is not less than the remaining wall thickness lower limit, it is recorded as safe and acceptable; When the defect volume after rounding is the same as the defect volume upper limit, or the remaining wall thickness after rounding is the same as the remaining wall thickness lower limit, it is recorded as needing repair; When the defect volume is greater than the defect volume upper limit, or the remaining wall thickness is less than the remaining wall thickness lower limit, it is recorded as unqualified; The defect detection conclusion is fed back to the previous step to form a closed loop: the scanning interval of the unknown sample dense area is reduced to 0.5mm, the center frequency is increased to 7MHz, and the interval of the high confidence area is increased to 2mm; for the defect position corresponding to the safe and acceptable, the frequency and position of the defect are recorded, and the manufacturing execution system is fed back to the production line to prompt the adjustment of the corresponding process, such as strengthening the cleaning control of a certain layer when inclusions frequently occur in the layer; the mold temperature is checked in the layered and concentrated area.

[0026] In this embodiment, the A-scan waveform of the test piece is collected, the potential defect area is determined based on the echo amplitude and time difference, and the A-scan waveform sequence containing spatial coordinates, original waveform and measurement point focusing parameters is output; after waveform preprocessing, the effective time window is intercepted, a plurality of sub-bands are generated through wavelet packet decomposition, the total energy of each sub-band is calculated and normalized to form a sub-band energy vector; based on the known defect sample, the Fisher criterion is used to evaluate the sub-band separability index and sort; the optimal energy dimension is adaptively determined through dimension cross-validation, the selected sub-band energy features are combined with phase centroid, phase dispersion and time difference features to form a composite feature vector; the double-channel lightweight deep network classification model is input, and the defect category and confidence are output and mapped to the C-scan frame image; based on the image backtracking three-dimensional coordinates, the defect volume and the remaining wall thickness are calculated in combination with the sound velocity, and the detection conclusion is output by comparing the material performance database; the present application solves the problem of defect misjudgment caused by echo waveform similarity, and realizes detection closed loop optimization.

[0027] The above formulas are all dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula closest to the actual situation, and the preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0028] The above embodiments can be realized all or partially by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized in the form of a computer program product all or partially.

[0029] Those skilled in the art can understand that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application of the technical solution and the constraints of the invention. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0030] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0031] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0032] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A nondestructive testing method for composite material defects, characterized in that: The method comprises the following steps: scanning a composite material test piece with a phased array transducer to obtain an A-scan waveform at each measuring point, determining a potential defect area based on the echo amplitude and echo time difference between various echoes in the A-scan waveform, and outputting an A-scan waveform sequence including spatial coordinates, an original A-scan waveform, and measuring point focusing parameters; wherein the A-scan waveform is a time domain echo signal acquired by scanning with the phased array transducer; wherein the measuring point focusing parameters include delay line parameters, array element aperture, and center frequency; After preprocessing the original A-scan waveform, an effective time window containing surface echoes, defect echoes, and bottom echoes is extracted. Multiple sub-bands are obtained through wavelet packet decomposition, and the total energy of each sub-band is calculated and normalized to form the corresponding sub-band energy vector. Based on samples of known defect types, the Fisher criterion is used to evaluate the separability index of each sub-band. The energy vectors of each sub-band are then sorted in descending order of separability index to form a priority list of energy vectors for each sub-band. An increasing dimensionality cross-validation method is adopted to adaptively determine the optimal energy dimension. The sub-band energy vector features corresponding to the selected optimal energy dimension are combined with the phase centroid, phase dispersion, and time difference features to form a composite feature vector. The obtained composite feature vector is input into the classification model to output the defect category and classification confidence, and the output result is mapped to a C-scan frame image with spatial coordinates as the dimension. The three-dimensional coordinates of the defect are traced back based on the C-scan frame image, and the defect volume and remaining wall thickness are calculated in combination with the sound velocity. The defect detection conclusion is output by comparing with the material performance database.

2. The nondestructive testing method for composite material defects according to claim 1, characterized in that: When scanning with a phased array transducer: an N-element phased array transducer is used, where N is the number of elements, the element spacing of the phased array transducer is P, and the center frequency is f0.

