Composite board detection method and system

Abstract

The invention provides a composite board detection method and system, and relates to the technical field of board detection, and the method comprises the steps: carrying out the ultrasonic scanning of a composite board, dividing the composite board into a plurality of acoustic characteristic layers, building a sound wave propagation model based on a recursive transfer matrix method, and calculating the reflection coefficient and transmission coefficient of each layer interface. A deep learning algorithm is combined with physical constraints to dynamically correct acoustic parameters, the corrected parameters are substituted into a model reconstruction signal, and defects are accurately positioned. And performing grid division on the defect position to analyze stress field distribution, and determining a stress concentration position. And fine scanning is carried out on a stress concentration area by adjusting probe parameters, and high-resolution data is obtained to update defect distribution information. And finally, outputting detection data including defect positions, stress concentration distribution and hazard levels, and realizing accurate detection and evaluation of the defects of the composite board.

Description

TECHNICAL FIELD

[0001] The present application relates to plate detection technology, and in particular to a composite board detection method and system. BACKGROUND

[0002] As a widely used structural material in modern industry, the defect detection of composite boards has always been the focus and difficulty in the field of non-destructive testing. Although the traditional ultrasonic detection method can detect obvious defects in composite boards, for complex multi-layer composite boards, the reflection and transmission of sound waves between different material interfaces will produce complex phenomena, resulting in overlapping and blurred echo signals, making it difficult to accurately distinguish defect signals and structure echoes. At the same time, the acoustic characteristics of each layer of the composite board are often difficult to accurately obtain, and the interface characteristics also affect the propagation of sound waves. These factors together lead to inaccurate defect positioning and insufficient sensitivity of the traditional detection method, especially in the detection of small defects in stress concentration areas.

[0003] The existing technology generally uses empirical parameter setting and fixed scanning strategy, which is difficult to adapt to the complexity of composite board structure and material. Although some advanced methods introduce mathematical models for sound wave propagation analysis, the model parameters are difficult to accurately set, resulting in deviations between theoretical prediction and actual situation. In addition, the existing technology focuses on the position information of defects, but ignores the stress field analysis around the defects, making it difficult to evaluate the potential harm of defects to the structure. With the increasing application of composite boards in key fields such as aerospace and energy, there is an urgent need to develop new detection methods that can accurately detect defects in composite boards and evaluate their harm degree, in order to ensure the safety and reliability of the structure. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a composite board detection method and system, which can solve the problems in the prior art.

[0005] In a first aspect of the embodiments of the present application, a composite board detection method is provided, comprising: scanning the composite board to obtain ultrasonic scan data, dividing the composite board into N acoustic characteristic layers, constructing a sound wave propagation model based on the recursive transfer matrix method, calculating the sound wave transfer coefficient matrix of the N acoustic characteristic layers, and calculating the reflection coefficient and transmission coefficient of each layer interface according to the sound wave transfer coefficient matrix; comparing the measured ultrasonic scan data with the theoretical waveform calculated based on the reflection coefficient and transmission coefficient, and dynamically correcting the acoustic parameters of each acoustic characteristic layer based on the deep learning algorithm and physical constraints; substituting the corrected acoustic parameters into the sound wave propagation model to reconstruct the signal of the ultrasonic scan data, and obtaining the real defect signal according to the real defect signal to determine the defect position coordinates; Grid division is performed at the defect position coordinates, stress field distribution data around the defect is calculated, acoustic characteristic change data of the defect area under the action of the stress field distribution data is obtained, and a stress concentration position is determined; The transmission power and scanning interval of the ultrasonic probe are adjusted according to the stress concentration position, fine scanning is performed, high-resolution data is obtained, and defect spatial distribution information is updated according to the high-resolution data; Output detection data, which includes defect position coordinate information, stress concentration distribution data and defect hazard level.

[0006] Optionally, The composite plate is divided into N acoustic characteristic layers, an acoustic wave propagation model is constructed based on a recursive transfer matrix method, an acoustic wave transfer coefficient matrix of the N acoustic characteristic layers is calculated, and the steps of calculating the reflection coefficient and the transmission coefficient of each layer interface according to the acoustic wave transfer coefficient matrix include: Empirical mode decomposition is performed on the ultrasonic scan data to obtain an intrinsic mode function, a phase function of the intrinsic mode function is calculated to obtain an instantaneous frequency feature, a time-frequency energy distribution matrix is constructed by combining the instantaneous frequency feature and the amplitude square of the intrinsic mode function, an interlayer interface position of the composite plate is determined based on the mutation feature of the time-frequency energy distribution matrix, and the composite plate is divided into N acoustic characteristic layers; Three-dimensional topographic parameters at the interlayer interface position are obtained, a stiffness calculation formula considering interface stress is constructed based on the three-dimensional topographic parameters, and an interface stiffness matrix representing interface contact characteristics is obtained; an ideal interface transfer matrix is constructed based on the wave impedance and thickness of each acoustic characteristic layer, and the interface stiffness matrix is multiplied by the ideal interface transfer matrix to obtain an acoustic wave transfer coefficient matrix; The acoustic wave transfer coefficient matrix is substituted into an acoustic wave propagation model, and the acoustic wave propagation model is constructed based on the material parameters of each acoustic characteristic layer; the wave number of the N acoustic characteristic layers is calculated according to the acoustic wave propagation model, the reflection coefficient and the transmission coefficient of the interlayer interface are calculated based on the wave number, and the calculation of the reflection coefficient and the transmission coefficient includes an interface characteristic correction term.

[0007] Optionally, The measured ultrasonic scan data is compared with the theoretical waveform calculated based on the reflection coefficient and the transmission coefficient, and the steps of dynamically correcting the acoustic parameters of each acoustic characteristic layer based on a deep learning algorithm and physical constraints include: The measured ultrasonic scan data is input into the waveform encoding layer of the bidirectional long short-term memory neural network and encoded into a waveform feature vector; inputting the waveform feature vector into a multi-scale attention mechanism layer, performing wavelet transform on the waveform feature vector to obtain a time-frequency feature matrix, constructing a local attention vector and a global attention vector based on the time-frequency feature matrix, adaptively fusing the local attention vector and the global attention vector to obtain a fused attention vector, and performing element-wise multiplication operation on the fused attention vector and the waveform feature vector to obtain a weighted feature vector; inputting the weighted feature vector into an acoustic parameter prediction network, the acoustic parameter prediction network outputting an acoustic parameter set, the acoustic parameter set including a wave velocity parameter, a density parameter and an attenuation coefficient parameter; constructing an optimization objective function based on a multi-level physical constraint function and a waveform reconstruction error; performing self-supervised pre-training on the bidirectional long short-term memory neural network using ultrasonic data of a defect-free area, the self-supervised pre-training using a waveform reconstruction task as a pre-training target; performing iterative optimization on the acoustic parameter set based on the optimization objective function, stopping iteration when the parameter change quantity of continuous multiple iterations is less than a preset change threshold, and obtaining a corrected acoustic parameter set.

[0008] Optionally, The step of constructing an optimization objective function includes: The multi-level physical constraint includes a wave equation constraint term, a parameter range constraint term and an interface continuity constraint term; the waveform reconstruction error includes a mean square error term and a phase error term; and the optimization objective function adopts a stage-by-stage weight configuration strategy based on an iteration process, and dynamically adjusts the weight coefficients of the waveform reconstruction error term and the physical constraint term according to the waveform reconstruction error value, the degree of violation of the physical constraint and the acoustic impedance difference of the material interface.

[0009] Optionally, substituting the corrected acoustic parameters into the acoustic wave propagation model to perform signal reconstruction on the ultrasonic scanning data, obtaining a real defect signal, and determining the defect position coordinates according to the real defect signal includes: substituting the corrected acoustic parameters and the acoustic transmission coefficient matrix into the acoustic wave propagation model, constructing an interlayer multiple reflection compensation factor based on the reflection coefficient of the adjacent interlayer interface and the interlayer acoustic wave propagation time delay, multiplying the interlayer multiple reflection compensation factor by the frequency domain response of the ultrasonic scanning data, and performing inverse Fourier transform to obtain a reconstructed signal; extracting the time domain amplitude feature and the phase feature of the reconstructed signal, performing difference operation on the time domain amplitude feature and the theoretical reflection time calculated based on the corrected acoustic parameters, performing difference operation on the phase feature and the theoretical reflection phase, and identifying a real defect signal according to the difference operation result; establish an acoustic wave propagation time equation based on the corrected acoustic parameters, and solve the acoustic wave propagation time equation according to time characteristics of the real defect signal to obtain defect position coordinates.

[0010] Optionally, grid division is performed at the defect position coordinates, stress field distribution data around the defect is calculated, acoustic characteristic variation data of the defect area under the action of the stress field distribution data is obtained, and the step of determining the stress concentration position comprises: An initial grid is constructed with the defect position coordinates as the center, a stress gradient index is obtained by calculating a square root of a sum of squares of stress partial derivatives in X and Y directions of each grid unit, and grid units with a stress gradient index greater than a preset stress gradient threshold value are refined and divided; Based on the refined and divided grid, a spatial derivative of a displacement component is calculated to obtain a strain value, and the strain value is multiplied by material elastic constants to obtain stress field distribution data; An acoustic velocity variation is calculated according to the stress field distribution data, the acoustic velocity variation is determined by the product of the principal stress components in three directions and the corresponding acoustic coefficient, and an acoustic impedance correction value is calculated according to the acoustic velocity variation; Acoustic characteristics corresponding to the acoustic velocity variation and the acoustic impedance correction value are extracted, a feature vector containing an acoustic velocity variation rate, an acoustic impedance variation rate and a frequency spectrum energy distribution is constructed, a weighted Euclidean distance between feature vectors is calculated, a neighborhood radius is determined according to an average distance of K nearest neighbors around each feature point, the neighborhood radius is used for density clustering of feature points, and a stress concentration position is determined by mapping the clustering result back to a spatial coordinate; A stress concentration coefficient is obtained by calculating a ratio of a maximum stress in the stress concentration position to a nominal stress, a deviation between a measured acoustic characteristic and a theoretical acoustic characteristic is substituted into an exponential decay function to calculate a confidence index, and the reliability of the stress concentration position is determined according to the confidence index.

