A laser-ultrasonic high-resolution imaging method for defects in rough surface materials

Abstract

This invention discloses a high-resolution laser-ultrasonic imaging method for defects in rough surface materials, belonging to the field of laser-ultrasonic nondestructive testing technology. The method includes acquiring time-domain laser-ultrasonic A-scan signals, initially extracting feature signals related to defects, and generating feature signal intensity curves related to the scan spatial location encoding. Based on a Gaussian mixture model, the feature signal intensity curves are modeled as a linear combination of multiple Gaussian distributions. Regularization techniques are introduced, and parameters of each Gaussian distribution are obtained through iterative calculation using an expectation-maximization clustering algorithm. Based on the parameters of each Gaussian distribution, the ultrasonic signal intensity is labeled, dividing it into labels related to defects and labels unrelated to defects, thereby generating a high-resolution defect image. Therefore, this high-resolution laser-ultrasonic imaging method for defects in rough surface materials can achieve sub-millimeter-level defect detection under high-roughness surface conditions and is suitable for stable detection of multi-scale defects under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of laser ultrasonic nondestructive testing technology, and in particular to a laser ultrasonic high-resolution imaging method for defects in rough surface materials. Background Technology

[0002] Laser Ultrasonic Testing (LUT), as a non-contact, high-resolution, and broadband excitation non-destructive testing method, has shown significant potential in detecting internal defects in various materials. This technology uses a pulsed laser to excite ultrasonic waves within a sample and a laser interferometer to receive the response signal, thereby achieving defect imaging and quantitative characterization. However, the performance of LUTs is highly dependent on the optical state of the tested component's surface. In many practical industrial scenarios, component surfaces often have high roughness—for example, the undulations of cladding channels in additive manufacturing, the texture of composite material mold forming surfaces, and the unpolished surfaces of castings and forgings. These rough surfaces strongly scatter the incident laser, leading to reduced ultrasonic excitation efficiency and a decreased signal-to-noise ratio. Simultaneously, they complicate the sound wave propagation path, causing waveform distortion, increased noise, and artifact interference, severely limiting the ability to identify and locate defects.

[0003] Currently, the mainstream strategies for laser ultrasonic testing on highly rough surfaces include signal denoising and feature enhancement. Algorithms such as wavelet transform, empirical mode decomposition, unsupervised learning, and deep learning have been introduced to improve the signal-to-noise ratio. However, existing methods still have significant limitations: First, there is a lack of systematic modeling and statistical analysis of the noise characteristics caused by highly rough surfaces under different materials and processes, making it difficult to establish a mapping relationship between noise, roughness, and process. Second, existing signal processing algorithms mostly rely on fixed thresholds or pre-trained models, resulting in poor adaptability when faced with significant fluctuations in roughness between different workpieces and regions. Furthermore, most existing methods exhibit a sharp performance decline at high roughness levels (e.g., Ra > 15 μm), and intelligent algorithms heavily rely on high-quality labeled data, facing limitations in generalization ability and complex model debugging in real industrial environments, hindering their engineering applicability.

[0004] Therefore, in order to address the problems of strong noise interference, large signal attenuation, and insufficient adaptability of existing methods in laser ultrasonic testing of high rough surfaces, there is an urgent need to develop a high-resolution and robust laser ultrasonic imaging and signal processing technology that can adapt to various materials and rough working conditions, so as to achieve reliable detection and accurate characterization of minute defects in metal additive parts, composite materials and traditional metal components. Summary of the Invention

[0005] The purpose of this invention is to provide a high-resolution laser ultrasonic imaging method for defects in rough surface materials, which can identify minute surface defects, improve detection resolution, and is also suitable for stable detection of multi-scale defects under complex working conditions.

[0006] To achieve the above objectives, the present invention provides a laser ultrasonic high-resolution imaging method for defects in rough surface materials, comprising the following steps:

[0007] S1. Use a laser ultrasonic scanning device to perform a traversal scan of the area to be tested and collect time-domain laser ultrasonic A-scan signals.

[0008] S2. Based on the time-domain laser ultrasonic A-scan signal, and according to the correspondence between ultrasonic propagation distance, wave velocity and wavelength, windowing is applied to initially extract the characteristic signals related to the defect.