3. The nondestructive testing method for composite material defects according to claim 1, characterized in that: Determining potential defect areas includes: calculating the mean surface echo amplitude, the mean bottom echo amplitude, and the standard deviation of noise in the defect-free area in the A-scan waveform; When the phased array transducer completes a scan line along the X direction at the initial scan pitch, the A-scan waveform of each measurement point is analyzed in real time. If any of the following conditions are met, the area is marked as a potential defect and the phased array transducer fine scan is triggered: When the absolute deviation between the surface echo amplitude or the bottom echo amplitude and the surface echo amplitude mean exceeds the amplitude anomaly threshold, it is marked as a potential defect area and the phased array transducer fine scanning is triggered; Or if the peak value of the A-scan waveform is greater than the defect echo judgment threshold between the surface echo and the bottom echo, it is marked as a potential defect area and triggers the phased array transducer to perform a fine scan.

4. The nondestructive testing method for composite material defects according to claim 3, characterized in that: Phased array transducer fine scanning includes: calculating the defect depth estimate based on the defect depth estimation formula as follows: h = d × (tb-ta)²; where tb-ta is the difference between the appearance time of the bottom echo and the appearance time of the surface echo, and d is the longitudinal wave speed of the tested composite material; increasing the center frequency to the initial center frequency value plus a preset increase value, and halving the number of array elements of the phased array transducer.

5. The nondestructive testing method for composite material defects according to claim 1, characterized in that: The wavelet packet decomposition includes: using typical wavelets to denoise the waveform and retain the main reflection information; intercepting only the effective time window [ta−Δtpre,c+Δtpost] containing the surface echo, defect echo and bottom echo, where Δtpre and Δtpost are reserved margin thresholds; decomposing the original A-scan waveform within the above effective time window using wavelet packets; setting the number of decomposition layers L to obtain 2L sub-bands; for each sub-band j, where j=1,2,..,2L represents sub-bands with different center frequencies; obtaining its corresponding time domain waveform through the wavelet packet reconstruction algorithm, recorded as the reconstructed sub-waveform xj[n], where n is the time domain measurement point index of the sub-waveform; The total energy Ej of each sub-band is calculated as the sum of the squares of the amplitudes of all measurement points of its corresponding reconstructed sub-waveform; the total energy Ej of each sub-band is divided by the sum of the total energies of all sub-bands j=12LEj to obtain the normalized relative energy value Mj; and the normalized relative energy values ​​Mj of all sub-bands are arranged in order to form a sub-band energy vector Mj=M1,M2,...M2L; where 2L is the number of sub-bands.

6. The nondestructive testing method for composite material defects according to claim 1, characterized in that: When evaluating the subband separability index: based on a pre-prepared sample of known defect types, the mean and standard deviation of each subband's energy are calculated. The Fisher criterion is used to define the subband's separability index. Specifically, for each subband, the mean difference in energy between all pairwise defect types is calculated, then divided by the sum of the energy standard deviations of these two defect types. Finally, these results for all pairwise defect types are added together to obtain the separability index for the subband. Subsequently, all subband energy vectors are sorted from largest to smallest according to their separability index value to form a priority list of subband energy vectors.

7. The nondestructive testing method for composite material defects according to claim 6, characterized in that: Adaptive determination of the optimal energy dimension includes: setting a maximum candidate dimension Kmax, and for each dimension k from 1 to the maximum candidate dimension Kmax, sequentially selecting the energy vectors of the first k sub-bands in the sub-band energy vector priority list as energy vector features; performing N-fold cross-validation on the k-dimensional energy vector features and recording the validation accuracy; if the validation accuracy of the current dimension k is less than a threshold ε compared to the previous dimension k-1 and this occurs twice in a row, the loop is terminated and the current dimension k is selected as the optimal energy dimension k*; If the above conditions are not met before the current dimension k reaches the set maximum candidate dimension Kmax, the set maximum candidate dimension Kmax is taken as the optimal energy dimension k*.