[0011] Optionally, The transmission power and the scanning interval of the ultrasonic probe are adjusted according to the stress concentration position, fine scanning is performed and high-resolution data is obtained, and the step of updating defect spatial distribution information according to the high-resolution data comprises: An acoustic wave energy attenuation compensation coefficient is calculated according to the depth value and the maximum stress value of the stress concentration position, the acoustic wave energy attenuation compensation coefficient is multiplied by the initial transmission power to obtain an optimized transmission power, and the scanning interval is determined based on the ratio of the ultrasonic wavelength to the probe diameter; construct a spiral scanning track with the stress concentration position as the center, the spiral scanning track being determined by an angle parameter and a radius parameter, and the angle parameter and the radius parameter being adjusted according to the scanning interval so that the spiral scanning track completely covers the stress concentration position; and performing ultrasonic scanning along the spiral scanning track to obtain high-resolution scanning data by using the optimized transmission power. calculating a boundary curvature value and a signal-to-noise ratio of the high-resolution scanning data, constructing a weighted fusion coefficient based on the boundary curvature value and the signal-to-noise ratio, fusing defect position coordinate information and fine scanning features by using the weighted fusion coefficient, and updating spatial distribution information of the defect, the spatial distribution information including a defect boundary contour and internal structure features.

[0012] In a second aspect, a composite plate detection system is provided, including: a first unit configured to scan a composite plate to obtain ultrasonic scanning data, divide the composite plate into N acoustic characteristic layers, construct an acoustic wave propagation model based on a recursive transfer matrix method, calculate an acoustic wave transfer coefficient matrix of the N acoustic characteristic layers, and calculate reflection coefficients and transmission coefficients of interfaces of each layer according to the acoustic wave transfer coefficient matrix; a second unit configured to compare actual ultrasonic scanning data with theoretical waveforms calculated based on the reflection coefficients and the transmission coefficients, and dynamically correct acoustic parameters of each acoustic characteristic layer based on a deep learning algorithm and physical constraints; a third unit configured to substitute the corrected acoustic parameters into the acoustic wave propagation model, reconstruct signals of the ultrasonic scanning data to obtain real defect signals, and determine defect position coordinates according to the real defect signals; a fourth unit configured to divide a grid at the defect position coordinates, calculate stress field distribution data around the defect, obtain acoustic feature change data of a defect region under the action of the stress field distribution data, and determine a stress concentration position; a fifth unit configured to adjust transmission power and scanning interval of an ultrasonic probe for the stress concentration position, perform fine scanning to obtain high-resolution data, and update defect spatial distribution information according to the high-resolution data; a sixth unit configured to output detection data, the detection data including defect position coordinate information, stress concentration distribution data, and defect hazard levels.

[0013] In a third aspect, a computer readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0014] The acoustic parameter correction process not only follows the physical law but also has data-driven adaptability, which significantly improves the accuracy and reliability of composite plate defect detection. In particular, by establishing the mapping relationship between acoustic characteristics and stress field distribution, not only can the defect be accurately located, but also the stress concentration around the defect can be analyzed, providing a scientific basis for defect hazard assessment and filling the gap of traditional detection methods in defect hazard assessment.

[0015] This paper also introduces an adaptive scanning strategy based on stress concentration area, which realizes fine detection of high-risk areas by dynamically adjusting ultrasonic probe parameters and scanning trajectory, greatly improving the detection resolution and sensitivity. The overall technical process of the method forms a complete system from acoustic propagation model construction, parameter optimization, signal reconstruction to stress analysis and hazard assessment, providing new ideas and methods for the field of composite plate nondestructive testing. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a composite plate detection method according to an embodiment of the present application is shown in Figure 2 A comparison chart of waveform reconstruction error changes with iteration number for different methods is shown in DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described below with reference to the drawings in the embodiments of the present application. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0018] Figure 1 A flowchart of a composite plate detection method according to an embodiment of the present application is shown in Figure 1 As shown in the figure, the method comprises: The composite plate is scanned to obtain ultrasonic scan data, the composite plate is divided into N acoustic characteristic layers, an acoustic propagation model is constructed based on the recursive transfer matrix method, the acoustic transfer coefficient matrix of the N acoustic characteristic layers is calculated, and the reflection coefficient and transmission coefficient of each layer interface are calculated according to the acoustic transfer coefficient matrix; The measured ultrasonic scan data is compared with the theoretical waveform calculated based on the reflection coefficient and transmission coefficient, and the acoustic parameters of each acoustic characteristic layer are dynamically corrected based on the deep learning algorithm and physical constraints; The corrected acoustic parameters are substituted into the acoustic propagation model, the ultrasonic scan data is reconstructed, the real defect signal is obtained, and the defect position coordinates are determined according to the real defect signal. grid division is performed on the defect position coordinates, stress field distribution data around the defect is calculated, acoustic characteristic variation data of the defect area under the action of the stress field distribution data is obtained, and a stress concentration position is determined; The transmission power and scanning interval of the ultrasonic probe are adjusted according to the stress concentration position, fine scanning is performed, high-resolution data is obtained, and defect spatial distribution information is updated according to the high-resolution data; Output detection data, which includes defect position coordinate information, stress concentration distribution data and defect hazard level. The product of the stress concentration coefficient and the defect volume is normalized to a quantitative index of 0-10, and the defect is divided into three levels of low hazard, medium hazard and high hazard according to the quantitative index.

[0019] Optionally, The composite plate is divided into N acoustic characteristic layers, an acoustic wave propagation model is constructed based on a recursive transfer matrix method, an acoustic wave transfer coefficient matrix of the N acoustic characteristic layers is calculated, and the steps of calculating the reflection coefficient and the transmission coefficient of each layer interface according to the acoustic wave transfer coefficient matrix include: The intrinsic mode function is obtained by empirical mode decomposition of the ultrasonic wave scanning data, the phase function of the intrinsic mode function is calculated to obtain the instantaneous frequency feature, the time-frequency energy distribution matrix is constructed by combining the instantaneous frequency feature and the amplitude square of the intrinsic mode function, the interlayer interface position of the composite plate is determined based on the mutation characteristics of the time-frequency energy distribution matrix, and the composite plate is divided into N acoustic characteristic layers; The three-dimensional topographic parameters at the interlayer interface position are obtained, the stiffness calculation formula considering the interface stress is constructed based on the three-dimensional topographic parameters, and the interface stiffness matrix representing the interface contact characteristics is obtained; the ideal interface transfer matrix is constructed based on the wave impedance and thickness of each acoustic characteristic layer, and the acoustic wave transfer coefficient matrix is obtained by multiplying the interface stiffness matrix and the ideal interface transfer matrix; The acoustic wave transfer coefficient matrix is substituted into the acoustic wave propagation model, the acoustic wave propagation model is constructed based on the material parameters of each acoustic characteristic layer, and is used to describe the propagation law of the acoustic wave in the composite plate; the wave number of the N acoustic characteristic layers is calculated according to the acoustic wave propagation model, the wave number represents the propagation characteristics of the acoustic wave in each layer, the reflection coefficient and the transmission coefficient of the interlayer interface are calculated based on the wave number, and the calculation of the reflection coefficient and the transmission coefficient includes an interface characteristic correction term, the interface characteristic correction term is an exponential function of the arithmetic average of the normal stiffness coefficient and the three-dimensional topographic parameters.

[0020] An exemplary C-scan of a composite plate is performed using an ultrasonic probe with a specific step size of 0.5 mm, a scanning frequency of 5 MHz, a sampling rate of 100 MHz, and a signal length of 1024 sampling points. The obtained raw scanning data is presented as a set of time-domain waveforms, which contains the reflection information of the internal structure and potential defects of the composite plate.

[0021] The ultrasonic scanning data is processed using empirical mode decomposition, which extracts intrinsic mode functions from the original signal through repeated screening and iteration. For example, the scanning data of a carbon fiber composite plate can be typically decomposed into 5-8 intrinsic mode functions, each representing a certain frequency component of the signal. For a composite plate with a thickness of 10 mm, it is usually necessary to perform 10 screening iterations to ensure the accuracy of the decomposition. The Hilbert transform is applied to the intrinsic mode functions to obtain the analytical signal, and then the first derivative of the phase function is calculated to obtain the instantaneous frequency. Taking a scanning case of a composite plate as an example, the calculated instantaneous frequency of the first intrinsic mode function shows a jump feature from 4.8 MHz to 4.2 MHz at the interface.

[0022] The time-frequency energy distribution matrix is constructed as follows: for each time point t, take the square of the amplitude of the ith intrinsic mode function at that time point as the weight, multiply it by the corresponding instantaneous frequency value, and fill it into the tth row and ith column of the matrix. In specific implementation, first calculate the square of the amplitude of each intrinsic mode function at each time point, then multiply the square of the amplitude at the same time point by the corresponding instantaneous frequency value to form the energy distribution matrix. For a typical case of an 8-layer composite plate, after obtaining a 1024x8 matrix, normalize it by dividing by the maximum value in the matrix.

[0023] The interlayer interface position of the composite plate is determined based on the sudden change feature of the time-frequency energy distribution matrix. In practice, the energy change threshold is set to 20%, and when the energy distribution changes at a rate exceeding the threshold on the time axis, it is identified as a potential interface position. For an 8-layer carbon fiber reinforced epoxy resin composite plate, typically 7 obvious interface positions can be detected, with time positions at 23 μs, 35 μs, 48 μs, 60 μs, 72 μs, 85 μs and 97 μs of the original signal. The time positions are converted to depth positions in combination with the sound velocity information to determine the thickness of each layer, and the N-layer division of the composite plate is realized.

[0024] The interlayer interface of the composite panel is sampled and measured by a surface roughness measuring instrument, and parameters such as peak height, valley depth and average roughness of the interface are recorded. Taking the third interface of a certain composite panel as an example, the average roughness measured is 2.3 μm, the peak-to-valley difference is 8.5 μm, and the standard deviation of the sampling point is 1.2 μm. Based on the three-dimensional topographic parameters, a stiffness calculation formula considering the interface stress is constructed. In practical application, the interface stiffness is positively correlated with the interface contact area, the material elastic modulus and the interface stress. For an interface with an average roughness of 2.3 μm, when the interface pressure is 5 MPa, the calculated normal stiffness coefficient is about 2.8×10 14 N / m 3 , and the tangential stiffness coefficient is about 9.3×10 13 N / m 3 .