[0009] S3. Calculate the intensity of the feature signal, generate the feature signal intensity curve related to the scanning spatial position encoding, and perform noise reduction processing on the feature signal intensity curve to improve the sensitivity and accuracy of defect detection.

[0010] S4. Based on the Gaussian mixture model, the intensity curve of the feature signal after noise reduction is modeled as a linear combination of several Gaussian distributions. Regularization technology is introduced, and the mean vector, covariance matrix and mixture weight of each Gaussian distribution are obtained by iterative calculation through the expectation-maximization clustering algorithm.

[0011] S5. Based on the mean vector and covariance matrix of each Gaussian distribution, the ultrasonic signal intensity is labeled and divided into labels related to defects and labels unrelated to defects. A spatial distribution matrix of signal intensity is constructed, and then a high-resolution defect image is generated.

[0012] Furthermore, in S2, the characteristic signals related to defects are extracted as follows: the characteristic signal of surface defects is the transmitted surface wave, and the characteristic signals of shallow surface and internal defects are the defect scattering echo signals based on the time-delay superposition technique.

[0013] Preferably, regularization techniques are introduced in S4, including adding small positive values ​​on the diagonal of the covariance matrix and adaptively adjusting the regularization parameters through a differential evolution algorithm.

[0014] Preferably, the differential evolution algorithm adaptively adjusts the regularization parameter, including adjusting the regularization parameter through mutation and crossover operations, and obtains the optimal regularization parameter by minimizing the negative log-likelihood value as the objective function.

[0015] Preferably, the expectation-maximization clustering algorithm in S4 includes:

[0016] S41. Randomly set the initial parameters for each Gaussian component: mean vector, covariance matrix, and mixture weights;

[0017] S42. Iterative optimization is performed through the expected step and the maximization step, updating the parameter values ​​until the parameter change is lower than the threshold or the preset number of iterations is reached.

[0018] Furthermore, in S5, labels unrelated to defects are assigned a value of 0 or 1, while labels related to defects are normalized within the range of [0, 1].

[0019] Therefore, the present invention employs the above-mentioned high-resolution laser ultrasonic imaging method for defects in rough surface materials, which has the following technical advantages:

[0020] This invention introduces regularization technology and uses the expectation-maximization clustering algorithm to automatically distinguish signals based on the statistical regularity of the Gaussian mixture distribution of defective and non-defective regions. This effectively achieves accurate boundary segmentation of deep defects, adaptive enhancement of weak shallow signals, and intelligent suppression of surface noise. At the same time, the method of this invention is applicable to surfaces with different roughness, effectively controlling the false detection rate while maintaining high detection accuracy. Even under high roughness surface conditions, it can still reliably identify sub-millimeter level defects.

[0021] The technical solution of the present invention will be further described in detail below with reference to embodiments and accompanying drawings. Attached Figure Description

[0022] Figure 1 This is an example of a laser ultrasonic high-resolution imaging method for detecting defects in rough surface materials, showing defect images of SLM workpieces with diameters of 2.85 mm, 1.9 mm, 0.95 mm, 0.5 mm, 0.4 mm, and 0.3 mm.

[0023] Figure 2 This is an example of a laser-ultrasonic high-resolution imaging method for detecting defects in rough surface materials. The LMD workpieces with diameters of 2.85 mm, 1.9 mm, 0.95 mm, 0.5 mm, and 0.4 mm are shown in the defect images.

[0024] Figure 3 This is an example of a laser-ultrasonic high-resolution imaging method for detecting defects in rough surface materials, showing a near-surface defect image of an SLM workpiece with a depth of 0.5 mm and a diameter of 0.5 mm. Detailed Implementation

[0025] The present invention will be explained in more detail through the following embodiments. The purpose of disclosing the present invention is to protect all changes and modifications within the scope of the present invention. The present invention is not limited to the following embodiments.