8. The nondestructive testing method for composite material defects according to claim 1, characterized in that: The classification model uses a dual-channel lightweight deep network structure and takes a composite feature vector as input to output defect categories and classification confidence, including: The input dimension of the energy-time difference channel is the optimal energy dimension k* plus the time difference feature. Through two layers of 1D convolution, the output is the abstract feature vector FE. Among them, the time difference feature includes the time difference between the surface echo and the defect echo, and the time difference between the surface echo and the bottom wall echo. The phase channel inputs the phase centroid and phase dispersion, which are mapped to an abstract feature vector Fφ of the same dimension as the abstract feature vector FE through a two-layer fully connected network. The phase centroid refers to the weighted average of the instantaneous phase in the time window containing the defect echo, with the weight being the signal amplitude at the corresponding moment. The phase dispersion refers to the deviation between the instantaneous phase and the phase centroid in the time window containing the defect echo. The square of these deviations is multiplied by the corresponding signal amplitude, averaged, and then squared to obtain the result. The fusion layer adds the abstract feature vector FE and the abstract feature vector Fφ element by element, obtains the output vector through a two-layer fully connected network, and finally outputs the category probability vector through Softmax activation; each defect category corresponds to a category probability vector; the defect category corresponding to the maximum category probability vector is selected as the defect category judgment result, and the category probability vector is recorded as the confidence level as the classification confidence level.

9. The nondestructive testing method for composite material defects according to claim 4, characterized in that: After the phased array transducer fine scanning is completed, it also includes: when a scan line is acquired, the A-scan waveform amplitude of all measuring points on the scan line at the same depth gate is mapped to pixel color, stacked along the scanning direction according to the pixel color to form a C-scan array, and spliced ​​to form a C-scan frame image; wherein, the depth gate includes the gate start time and gate width, and the gate start time is the moment corresponding to the estimated defect depth.

10. The non-destructive testing method for composite material defects according to claim 9, characterized in that: The defect detection conclusion generation includes: based on the defect category classification result, regenerating the C-scan frame image to obtain a new superimposed defect category determination result Ci, the classification confidence and the coordinate C-scan frame image corresponding to the defect category determination result; based on the new C-scan frame image, tracing back the A-scan waveform, calculating the defect center depth, finding its center coordinates, tracing back the original A-scan waveform, and extracting the surface echo appearance time, the defect echo appearance time and the bottom echo appearance time; Combining the water medium sound velocity and propagation time, the propagation time is obtained by subtracting the surface echo appearance time from the defect echo appearance time and multiplying it by the propagation path ratio in water; the calculation is split according to the sound wave propagation path: The distance a sound wave travels in water is the product of the speed of sound in the water medium and the propagation time; The propagation distance of the sound wave in the tested composite material specimen is the product of the longitudinal wave speed of the tested composite material and the propagation time; The distance the sound wave propagates in water and half of the sum of the distance the sound wave propagates in the tested composite material specimen is taken as the depth of the defect center; the defect depth estimation value and its corresponding two-dimensional coordinates are combined with the calculated defect center depth to form a three-dimensional point set; the three-dimensional point set belonging to the same defect type is clustered using a spatial clustering algorithm and then a three-dimensional surface reconstruction is performed. After the reconstruction is completed, the defect volume is calculated by voxel integral summation, and the maximum Euclidean distance between any two points in the area is calculated to obtain the maximum radial dimension of the XY plane and the residual wall thickness obtained by subtracting the defect center depth and the maximum defect depth from the material thickness; wherein the maximum defect depth is the maximum depth in the three-dimensional point set of the same defect type; the above three-dimensional parameters, namely the defect volume, the maximum radial dimension of the XY plane, and the residual wall thickness are input into the material performance database, and the defect volume and the residual wall thickness are compared with the set defect volume upper limit and residual wall thickness lower limit to generate a defect detection conclusion; when the defect volume is not greater than the defect volume upper limit and the residual wall thickness is not less than the residual wall thickness lower limit, it is recorded as safe and acceptable; When the defect volume after rounding is the same as the upper limit of the defect volume, or the remaining wall thickness after rounding is the same as the lower limit of the remaining wall thickness, it is recorded as requiring repair; When the defect volume is greater than the upper limit of the defect volume, or the remaining wall thickness is less than the lower limit of the remaining wall thickness, it is recorded as unqualified.

Images