[0025] The construction of the interface stiffness matrix needs to consider the stiffness characteristics in the normal and tangential directions. For a typical composite panel interface, the matrix dimension is 2×2, the diagonal elements are the normal and tangential stiffness coefficients respectively, and the non-diagonal elements are zero in ideal cases. In actual calculation, the non-diagonal elements can be appropriately adjusted according to the anisotropic characteristics of the interface, and the value is usually not more than 5% of the diagonal elements.

[0026] Based on the wave impedance and thickness of each acoustic characteristic layer, an ideal interface transfer matrix is constructed. For carbon fiber composite panels, the typical acoustic impedance value ranges from 6×10 6 kg / (m 2 ·s) to 9×10 6 kg / (m 2 ·s). Taking an 8-layer composite panel as an example, the acoustic impedance of each layer from top to bottom is 7.2, 8.5, 8.3, 7.8, 8.1, 8.6, 7.9 and 8.2 (×10 6 kg / (m 2 ·s) respectively, and the thickness of each layer is 1.2, 1.3, 1.25, 1.3, 1.2, 1.35, 1.3 and 1.1 millimeters respectively.

[0027] The interface stiffness matrix is multiplied by the ideal interface transfer matrix to obtain the sound wave transfer coefficient matrix. For each interface, the calculated transfer coefficient matrix has a dimension of 4×4, which includes the reflection, transmission and mode conversion characteristics of sound waves at the interface. In actual calculation, the main diagonal elements of the transfer coefficient matrix of the third interface have typical values of 0.92 and 0.88, indicating high transmission efficiency, while the non-diagonal elements have typical values less than 0.15, indicating weak mode conversion effect.

[0028] The acoustic transmission coefficient matrix is substituted into an acoustic wave propagation model, which is constructed based on the material parameters of each acoustic property layer. The model input parameters include the longitudinal wave velocity (typical value 5900-6500 m / s), transverse wave velocity (typical value 3100-3400 m / s), density (typical value 1550-1650 kg / m 3 ) and attenuation coefficient (typical value 20-35 Np / m) of each layer. For an incident wave of a specific frequency (such as 5 MHz), the model can calculate the propagation characteristics of the acoustic wave in each layer and the reflection / transmission behavior at the interface.

[0029] The wave number of each acoustic property layer is calculated according to the acoustic wave propagation model, which represents the propagation characteristics of the acoustic wave in each layer. The typical real part of the longitudinal wave number is about 5000-5500 rad / m, and the imaginary part is about 10-20 rad / m; the real part of the transverse wave number is about 9200-9800 rad / m, and the imaginary part is about 25-40 rad / m. The real part of the wave number determines the wavelength, and the imaginary part determines the attenuation characteristics. The reflection coefficient and the transmission coefficient are calculated using a method that takes into account the difference in wave impedance of the materials on both sides of the interface and the stiffness of the interface. For the interface between the mth layer and the m+1th layer, the reflection coefficient is calculated as the ratio of the difference in wave impedance of the materials on both sides to the sum of the wave impedances, multiplied by an interface characteristic correction term; the transmission coefficient is calculated as twice the product of the wave impedances divided by the sum of the wave impedances, multiplied by the interface characteristic correction term. The interface characteristic correction term is determined by an exponential function of the normal stiffness coefficient and the arithmetic mean of the three-dimensional topography parameters, which takes into account the influence of the interface contact state on the transmission of acoustic waves. For the third interface, the calculated longitudinal wave reflection coefficient is about 0.08, and the transmission coefficient is about 0.92; the transverse wave reflection coefficient is about 0.12, and the transmission coefficient is about 0.88.

[0030] The method can accurately identify the interlayer interface in a complex composite plate using empirical mode decomposition and time-frequency analysis methods, overcoming the limitations of traditional methods in handling multi-layer superimposed signals. The acoustic wave propagation model based on the recursive transfer matrix method can efficiently handle wave propagation problems in multi-layer structures, with high computational efficiency and good accuracy.

[0031] Optionally, Comparing the measured ultrasonic scan data with the theoretical waveform calculated based on the reflection coefficient and the transmission coefficient, and dynamically correcting the acoustic parameters of each acoustic property layer based on a deep learning algorithm and physical constraints, the steps include: Inputting the measured ultrasonic scan data into the waveform encoding layer of the bidirectional long short-term memory neural network, which encodes the measured ultrasonic scan data into a waveform feature vector; inputting the waveform feature vector into a multi-scale attention mechanism layer, performing wavelet transform on the waveform feature vector to obtain a time-frequency feature matrix, constructing a local attention vector and a global attention vector based on the time-frequency feature matrix, adaptively fusing the local attention vector and the global attention vector to obtain a fused attention vector, and performing element-by-element multiplication operation on the fused attention vector and the waveform feature vector to obtain a weighted feature vector; inputting the weighted feature vector into an acoustic parameter prediction network, and outputting an acoustic parameter set including a wave velocity parameter, a density parameter, and an attenuation coefficient parameter from the acoustic parameter prediction network; constructing an optimization objective function based on a multi-level physical constraint function and a waveform reconstruction error; performing self-supervised pre-training on the bidirectional long short-term memory neural network using ultrasonic data of a defect-free area, wherein the self-supervised pre-training uses a waveform reconstruction task as a pre-training target; performing iterative optimization on the acoustic parameter set based on the optimization objective function, stopping iteration when a parameter change amount of continuous multiple iterations is less than a preset change threshold, and obtaining a corrected acoustic parameter set.

[0032] For example, the measured ultrasonic scanning data is input into the waveform encoding layer of the bidirectional long short-term memory neural network. The network adopts a four-layer structure, each layer containing 128 hidden units, and using a tanh activation function. The input data is an ultrasonic A-scan signal with a length of 1024, which is preprocessed by time domain normalization and then sent into the network. The waveform encoding layer captures the timing features and contextual relationships of the waveform through information transmission in the forward and backward directions, and finally encodes the measured waveform into a 256-dimensional feature vector. For example, after encoding the ultrasonic signal of a certain carbon fiber composite plate, the high-frequency noise component in the signal is effectively suppressed, and the key features are enhanced. The first 128 dimensions of the feature vector mainly represent low-frequency structural information, and the last 128 dimensions mainly represent high-frequency detail information.

[0033] The waveform feature vector is input into the multi-scale attention mechanism layer. The layer first performs continuous wavelet transform on the waveform feature vector, uses the Mexican hat wavelet as the mother wavelet, sets the scale parameter to 1 to 32 with a step of 1, calculates 32 scales in total, and obtains a time-frequency feature matrix with a size of 256x32. In practical applications, for an ultrasonic signal with a center frequency of 5MHz, the time-frequency feature matrix after wavelet transform has the most concentrated energy in the 4-6MHz frequency band and the 20-60μs time interval, showing obvious interlayer interface features.

[0034] The local attention vector and the global attention vector are constructed based on the time-frequency feature matrix. The local attention calculation process is to divide the time-frequency feature matrix into 4x4 sub-blocks, each sub-block size is 64x8, calculate the energy mean value of each sub-block as the representative value of the sub-block, and then convert the 16 representative values into local attention scores through the softmax function to form a 16-dimensional local attention vector. The global attention calculation process is to calculate the average value of the row direction and the column direction of the entire time-frequency feature matrix to obtain a 256-dimensional row vector and a 32-dimensional column vector. The two vectors are spliced and reduced to 64 dimensions through a fully connected layer, and then the global attention vector is obtained through the softmax function. The local attention vector and the global attention vector are adaptively fused to obtain a fused attention vector. The fusion process is realized through a gating mechanism, and a two-layer perceptron network is used to calculate the fusion weight, the input is the splicing of the local attention vector and the global attention vector, and the output is a fusion coefficient between 0 and 1. When the signal noise is low, the global attention weight is high (typical value is 0.7-0.8); and when the signal contains defect information with obvious local features, the local attention weight will increase accordingly (typical value is 0.6-0.7). The dimension of the fused attention vector is 256, which is consistent with the original waveform feature vector.

[0035] The fused attention vector and the waveform feature vector are multiplied element by element to obtain a weighted feature vector. This step realizes the selective enhancement of the original features, enabling the network to focus on the key information in the waveform.

[0036] The weighted feature vector is input into the acoustic parameter prediction network, which adopts a three-layer fully connected structure with hidden layer sizes of 128, 64 and 32, and a ReLU activation function. The network outputs an acoustic parameter set, including the wave velocity parameter, the density parameter and the attenuation coefficient parameter of each layer. For an 8-layer composite panel, the output dimension is 8x3=24, corresponding to the 3 acoustic parameters of the 8 layers. In practical applications, the wave velocity parameter range predicted by the prediction network is usually 5900-6500 m / s, the density parameter range is 1550-1650 kg / m 3 , and the attenuation coefficient parameter range is 20-35 Np / m.

[0037] An optimization objective function is constructed based on a multi-level physical constraint function and a waveform reconstruction error. The physical constraints include three levels: first, the wave equation constraint ensures that the acoustic parameters satisfy the wave propagation law, with a weight of 0.3; second, the parameter range constraint ensures that the acoustic parameters are within the physically reasonable range, with a weight of 0.2; and finally, the interface continuity constraint ensures that the displacement and stress continuity conditions are met at the layer interface, with a weight of 0.2. The waveform reconstruction error includes the mean square error of the time-domain waveform amplitude (typical weight is 0.2) and the phase error based on Hilbert transform (typical weight is 0.1).