[0026] Example 1

[0027] This invention provides a high-resolution laser ultrasonic imaging method for detecting material defects on rough surfaces. Taking metal additive manufacturing as an example, defect-free regions with surface roughnesses of 8.4 μm (selective laser melting (SLM) sample) and 37.5 μm (laser metal deposition (LMD) sample) were selected, and 5000 sampling points were randomly selected with a point spacing greater than 0.4 mm. Statistical analysis of the transmitted Rayleigh wave intensity data of the sampling points revealed that the transmitted Rayleigh wave intensity of both roughness samples conformed to a Gaussian distribution, and the average signal of the SLM sample with the smaller roughness (mean -5.28 dB) decreased by 31.20% compared to the LMD sample (mean -7.23 dB). Despite this, the standard deviations of the two samples were not significantly different, at 2.18 dB and 2.5 dB, respectively. This indicates that the irregular morphology of the rough surface enhances ultrasonic scattering, leading to signal energy attenuation and increasing the random fluctuation of the signal in spatial distribution. Furthermore, this invention also performed signal statistical analysis within the scanning region containing defects. Compared to the flawless sample, the signal intensity distribution of the defective sample exhibits a bi-Gaussian distribution: one Gaussian distribution with a lower mean corresponding to the flawless region, and another with a higher mean corresponding to the defective region. The mean and variance of the flawless region are consistent with those of the flawless sample; however, the signal mean values ​​of both roughness samples in the defective region are significantly lower than those in the flawless region (the mean of the SLM sample decreases to -18 dB, and the mean of the LMD sample is -15 dB), and the broadening of the distribution is significantly increased (the variance of the SLM sample increases to 3.05 dB, and the variance of the LMD sample increases to 3.13 dB). This is due to the reflection effect of the defect boundary, which leads to the amplitude attenuation of the directly transmitted surface wave signal.

[0028] In summary, the signal intensity distributions in both defective and defect-free regions exhibit Gaussian characteristics, but there are significant differences in their parameters. Therefore, this embodiment constructs a Gaussian mixture model (GMM) to estimate key parameters such as mean and variance, thereby achieving automatic differentiation between defective and defect-free regions. This effectively mitigates the impact of rough surface interference on detection accuracy and significantly improves the reliability of defect detection.

[0029] Gaussian Mixture Models (GMMs) can effectively model the signal intensity of defect features in a detection region as a linear combination of multiple Gaussian distributions, with the weight coefficients of each component reflecting the contribution of different distributions. However, parameter estimation of GMMs faces two core challenges: first, the latent Gaussian distribution to which the data points belong is a latent variable; second, the log-likelihood function of the model involves a linear combination of multiple Gaussian components, leading to analytical obstacles in its differentiation process. These characteristics make it difficult to directly apply traditional maximum likelihood estimation methods. To address this, this embodiment introduces the Expectation-Maximization (EM) clustering algorithm, which not only effectively solves the parameter estimation problem of Gaussian Mixture Models but also outputs the latent distribution probability of samples, thereby revealing the inherent clustering structure and generation mechanism of the data.

[0030] Assuming the input sample obey A Gaussian distribution with unknown parameters, each Gaussian distribution corresponding to a different mean. Covariance Matrix The Expectation-Maximization (EM) algorithm consists of three stages, as follows:

[0031] (1) Parameter initialization: Randomly set the initial parameters of each Gaussian component: mean vector Covariance matrix and mixed weights ,satisfy The constraints.

[0032] (2) Iterative optimization: This involves two core steps that are executed alternately:

[0033] Expected step (E-step): based on current parameters Calculate the posterior probability (responsibility value) of the latent variables. ):

[0034] ;

[0035] in, Characterizing the first The sample is from the first The posterior probability generated by each Gaussian component reflects the degree of membership of a data point to each component. As an index variable, it is used to represent the first index in the Gaussian mixture model. Each Gaussian component has a value ranging from 1 to... ; It is the first A Gaussian distribution.

[0036] Maximize step (M-step): Update parameters using the responsibility value obtained from the E-step.

[0037] ;

[0038] ;

[0039] ;

[0040] In the formula, , , For the first The parameters are updated using a Gaussian distribution.

[0041] The calculation terminates when the parameter change is below the threshold or when the preset number of iterations is reached.