[0038] The optimization objective function adopts a phased weight configuration strategy. In the early stage of iteration (usually the first 50 iterations), the waveform reconstruction error weight is high (0.8), and the physical constraint weight is low (0.2), to quickly approach the measured waveform; when the waveform reconstruction error falls below the threshold (usually 20% of the initial error), the physical constraint weight is increased to 0.5, and the waveform reconstruction error weight is reduced to 0.5, to ensure the physical reasonableness of the solution; when a region with an acoustic impedance difference greater than 30% at the material interface is detected, the waveform reconstruction error weight of the region is increased by 50%, to strengthen the fitting of the abnormal region; when any physical constraint violates the degree exceeds the preset threshold (usually 10% of the initial violation degree), the constraint term weight is doubled until the constraint violation degree falls below the threshold.

[0039] Before applying the deep learning algorithm, the network needs to be pre-trained. The bidirectional long short-term memory network is pre-trained using the ultrasonic wave data of the defect-free region in a self-supervised manner, and the waveform reconstruction task is used as the pre-training target. Specifically, the ultrasonic A-scan signal of the defect-free region is input into the network, the original signal is reconstructed through the encoding-decoding structure, and the training is performed by minimizing the reconstruction error. The pre-training adopts a batch size of 64, a learning rate of 0.001, and 100 training rounds, and uses the Adam optimizer. The pre-trained network can effectively capture the essential features of the ultrasonic signal, laying a foundation for the subsequent parameter prediction task.

[0040] The acoustic parameter set is iteratively optimized based on the optimization objective function, and the gradient descent method is used, with an initial learning rate of 0.01 and a cosine annealing strategy for dynamic adjustment. When the parameter change amount of 5 consecutive iterations is less than the preset change threshold (usually 0.1% of the initial parameter), the iteration is stopped, and the corrected acoustic parameter set is obtained. In typical applications, the optimization process usually converges after 200-300 iterations, and the error between the final acoustic parameters and the actual physical parameters is usually less than 5%.

[0041] The method combines the advantages of deep learning and physical constraints to achieve accurate correction of the acoustic parameters of the composite plate. Compared with traditional methods, it has significantly improved adaptability and accuracy. The bidirectional long short-term memory network captures the temporal features of the waveform, the multi-scale attention mechanism enhances the representation ability of key information, and the physical constraints ensure that the parameter correction result conforms to the sound wave propagation law. The adaptive fusion strategy can dynamically adjust the weights of local and global information according to the signal characteristics, so that the algorithm can maintain stable performance in different noise environments and defect types.

[0042] Optionally, The steps of constructing the optimization objective function include: The multi-level physical constraints include: a wave equation constraint term constructed based on sound wave propagation theory, used to ensure that the set of acoustic parameters satisfies the wave equation; a parameter range constraint term constructed based on material acoustic characteristics, used to ensure that the set of acoustic parameters is within a physically feasible range; an interface continuity constraint term constructed based on interlayer interface characteristics, used to ensure that the sound wave satisfies the displacement and stress continuity conditions when propagating across the interlayer interface; The waveform reconstruction error includes: a mean square error term based on the amplitude of the time-domain waveform, used to represent the amplitude difference between the measured waveform and the reconstructed waveform; a phase error term based on Hilbert transform, used to represent the phase difference between the measured waveform and the reconstructed waveform. The optimization objective function adopts a staged weight configuration strategy based on the iterative process, and dynamically adjusts the weight coefficients of the waveform reconstruction error term and the physical constraint term according to the waveform reconstruction error value, the degree of violation of the physical constraint, and the acoustic impedance difference of the material interface.

[0043] For example, the wave equation constraint term is implemented by checking whether the acoustic parameters satisfy the propagation equation of sound waves in anisotropic medium. For each acoustic characteristic layer, the longitudinal wave velocity, transverse wave velocity and density parameters of the layer are substituted into the sound wave propagation equation to check the residual value. The residual value is calculated using the difference method, and for a 5MHz ultrasonic wave, the spatial step is set to 0.05mm and the time step is set to 0.01μs. When the residual value exceeds a preset threshold (such as 1% of the longitudinal wave velocity), the corresponding wave equation constraint term is increased. In practice, for a typical carbon fiber composite plate, when the longitudinal wave velocity varies in the range of 5900-6500m / s, the weight coefficient of the wave equation constraint term is set to 0.3; when the velocity value exceeds this range, the weight coefficient increases to 0.5 to force the parameter regression to a reasonable range.

[0044] The parameter range constraint term constructed based on material acoustic characteristics is the second layer of physical constraints. This constraint ensures that each acoustic parameter is within a physically feasible range. For carbon fiber composite materials, the reasonable range of longitudinal wave velocity is 5500-7000m / s, the reasonable range of transverse wave velocity is 3000-3800m / s, the reasonable range of density is 1450-1750kg / m³, and the reasonable range of attenuation coefficient is 15-50Np / m. The parameter range constraint term is implemented in the form of soft constraint, and when the parameter value approaches the boundary, the value of the constraint term increases gently; when the parameter value exceeds the boundary, the value of the constraint term increases sharply. In specific implementation, when the parameter deviates from the center value by 10% or less, the constraint term growth coefficient is 0.05; when it deviates by 10%-20%, the coefficient is 0.2; when it deviates by 20%-30%, the coefficient is 0.5; when it deviates by more than 30%, the coefficient is 1.0. This segmented setting allows the parameter to have a certain optimization space while ensuring the reasonableness of the parameter.

[0045] The interface continuity constraint term constructed based on the interlayer interface characteristics is the third layer of physical constraints. This constraint ensures that the displacement and stress continuity conditions are met when the sound wave propagates at the interlayer interface. At the interface, the normal displacement, tangential displacement, normal stress, and tangential stress of the upper and lower layers need to meet the continuity conditions. In implementation, for each interface, the displacement and stress difference values on both sides of the interface are calculated, and when the difference value is greater than the preset threshold, the corresponding interface continuity constraint term is increased. For the third interface, the displacement continuity threshold is set to 1 μm, and the stress continuity threshold is set to 0.1 MPa. In practical applications, when the material properties on both sides of the interface differ greatly (such as the fiber direction changes more than 45 degrees), the weight of the continuity constraint needs to be adjusted accordingly, and the typical value is 0.15-0.25.

[0046] The waveform reconstruction error includes two parts: a mean square error term based on the time domain waveform amplitude and a phase error term based on Hilbert transform. The mean square error term is used to represent the amplitude difference between the measured waveform and the reconstructed waveform, and the calculation method is to sum the square of the amplitude difference of the corresponding sampling points of the two waveforms, and then divide by the number of sampling points. For a 1024-point A-scan signal, the mean square error usually decreases rapidly in the first 20 iterations, from the initial value (such as 0.5) to a lower level (such as 0.05). The phase error term based on Hilbert transform is used to represent the phase difference between the measured waveform and the reconstructed waveform. First, Hilbert transform is performed on the measured waveform and the reconstructed waveform to obtain the corresponding analytic signals, then the absolute value of the phase difference between the two analytic signals is calculated, and finally the phase difference of all sampling points is summed and normalized. In spectral analysis, the phase difference in the range of 3-7 MHz is mainly concerned, and the phase error in this range has the greatest impact on the reconstruction quality. In practice, the weight coefficient of the phase error term is usually set to half of the mean square error term, such as 0.1, because the phase information has an important influence on the shape of the reconstructed waveform, but its contribution to the overall fitting degree is less than that of the amplitude information.

[0047] The optimization objective function adopts a phased weight configuration strategy based on the iteration process to dynamically adjust the weight coefficients of each term. In specific implementation, it is divided into three stages: In the first stage (the number of iterations is less than 50), the weight of the waveform reconstruction error term is higher, and the weight of the physical constraint term is lower. The typical configuration is: the mean square error term weight is 0.5, the phase error term weight is 0.2, the wave equation constraint term weight is 0.1, the parameter range constraint term weight is 0.1, and the interface continuity constraint term weight is 0.1. The goal of this stage is to quickly reduce the waveform reconstruction error and make the theoretical waveform close to the measured waveform.

[0048] When the mean square error is reduced to less than 20% of the initial value (for example, from 0.5 to less than 0.1), the second stage is entered. At this time, the weight of the waveform reconstruction error term is appropriately reduced, and the weight of the physical constraint term is increased. A typical configuration is: the mean square error term weight is 0.3, the phase error term weight is 0.15, the wave equation constraint term weight is 0.2, the parameter range constraint term weight is 0.15, and the interface continuity constraint term weight is 0.2. The second stage focuses on the physical reasonableness of the parameters and prevents overfitting.

[0049] When a region where the acoustic impedance difference at the material interface is greater than 30% is detected, the third stage is entered, and the processing strategy for the region is adjusted. At this time, the mean square error term weight of the region is increased by 50%, and the interface continuity constraint term weight is reduced by 30%. For example, if the acoustic impedance ratio of an interface is 1.35 (exceeding the threshold of 1.3), the mean square error term weight of the region near the interface is increased from 0.3 to 0.45, and the interface continuity constraint term weight is reduced from 0.2 to 0.14. This adjustment takes into account that at interfaces with large acoustic impedance differences, reflection and transmission phenomena are more complex, waveform fitting is more difficult, and interface continuity conditions may not be fully met.

[0050] During the entire optimization process, an adaptive adjustment mechanism is also set. When the violation of any physical constraint exceeds a predetermined threshold, the weight of the constraint term is doubled until the violation of the constraint is reduced to below the threshold. For example, when the longitudinal wave velocity of a layer exceeds the reasonable range (7000 m / s) by more than 10%, the weight of the parameter range constraint term is increased from 0.15 to 0.3. This mechanism ensures that physically unreasonable results do not occur during parameter optimization.

[0051] Figure 2 A comparison chart of waveform reconstruction error for different methods with iteration number is shown in the figure, where the horizontal axis represents the iteration number (0-250 times), and the vertical axis represents the waveform reconstruction error (0-0.5). The error reduction speed of genetic algorithm (dashed line) and simplex method (dotted line) gradually slows down, while the error of the present invention (solid line) is reduced to 0.23 in the first 50 iterations. After entering the second stage, due to the use of a multi-stage weight configuration strategy based on iteration progress, the weight coefficients of each error term and constraint term are dynamically adjusted, and the error reduction speed remains fast, ultimately reaching 0.01, fully demonstrating that the multi-level physical constraint optimization method of the present invention has higher parameter correction accuracy.

[0052] The present method combines multi-level physical constraints with waveform reconstruction error, effectively solving the problem of acoustic parameter correction of composite panels. Through the comprehensive consideration of time domain amplitude error and phase error, the accuracy of waveform reconstruction is improved; through the multi-stage weight configuration strategy, the balance between reconstruction accuracy and physical reasonableness is achieved.

[0053] Optionally, The corrected acoustic parameters are substituted into the acoustic wave propagation model to reconstruct the signal of the ultrasonic scan data, and the real defect signal is obtained. The step of determining the defect position coordinates according to the real defect signal comprises: The corrected acoustic parameters and the acoustic wave transmission coefficient matrix are substituted into the acoustic wave propagation model, an interlayer multiple reflection compensation factor is constructed based on the reflection coefficient of the adjacent interlayer interface and the interlayer acoustic wave propagation time delay, the interlayer multiple reflection compensation factor is multiplied by the frequency domain response of the ultrasonic scan data, and then inverse Fourier transform is performed to obtain a reconstructed signal. The time domain amplitude feature and the phase feature of the reconstructed signal are extracted, the time domain amplitude feature is differentially operated with the theoretical reflection time calculated based on the corrected acoustic parameters, the phase feature is differentially operated with the theoretical reflection phase, and the real defect signal is identified according to the differential operation result. The acoustic wave propagation time equation is established based on the corrected acoustic parameters, and the defect position coordinates are obtained by solving the acoustic wave propagation time equation according to the time characteristics of the real defect signal.

[0054] For example, the corrected acoustic parameters and the acoustic wave transmission coefficient matrix are substituted into the acoustic wave propagation model. The acoustic parameters include the longitudinal wave velocity, the transverse wave velocity, the density and the attenuation coefficient of each layer, and the acoustic wave transmission coefficient matrix contains the characteristics of the acoustic wave propagation at the interface. Taking an 8-layer carbon fiber composite plate as an example, the acoustic parameters of the third layer after the foregoing correction are: longitudinal wave velocity 6320 m / s, transverse wave velocity 3280 m / s, density 1620 kg / m 3 , and attenuation coefficient 28 Np / m. The main diagonal elements of the acoustic wave transmission coefficient matrix of the interface between the third layer and the fourth layer are 0.92 and 0.88, indicating a high transmission efficiency.

[0055] For each interface between two adjacent layers, the amplitude attenuation of multiple reflections is calculated according to the interface reflection coefficient, and the propagation time delay is calculated according to the thickness and acoustic velocity of each layer. For example, the thickness of the third layer is 1.25 mm, the longitudinal wave velocity is 6320 m / s, and the propagation time delay is 0.40 μs; the reflection coefficient of the interface between the third layer and the fourth layer is 0.08, and the corresponding one-time reflection attenuation is 92%. Based on these parameters, the interlayer multiple reflection compensation factor is constructed by superimposing the responses of multiple reflections. Generally, multiple reflections within 3 times are considered, because the reflection intensity of higher times has been attenuated to less than 1% of the original signal and can be ignored.

[0056] The original ultrasonic A-scan signal is subjected to Fourier transform to obtain a frequency domain response; then the frequency domain response is multiplied by a multiple reflection compensation factor to realize frequency domain compensation; and finally inverse Fourier transform is performed to obtain a time domain reconstructed signal. The frequency domain compensation operation is crucial for effectively separating the interlayer multiple reflection and the real defect signal. In practical applications, for an ultrasonic wave with a center frequency of 5 MHz, compensation processing is mainly performed in the range of 2-8 MHz, and frequency components outside this range have less effect on the compensation effect.

[0057] Compared with the original signal, the interlayer multiple reflection is effectively suppressed in the reconstructed signal. For example, in a typical 8-layer composite plate test, the amplitude of the multiple reflection signal at the interface between the third layer and the fourth layer in the original signal is about 15% of the main reflection signal, while the amplitude of this multiple reflection signal after reconstruction is reduced to less than 3%, and the signal-to-noise ratio is increased by about 5 times. The peak value in the reconstructed signal is more prominent, and the distinction between the defect signal and the structure signal is significantly improved.

[0058] The time domain amplitude feature and the phase feature of the reconstructed signal are extracted. The time domain amplitude feature includes the local maximum value of the signal and its occurrence time, and a sliding window method is used, with the window size set to 1.5 times the ultrasonic wave period (a typical value is 0.3 μs), and the window sliding step is the sampling interval (a typical value is 0.01 μs). The amplitude maximum value is detected in each window, and the time position and amplitude size are recorded. For a typical case of an 8-layer composite plate, 7 main peak values corresponding to the interlayer interfaces and a number of secondary peak values can be detected. The phase feature is calculated by Hilbert transform to calculate the instantaneous phase, and the time position and jump amplitude of the phase jump point are recorded. Phase jump is an important feature for identifying defect signals, because defects often cause abnormal changes in the phase of sound waves.

[0059] The time domain amplitude feature is subjected to difference operation with the theoretical reflection time calculated based on the modified acoustic parameters. The theoretical reflection time is calculated based on the thickness and sound speed of each layer, for example, for the above-mentioned 8-layer composite plate, the theoretical reflection times of the interfaces of the first three layers are 24.2 μs, 36.1 μs and 48.3 μs respectively. The actual detected peak time is differentiated with the theoretical time to obtain the time deviation. The time deviation of the normal interlayer reflection is usually within ±0.2 μs, while the defect signal often shows obvious time advance or delay, with a deviation range of ±0.5 μs to ±2.0 μs.

[0060] The phase feature is differenced with the theoretical reflection phase. The theoretical reflection phase is calculated according to the acoustic impedance ratio of the materials on both sides of the interface. When the acoustic impedance ratio is greater than 1, the phase change is π, and when the acoustic impedance ratio is less than 1, the phase change is 0. For example, the acoustic impedance ratio of the third interface is 1.06, and the theoretical reflection phase changes by π. The actual detected phase jump is differenced with the theoretical phase change to obtain the phase deviation. The phase deviation of the normal interlayer reflection is usually within ±π / 6, while the phase deviation of the defect signal is often above ±π / 3, and some are even close to π.

[0061] Real defect signals are identified according to the difference operation result. The identification of the defect signal adopts double criteria: one is that the time deviation exceeds ±0.3μs, and the other is that the phase deviation exceeds ±π / 4. When both conditions are met, it is determined as a defect signal. Different threshold combinations are set for different types of defects such as delamination and inclusion. For example, the time deviation threshold of delamination defects is set to ±0.3μs, and the phase deviation threshold is set to ±π / 4; while the time deviation threshold of inclusion defects is set to ±0.5μs, and the phase deviation threshold is set to ±π / 3. This differentiated setting improves the recognition accuracy of different types of defects.

[0062] An acoustic wave propagation time equation is established based on the corrected acoustic parameters. The equation describes the relationship between the acoustic wave propagation time from the transducer emission point to the defect and back to the receiving point and the defect position coordinates. The equation contains the thickness and sound speed information of each layer, as well as the position coordinates of the transducer. For the common case of vertical incidence, the equation is simplified to the sum of the propagation times of the acoustic wave in each layer. For example, if the defect is located in the fourth layer, the acoustic wave needs to pass through the first three layers in turn, reflect and then return, and the total propagation time is the sum of the time of each segment.

[0063] The defect position coordinates are obtained by solving the acoustic wave propagation time equation according to the time characteristics of the real defect signal. The Newton iteration method is used in the solving process, and the initial value is set as the estimated depth corresponding to the time. The iteration convergence condition is that the coordinate change of adjacent two iterations is less than 0.01mm or the iteration number reaches 20 times. For the defect signal with a time of 58.6μs, the calculated defect depth is 3.85mm, located in the middle of the fourth layer, and the lateral coordinate is consistent with the probe position. When using oblique incidence or focused probe, the equation solving needs to consider the change of the acoustic wave propagation path, and the calculation process is more complex, but the principle is the same.

[0064] To verify the accuracy of the positioning results, cross-validation can be performed using multiple scans at different locations. For example, a defect in a carbon fiber composite panel was scanned at five different locations, and the calculated defect depths were 3.83mm, 3.86mm, 3.85mm, 3.84mm, and 3.87mm, respectively. The average was 3.85mm, and the standard deviation was 0.014mm, indicating good consistency and reliability of the positioning results. Compared with the subsequent slice verification results, the positioning error was less than 0.1mm, meeting the requirements of engineering applications.

[0065] This method has three significant advantages: by constructing an interlayer multiple reflection compensation factor, it effectively eliminates the multiple reflection interference in complex multi-layer structures and improves the signal-to-noise ratio of the defect signal; combining the dual criteria of time domain amplitude characteristics and phase characteristics, it significantly improves the accuracy of defect identification and reduces the misjudgment rate; the sound wave propagation time equation established based on precisely corrected acoustic parameters makes the calculation of defect position coordinates more accurate.

[0066] Optional, The steps of meshing the coordinates of the defect position and calculating stress field distribution data around the defect; obtaining acoustic characteristic change data of the defect area under the action of the stress field distribution data and determining the stress concentration position include: An initial grid is constructed with the defect location coordinates as the center, and the square root of the sum of the squares of the stress partial derivatives in the X and Y directions of each grid cell is calculated to obtain a stress gradient index. The grid cells whose stress gradient index is greater than a preset stress gradient threshold are refined and segmented; Calculating the spatial derivative of the displacement component based on the refined and segmented grid to obtain a strain value, and multiplying the strain value by the material elastic constant to obtain stress field distribution data; Calculating a sound velocity variation according to the stress field distribution data, wherein the sound velocity variation is determined by multiplying the principal stress components in three directions by the corresponding acoustoelastic coefficients, and calculating an acoustic impedance correction value according to the sound velocity variation; Extracting acoustic features corresponding to the sound velocity change and the acoustic impedance correction value, and constructing a feature vector including the sound velocity change rate, the acoustic impedance change rate, and the spectral energy distribution; calculating the weighted Euclidean distance between the feature vectors, determining a neighborhood radius based on the average distance of the K nearest neighboring points around each feature point, performing density clustering on the feature points using the neighborhood radius, and mapping the clustering results back to spatial coordinates to determine the stress concentration location; The ratio of the maximum stress to the nominal stress in the stress concentration location is calculated to obtain a stress concentration coefficient. The deviation between the measured acoustic characteristics and the theoretical acoustic characteristics is substituted into an exponential decay function to calculate a confidence index. The reliability of the stress concentration location is determined based on the confidence index.

[0067] An exemplary initial mesh is constructed with the defect location coordinates as the center, using a rectangular mesh structure, with the defect as the center, and expanding outward to a region 5 times the size of the defect. For a circular defect with a diameter of 2 mm, the initial mesh covers a range of 10 mm x 10 mm, and the initial mesh size is 0.5 mm x 0.5 mm, including a total of 400 mesh units. To capture regions with significant stress gradient changes, the square root of the sum of the squares of the stress partial derivatives in the X and Y directions for each mesh unit is calculated as a stress gradient index. The stress partial derivative is calculated by the central difference method, and the difference between the stress values of adjacent mesh points is divided by the mesh spacing. For example, the stress gradient index of a mesh unit located at the edge of the defect is usually in the range of 25-35 MPa / mm, while in the area far from the defect, the index is usually less than 5 MPa / mm.

[0068] The mesh units with a stress gradient index greater than a preset stress gradient threshold value are refined and divided. In practical applications, the stress gradient threshold value is set to 20 MPa / mm, and the mesh units higher than this threshold value will be equally divided into four sub-meshes. The refinement process can be iterated, and for areas with particularly high stress gradients (such as the tip of the defect), up to 3 times of refinement can be performed, so that the minimum mesh size reaches 0.0625 mm x 0.0625 mm. In a certain circular defect case, the initial 400 mesh units are increased to 1260 after refinement, of which the mesh density in the 0.5 mm range around the defect is the highest, covering about 680 mesh units.

[0069] The strain value is obtained by calculating the spatial derivative of the displacement component based on the refined and divided mesh. The displacement field is calculated using the finite element method, and the boundary condition is set as the working stress of the composite plate, and the typical value is a tensile stress of 50 MPa. The displacement field is solved using a quadratic interpolation function to improve accuracy. For anisotropic carbon fiber composites, the angle between the material principal axis direction and the loading direction needs to be considered in the calculation of the displacement field. For example, when the fiber direction and the loading direction are at an angle of 45 degrees, the displacement field exhibits obvious asymmetry, and the maximum displacement around the defect can reach 1.8 times the far-field displacement.

[0070] The stress field distribution data is obtained by multiplying the strain value by the material elastic constant. For orthotropic carbon fiber composites, the elastic constant matrix contains 9 independent components. For a typical carbon fiber / epoxy composite material, the longitudinal elastic modulus is 140 GPa, the transverse elastic modulus is 10 GPa, the in-plane shear modulus is 5.5 GPa, and the principal Poisson's ratio is 0.3. The stress calculation considers the anisotropic properties of the material, and the stress tensor at each point is obtained. At the edge of the circular defect, the maximum principal stress can reach 3 times the far-field stress, forming a significant stress concentration. For the tip of the defect, the stress concentration is more serious, and the maximum principal stress can reach 5-7 times the far-field stress.

[0071] The velocity variation is calculated from the stress field distribution data. The velocity variation is determined by the product of the principal stress components and the corresponding acoustic coefficients. For carbon fiber composites, the typical value of the longitudinal wave acoustic coefficient is -10 x 10 -11 Pa -1 , and the typical value of the transverse wave acoustic coefficient is -15 x 10 -11 Pa -1 . This means that when the material is subjected to a tensile stress of 50 MPa, the longitudinal wave velocity changes by about -0.5%, and the transverse wave velocity changes by about -0.75%. In the stress concentration area, this change is more significant, for example, at the defect tip, the longitudinal wave velocity decreases by up to -3.5%, and the transverse wave velocity decreases by up to -5.2%. The acoustic impedance correction value is determined by the velocity variation and the density variation. Due to the small change in the volume of the material under stress, the density also changes accordingly, but this change is usually small, and for a stress level of 50 MPa, the density change is about -0.15%. Therefore, the acoustic impedance change is mainly determined by the velocity change. In the high stress area of the defect tip, the acoustic impedance decreases by up to -3.7%, which is sufficient to produce a measurable difference in the reflection signal in ultrasonic testing.

[0072] The acoustic features corresponding to the velocity variation and the acoustic impedance correction value are extracted, and a feature vector containing the velocity variation rate, the acoustic impedance variation rate, and the spectral energy distribution is constructed. The velocity variation rate is directly calculated from the above calculation; the acoustic impedance variation rate is determined by the velocity variation rate and the density variation rate; the spectral energy distribution is obtained by comparing the ultrasonic spectra before and after stress. For 5 MHz center frequency ultrasonic waves, the spectral changes caused by stress mainly manifest as a small shift in the center frequency (usually within ±0.2 MHz) and a change in the frequency band width (usually within ±5%). The feature vector of each grid point contains 5 components: the longitudinal wave velocity variation rate, the transverse wave velocity variation rate, the acoustic impedance variation rate, the spectral center frequency shift, and the frequency band width change rate.

[0073] The weighted Euclidean distance between the feature vectors is calculated, and the neighborhood radius is determined according to the average distance of the K nearest neighbors around each feature point. In the weighted Euclidean distance calculation, the weights of each feature component are set according to their sensitivity, and the typical values are as follows: the longitudinal wave velocity variation rate weight is 0.3, the transverse wave velocity variation rate weight is 0.3, the acoustic impedance variation rate weight is 0.2, the spectral center frequency shift weight is 0.1, and the frequency band width change rate weight is 0.1. The number of nearest neighbors K is set to 5% of the total number of points, for example, for 1260 grid points, K is 63. The neighborhood radius is the average distance of the K nearest neighbors, and for different areas, the neighborhood radius is different, and the neighborhood radius in the stress concentration area is usually smaller, about 0.7 times the average distance in the feature space.

[0074] The stress concentration locations are determined by mapping the density clustering results back to spatial coordinates. The density clustering algorithm is based on the density reachability of points, which divides high density areas into one cluster. The clustering parameters include the minimum number of points, which is set to K / 3, i.e., 21 points, and the distance threshold, which is set to the neighborhood radius calculated above. The clustering results usually get 3-5 clusters, in which the densest cluster corresponds to the most significant stress concentration area. Mapping the clustering results back to spatial coordinates can get the accurate location and shape of the stress concentration area. In a certain circular defect case, the most significant stress concentration area is located at the ±45 degree position of the defect edge, forming two local areas, each with a size of about 0.8mm x 0.5mm.

[0075] The stress concentration factor is calculated by the ratio of the maximum stress in the stress concentration location to the nominal stress. For the above circular defect case, the maximum stress in the stress concentration location is 152MPa, and the far-field nominal stress is 50MPa, so the stress concentration factor is 3.04. Different shaped defects have different stress concentration factors, for example, when the long axis of an elliptical defect is perpendicular to the loading direction, the stress concentration factor can be as high as 4.2; when the long axis is parallel to the loading direction, the stress concentration factor drops to about 2.1.

[0076] The deviation of the measured acoustic features from the theoretical acoustic features is substituted into the exponential decay function to calculate the confidence index. The measured acoustic features come from the reflection waveform analysis of the ultrasonic scanning data, including reflection amplitude, phase and spectral changes; the theoretical acoustic features come from the above stress field calculation results. The deviation is expressed by the normalized root mean square error, ranging from 0.05 to 0.2. The confidence index calculation uses an exponential decay function form, when the deviation is 0.05, the confidence index is 0.95; when the deviation is 0.1, the confidence index is 0.90; when the deviation is 0.2, the confidence index drops to 0.82. This calculation method ensures that the confidence index is inversely proportional to the consistency of the measured data and the theoretical prediction. The area with a confidence index greater than 0.9 is considered to be a high-reliability stress concentration location; the area with a confidence index between 0.8 and 0.9 is considered to be a medium-reliability stress concentration location; the area with a confidence index lower than 0.8 needs additional verification means. For example, for key structural components, it is usually required that the confidence index of the stress concentration location is not less than 0.85 to ensure the reliability of the evaluation results.

[0077] The adaptive mesh refinement strategy of the method improves the calculation efficiency and accuracy, finer mesh is used in the area with large stress gradient, and coarser mesh is maintained in the area with gentle stress change, so that the calculation resources are reasonably allocated; the quantitative relationship between the stress field and the acoustic characteristics is established through the acoustic effect, and the organic combination of the ultrasonic detection result and the stress analysis is realized; the density clustering algorithm based on multi-dimensional features improves the robustness of the stress concentration position identification and reduces the influence of noise and measurement error.

[0078] Optionally, The steps of adjusting the transmission power and scanning interval of the ultrasonic probe according to the stress concentration position, performing fine scanning and obtaining high-resolution data, and updating the defect spatial distribution information according to the high-resolution data include: According to the depth value and the maximum stress value of the stress concentration position, a sound wave energy attenuation compensation coefficient is calculated, the sound wave energy attenuation compensation coefficient is multiplied by the initial transmission power to obtain an optimized transmission power; the scanning interval is determined based on the ratio of the ultrasonic wavelength to the probe diameter; A spiral scanning track is constructed with the stress concentration position as the center, the spiral scanning track is determined by an angle parameter and a radius parameter, and the angle parameter and the radius parameter are adjusted according to the scanning interval so that the spiral scanning track completely covers the stress concentration position; high-resolution scanning data is obtained by performing ultrasonic scanning along the spiral scanning track using the optimized transmission power; The boundary curvature value and the signal-to-noise ratio of the high-resolution scanning data are calculated, and a weighted fusion coefficient is constructed based on the boundary curvature value and the signal-to-noise ratio; the weighted fusion coefficient is used to fuse the defect position coordinate information and the fine scanning features, and the spatial distribution information of the defect is updated, the spatial distribution information includes the defect boundary contour and the internal structure features.

[0079] For example, when the sound wave propagates in the composite material, the energy attenuation occurs, which is related to the material properties, the sound wave frequency and the propagation distance. The calculation of the sound wave energy attenuation compensation coefficient considers two main factors: one is the geometric diffusion and material absorption attenuation caused by depth, and the other is the acoustic impedance change caused by stress concentration. For the depth factor, the attenuation compensation coefficient increases with the increase of depth, and an exponential relationship is used to describe it; for the stress concentration factor, the attenuation compensation coefficient is proportional to the maximum stress value. For example, for a stress concentration area located at a depth of 3.5 mm and with a maximum stress of 150 MPa, the calculated attenuation compensation coefficient is 1.42, which means that the transmission power needs to be increased by 42% to obtain sufficient echo signal strength.

[0080] The initial transmit power of the ultrasonic probe is usually set to an excitation voltage of 50-100 V, corresponding to an acoustic power of 0.1-0.2 W. For example, when the initial transmit power is 80 V, the optimized transmit power is 113.6 V after multiplying the attenuation compensation factor of 1.42. The upper limit is set to twice the initial power. When the calculation result exceeds this upper limit, the upper limit is used, and the signal averaging number is increased accordingly to improve the signal-to-noise ratio.

[0081] The scan pitch is determined based on the ratio of the ultrasonic wavelength to the probe diameter. For carbon fiber composite materials, the longitudinal wave speed is about 6000 m / s, and the wavelength corresponding to a frequency of 5 MHz is 1.2 mm. The probe diameter is usually 6-12 mm, for example, when a probe with a diameter of 8 mm is used, the ratio of the wavelength to the probe diameter is 0.15. According to the acoustic field theory and the sampling theorem, the scan pitch should be no more than half of the wavelength, i.e. 0.6 mm. However, considering that the stress concentration area requires higher resolution, the actual scan pitch is further reduced to one quarter of the wavelength, i.e. 0.3 mm. Such a setting ensures a good balance between high frequency details and spatial resolution.

[0082] A spiral scan trajectory is constructed centered on the stress concentration position, which has higher efficiency and more uniform spatial coverage than traditional grid scanning. The spiral trajectory is determined by an angle parameter and a radius parameter, the angle parameter represents the rotation angle around the center point, and the radius parameter represents the distance to the center point. The angle parameter starts from 0 degrees and increases by a fixed step until the required number of turns is completed; the radius parameter increases linearly with the angle. For a specific spiral trajectory, the distance between adjacent trajectory points can be controlled to be equal to or slightly less than the set scan pitch of 0.3 mm by adjusting the angle step and the radius growth rate. Adjust the angle parameter and the radius parameter according to the scan pitch to make the spiral scan trajectory completely cover the stress concentration position. For example, for an elliptical stress concentration area with an area of 0.8 mm x 0.5 mm, considering the edge effect, the scanning range needs to be expanded to 1.2 mm x 0.9 mm. Take the center of the ellipse as the starting point of the spiral, set the angle step to 15 degrees, and the radius growth rate to 0.02 mm / 15 degrees. The distance between the trajectory points of one spiral (360 degrees) is about 0.29 mm, which is less than the set scan pitch of 0.3 mm, ensuring sufficient spatial coverage. A total of 6 spirals are needed to completely cover the target area, corresponding to a maximum radius of 0.72 mm, generating a total of 144 scan points.

[0083] The ultrasonic wave is scanned along the spiral scanning trajectory to obtain high-resolution scanning data by using the optimized transmit power. During the scanning process, the probe stops at each trajectory point and transmits the ultrasonic wave, and receives the echo signal. To improve the signal-to-noise ratio, multiple repeated measurements are performed at each point and the average value is taken, usually set to 16-32 times. For the case of the above-mentioned optimized transmit power of 113.6V, 24 times of signal averaging can improve the signal-to-noise ratio by about 4.9 times. The original data obtained by scanning is the A-scan signal corresponding to each trajectory point, usually containing 1024-2048 sampling points, with a sampling rate of 100MHz, and the corresponding time window is 10.24-20.48μs.

[0084] For high-resolution scanning data, the boundary curvature value reflects the shape characteristics of the defect boundary, and is an important indicator for distinguishing different types of defects. The curvature calculation is based on the difference in echo time of adjacent scanning points. The specific process is: first, extract the time position of the defect echo in each A-scan signal to construct a defect depth distribution map; then, perform smoothing and edge detection on the depth distribution map to obtain the defect boundary; finally, calculate the curvature value of each point on the boundary. For circular defects, the boundary curvature value is relatively uniform, and the standard deviation is less than 0.2; while for crack-like defects, the curvature value increases significantly at the tip, and the maximum curvature value can be 3-5 times the average value. The signal-to-noise ratio is calculated as the ratio of the amplitude of the defect echo signal to the standard deviation of the background noise, and the signal-to-noise ratio of high-quality scanning data is usually greater than 20dB. The weighted fusion coefficient is used to control the weight of the initial defect position information and the fine scanning feature in the fusion process. The construction method is: the area with high boundary curvature value (such as the defect tip) gives higher weight to the fine scanning feature, while the area with low curvature value (such as the flat edge of the defect) gives higher weight to the initial defect position information; at the same time, the weight of the fine scanning feature increases in the area with high signal-to-noise ratio, and the weight of the initial information increases in the area with low signal-to-noise ratio. In practical application, when the curvature value exceeds 1.5 and the signal-to-noise ratio exceeds 25dB, the weight of the fine scanning feature can reach 0.85; when the curvature value is less than 0.5 and the signal-to-noise ratio is less than 15dB, the weight of the initial defect position information increases to 0.7.

[0085] The weighted fusion coefficient is used to fuse the defect position coordinate information and the fine scanning feature, and update the spatial distribution information of the defect. The fusion process uses weighted averaging, and for each point on the defect boundary, the final position coordinates are the weighted average of the initial position and the fine scanning position. For example, the initial depth of a certain point is 3.85mm, the fine scanning depth is 3.92mm, and the weighted fusion coefficient is 0.75 (fine scanning weight), then the fused depth is 3.9025mm. For the internal structural features of the defect, they are mainly extracted from the fine scanning data, including internal echo intensity distribution, phase variation and spectral features. Through these features, the nature of the defect can be judged, such as whether it contains inclusions, whether there are secondary cracks, etc.

[0086] The updated spatial distribution information of the defect includes the defect boundary outline and internal structural features. The boundary outline is represented as a 3D point cloud, describing the shape and size of the defect. For example, for an actual elliptical defect, the updated major axis length is 2.24mm, the minor axis length is 1.38mm, and the area is 2.43mm², which improves the dimensional accuracy by approximately 40% compared to the initial estimate. The internal structural features include the acoustic impedance distribution map, phase jump map, and spectral energy distribution map, which together reveal the internal composition and properties of the defect. For example, low values ​​in the central area of ​​the acoustic impedance distribution map (more than 30% lower than the surrounding material) indicate that the defect may have a cavity inside; while high values ​​in certain areas (more than 15% higher than the surrounding material) indicate the presence of high-density impurities.

[0087] For specific defect types, characteristic parameters can also be extracted for quantitative characterization. For example, for delamination defects, their area, perimeter, thickness and average acoustic impedance value are extracted; for crack-type defects, their length, opening width, tip radius and expansion direction are extracted. These parameters, combined with the mechanical properties of the material, can be used to evaluate the impact of defects on structural integrity. In a case study of a carbon fiber / epoxy resin composite plate, a crack defect with a length of 2.24 mm and a tip radius of 0.08 mm was found. The angle between its expansion direction and the principal stress direction was 15 degrees. The evaluation results showed that the defect was at risk of continuing to expand under the working load and needed to be repaired or monitored more closely.

[0088] This method is based on the acoustic wave energy attenuation compensation technology at the stress concentration location, which improves the detection sensitivity of defects in deep and high stress areas; the spiral scanning trajectory design not only improves the scanning efficiency, but also achieves uniform coverage of the defect area; the weighted fusion strategy based on boundary curvature and signal-to-noise ratio fully utilizes the complementary advantages of initial scanning and fine scanning, and the accuracy of boundary shape characterization is significantly enhanced.

[0089] In a second aspect, a composite plate detection system is provided, comprising: The first unit is configured to scan the composite plate to obtain ultrasonic scanning data, divide the composite plate into N acoustic characteristic layers, construct an acoustic wave propagation model based on a recursive transfer matrix method, calculate an acoustic wave transfer coefficient matrix of the N acoustic characteristic layers, and calculate a reflection coefficient and a transmission coefficient of an interface of each layer based on the acoustic wave transfer coefficient matrix; The second unit is used to compare the measured ultrasonic scanning data with the theoretical waveform calculated based on the reflection coefficient and the transmission coefficient, and dynamically correct the acoustic parameters of each acoustic characteristic layer based on the deep learning algorithm and physical constraints; a third unit configured to substitute the corrected acoustic parameters into the acoustic wave propagation model, reconstruct a signal of the real defect from the ultrasonic scan data, and determine a defect position coordinate according to the signal of the real defect; a fourth unit configured to divide a grid at the defect position coordinate, calculate stress field distribution data around the defect, obtain acoustic characteristic variation data of the defect region under the action of the stress field distribution data, and determine a stress concentration position; a fifth unit configured to adjust a transmission power and a scan interval of an ultrasonic probe for the stress concentration position, perform fine scanning and obtain high-resolution data, and update defect spatial distribution information according to the high-resolution data; a sixth unit configured to output detection data, the detection data including defect position coordinate information, stress concentration distribution data, and defect hazard level.

[0090] In a third aspect, a computer readable storage medium is provided, and computer program instructions are stored on the computer readable storage medium. When the computer program instructions are executed by a processor, the foregoing method is implemented.

Claims

1. A method for detecting a composite board, characterized in that: include: Scanning a composite plate to obtain ultrasonic scanning data, dividing the composite plate into N acoustic characteristic layers, constructing an acoustic wave propagation model based on a recursive transfer matrix method, calculating an acoustic wave transfer coefficient matrix for the N acoustic characteristic layers, and calculating a reflection coefficient and a transmission coefficient at an interface of each layer based on the acoustic wave transfer coefficient matrix; Comparing the measured ultrasonic scanning data with the theoretical waveform calculated based on the reflection coefficient and transmission coefficient, and dynamically correcting the acoustic parameters of each acoustic characteristic layer based on a deep learning algorithm and physical constraints; Substituting the corrected acoustic parameters into the sound wave propagation model, performing signal reconstruction on the ultrasonic scanning data to obtain a true defect signal, and determining the defect position coordinates based on the true defect signal; Performing grid division on the coordinates of the defect position and calculating stress field distribution data around the defect; obtaining acoustic characteristic change data of the defect area under the action of the stress field distribution data and determining the stress concentration position; Adjusting the transmitting power and scanning interval of the ultrasonic probe according to the stress concentration position, performing fine scanning and acquiring high-resolution data, and updating the defect spatial distribution information according to the high-resolution data; Output detection data, which includes defect location coordinate information, stress concentration distribution data and defect hazard level.

2. The method according to claim 1, characterized in that The steps of dividing the composite plate into N acoustic characteristic layers, constructing an acoustic wave propagation model based on a recursive transfer matrix method, calculating an acoustic wave transfer coefficient matrix of the N acoustic characteristic layers, and calculating the reflection coefficient and transmission coefficient of each layer interface according to the acoustic wave transfer coefficient matrix include: Performing empirical mode decomposition on the ultrasonic scanning data to obtain an intrinsic mode function, calculating a phase function of the intrinsic mode function to obtain an instantaneous frequency characteristic, combining the instantaneous frequency characteristic with the square of the amplitude of the intrinsic mode function to construct a time-frequency energy distribution matrix, determining the position of the interlayer interface of the composite plate based on the mutation characteristics of the time-frequency energy distribution matrix, and dividing the composite plate into N acoustic characteristic layers; Acquiring three-dimensional morphological parameters at the interlayer interface position, constructing a stiffness calculation formula that takes into account interface stress based on the three-dimensional morphological parameters, and obtaining an interface stiffness matrix that characterizes interface contact characteristics; constructing an ideal interface transfer matrix based on the wave impedance and thickness of each acoustic characteristic layer, and multiplying the interface stiffness matrix by the ideal interface transfer matrix to obtain an acoustic wave transfer coefficient matrix; The acoustic wave transfer coefficient matrix is ​​substituted into an acoustic wave propagation model, where the acoustic wave propagation model is constructed based on the material parameters of each acoustic characteristic layer; the wave numbers of the N acoustic characteristic layers are calculated according to the acoustic wave propagation model, and the reflection coefficient and transmission coefficient of the interface between the layers are calculated based on the wave numbers.

3. The method according to claim 1, characterized in that Comparing the measured ultrasonic scanning data with the theoretical waveform calculated based on the reflection coefficient and the transmission coefficient, and dynamically correcting the acoustic parameters of each acoustic characteristic layer based on a deep learning algorithm and physical constraints includes: The measured ultrasonic scanning data is input into the waveform encoding layer of the bidirectional long short-term memory neural network and encoded into a waveform feature vector; Inputting the waveform feature vector into a multi-scale attention mechanism layer, performing a wavelet transform on the waveform feature vector to obtain a time-frequency feature matrix, constructing a local attention vector and a global attention vector based on the time-frequency feature matrix, adaptively fusing the local attention vector with the global attention vector to obtain a fused attention vector, and performing an element-by-element product operation on the fused attention vector and the waveform feature vector to obtain a weighted feature vector; Inputting the weighted feature vector into an acoustic parameter prediction network, the acoustic parameter prediction network outputting an acoustic parameter set, the acoustic parameter set including a wave velocity parameter, a density parameter, and an attenuation coefficient parameter; Constructing an optimization objective function based on multi-level physical constraint functions and waveform reconstruction errors; Using ultrasonic data from a defect-free area to perform self-supervised pre-training on the bidirectional long short-term memory neural network, the self-supervised pre-training uses a waveform reconstruction task as a pre-training objective; The acoustic parameter set is iteratively optimized based on the optimization objective function, and the iteration is stopped when the parameter change amount of multiple consecutive iterations is less than a preset change threshold, thereby obtaining a revised acoustic parameter set.

4. The method according to claim 3, characterized in that The steps to construct the optimization objective function include: The multi-level physical constraints include: wave equation constraints based on acoustic wave propagation theory; parameter range constraints based on material acoustic properties; and interface continuity constraints based on interlayer interface properties. The waveform reconstruction error includes: a mean square error term based on the time domain waveform amplitude; a phase error term based on the Hilbert transform; The optimization objective function adopts a phased weight configuration strategy based on an iterative process, and dynamically adjusts the weight coefficients of the waveform reconstruction error term and the physical constraint term according to the waveform reconstruction error value, the degree of physical constraint violation, and the acoustic impedance difference of the material interface.

5. The method according to claim 2, characterized in that Substituting the corrected acoustic parameters into the sound wave propagation model, reconstructing the ultrasonic scanning data to obtain a true defect signal, and determining the defect position coordinates according to the true defect signal include: Substituting the corrected acoustic parameters and the acoustic wave transfer coefficient matrix into the acoustic wave propagation model, constructing an interlayer multiple reflection compensation factor based on the reflection coefficient of the interface between adjacent layers and the interlayer acoustic wave propagation delay, multiplying the interlayer multiple reflection compensation factor by the frequency domain response of the ultrasonic scanning data, and then performing an inverse Fourier transform to obtain a reconstructed signal; Extracting the time-domain amplitude feature and phase feature of the reconstructed signal, performing a differential operation on the time-domain amplitude feature and the theoretical reflection time calculated based on the corrected acoustic parameters, performing a differential operation on the phase feature and the theoretical reflection phase, and identifying the real defect signal based on the differential operation results; An acoustic wave propagation time equation is established based on the corrected acoustic parameters, and the acoustic wave propagation time equation is solved according to the time characteristics of the real defect signal to obtain the defect position coordinates.

6. The method according to claim 1, characterized in that Performing grid division on the coordinates of the defect position and calculating stress field distribution data around the defect; The steps of obtaining acoustic characteristic change data of the defect area under the action of the stress field distribution data and determining the stress concentration location include: An initial grid is constructed with the defect location coordinates as the center, and the square root of the sum of the squares of the stress partial derivatives in the X and Y directions of each grid cell is calculated to obtain a stress gradient index. The grid cells whose stress gradient index is greater than a preset stress gradient threshold are refined and segmented; Calculating the spatial derivative of the displacement component based on the refined and segmented grid to obtain a strain value, and multiplying the strain value by the material elastic constant to obtain stress field distribution data; Calculating a sound velocity variation according to the stress field distribution data, wherein the sound velocity variation is determined by multiplying the principal stress components in three directions by the corresponding acoustoelastic coefficients, and calculating an acoustic impedance correction value according to the sound velocity variation; Extracting acoustic features corresponding to the sound velocity change and the acoustic impedance correction value, and constructing a feature vector including the sound velocity change rate, the acoustic impedance change rate, and the spectral energy distribution; calculating the weighted Euclidean distance between the feature vectors, determining a neighborhood radius based on the average distance of the K nearest neighboring points around each feature point, performing density clustering on the feature points using the neighborhood radius, and mapping the clustering results back to spatial coordinates to determine the stress concentration location; The ratio of the maximum stress to the nominal stress in the stress concentration location is calculated to obtain a stress concentration coefficient. The deviation between the measured acoustic characteristics and the theoretical acoustic characteristics is substituted into an exponential decay function to calculate a confidence index. The reliability of the stress concentration location is determined based on the confidence index.

7. The method according to claim 1, characterized in that The steps of adjusting the transmitting power and scanning spacing of the ultrasonic probe according to the stress concentration position, performing fine scanning and acquiring high-resolution data, and updating the defect spatial distribution information according to the high-resolution data include: Calculating an acoustic wave energy attenuation compensation coefficient based on the depth value and the maximum stress value of the stress concentration position, and multiplying the acoustic wave energy attenuation compensation coefficient by the initial transmission power to obtain an optimized transmission power; determining a scanning interval based on a ratio of the ultrasonic wavelength to the probe diameter; Constructing a spiral scanning trajectory with the stress concentration location as the center, the spiral scanning trajectory being determined by an angle parameter and a radius parameter, and adjusting the angle parameter and the radius parameter according to the scanning interval so that the spiral scanning trajectory completely covers the stress concentration location; performing ultrasonic scanning along the spiral scanning trajectory using the optimized transmission power to obtain high-resolution scanning data; Calculate the boundary curvature value and signal-to-noise ratio of the high-resolution scanning data, and construct a weighted fusion coefficient based on the boundary curvature value and the signal-to-noise ratio; use the weighted fusion coefficient to fuse the defect position coordinate information and the fine scanning features, and update the spatial distribution information of the defect, wherein the spatial distribution information includes the defect boundary contour and internal structure features.

8. A composite board detection system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is configured to scan the composite plate to obtain ultrasonic scanning data, divide the composite plate into N acoustic characteristic layers, construct an acoustic wave propagation model based on a recursive transfer matrix method, calculate an acoustic wave transfer coefficient matrix of the N acoustic characteristic layers, and calculate a reflection coefficient and a transmission coefficient of an interface of each layer based on the acoustic wave transfer coefficient matrix; The second unit is used to compare the measured ultrasonic scanning data with the theoretical waveform calculated based on the reflection coefficient and the transmission coefficient, and dynamically correct the acoustic parameters of each acoustic characteristic layer based on the deep learning algorithm and physical constraints; The third unit is configured to substitute the corrected acoustic parameters into the sound wave propagation model, perform signal reconstruction on the ultrasonic scanning data, obtain a true defect signal, and determine the defect position coordinates according to the true defect signal; The fourth unit is used to perform grid division at the coordinates of the defect position, calculate stress field distribution data around the defect, obtain acoustic characteristic change data of the defect area under the action of the stress field distribution data, and determine the stress concentration position; A fifth unit is configured to adjust the transmission power and scanning interval of the ultrasonic probe according to the stress concentration position, perform fine scanning and obtain high-resolution data, and update the defect spatial distribution information according to the high-resolution data; The sixth unit is used to output detection data, where the detection data includes defect location coordinate information, stress concentration distribution data and defect hazard level.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images