[0042] The advantages of the EM method lie in its decomposition of complex joint optimization problems into analytically solvable alternating steps, maintaining probabilistic interpretability through a soft allocation mechanism, and providing reliable convergence guarantees, making it a classic method for handling probabilistic models with latent variables. However, this algorithm relies on the covariance matrix to describe the shape of the data distribution. In practical applications, due to noise or insufficient sample size, the covariance matrix may exhibit singularity or instability, leading to inaccurate parameter estimation and affecting the accuracy of clustering results. To address this issue, this embodiment also proposes an adaptive regularization parameter tuning strategy based on differential evolution to solve the singularity problem of the covariance matrix and effectively alleviate the instability in the EM clustering process, as follows:

[0043] First, a regularization technique is introduced by adding small positive values ​​to the diagonal of the covariance matrix to ensure its invertibility and prevent overfitting. The expression is as follows:

[0044] ;

[0045] In the formula, Represents the regularized sample Given model parameters , , The probability density function under; Indicates sample Given model parameters , , The probability density function under the given conditions.

[0046] Then, the regularization parameter is adaptively adjusted using the differential evolution algorithm. And minimize the negative log-likelihood value The objective function includes:

[0047] (1) Initialize the population: Randomly generate the initial value of the regularization parameter for the population. This serves as the starting point for subsequent iterative optimizations.

[0048] (2) Mutation operation: In each iteration, three different individuals are randomly selected. (in , , (For different indices), generate mutation vectors according to the difference formula. :

[0049] ;

[0050] In the formula, It is a variable factor that controls the magnitude of variation.

[0051] (3) Crossover operation: Combine the mutation vector with the current individual to generate new candidate solutions. To improve the diversity of optimizations:

[0052] ;

[0053] In the formula, It is the crossover probability, which controls the mixing ratio of the mutation vector with the current individual.

[0054] In summary, this invention provides a regularized EM clustering imaging method, as follows:

[0055] (1) Signal acquisition: In the area to be measured, each imaging scattering point is denoted as P( ), Used to refer to imaging scattering points , , These represent the x-axis, y-axis, and z-axis coordinate indices of the imaging scattering point, respectively, and each imaging scattering point corresponds to a fixed number n. A laser ultrasonic traversal grid scan is performed on the area to be tested, and the set of scanned detection points is denoted as... The original time-domain laser-ultrasound A-scan signal was acquired. Used to refer to scan points and These represent the x-axis and y-axis coordinate indices of the scan point, respectively.

[0056] (2) Signal strength calculation: The characteristic signals of surface defects (transmitted Rayleigh waves) and subsurface defects (scattered signals) are extracted from the original signal using a rectangular window function. The window position and width of the rectangular window are adaptively adjusted according to a preset physical model, such as the ultrasonic propagation distance, wave speed and wavelength model, in order to accurately extract the defect characteristic signals.

[0057] (3) Dimensional transformation: After considering geometric diffraction attenuation compensation, the intensity values ​​of ultrasonic characteristic signals at different locations are calculated in decibels (dB). Based on the mapping relationship between ultrasonic characteristic signal intensity and spatial location, a three-dimensional spatial distribution of ultrasonic signal intensity is formed. The three-dimensional spatial distribution map of ultrasonic signal intensity is transformed into a one-dimensional characteristic signal intensity curve based on the scattering imaging point number. , where n is the number of the imaging scattering point.

[0058] (4) Wavelet noise reduction: In this embodiment, the wavelet basis function is preferred. Wavelet reconstruction technology is used to perform wavelet decomposition and soft thresholding on the ultrasonic feature signal intensity curve related to spatial location number, suppressing high-frequency noise and preserving effective signal feature information.

[0059] (5) Adaptive Regularization: An adaptive regularization parameter tuning strategy based on differential evolution makes the objective function Reaching a stable value yields the optimal regularization parameter. This ensures the stability of the covariance matrix and addresses the singularity problem.

[0060] (6) EM clustering algorithm: Based on the expectation-maximization (EM) algorithm, the optimal regularization parameter is applied to the EM clustering algorithm. The expectation step and the maximization step are executed repeatedly until the parameters converge.

[0061] In practical applications, for surface defects, a Gaussian distribution with a higher mean corresponds to the "Noise" label, while a Gaussian distribution with a lower mean corresponds to the "Defect signal" label. Conversely, for shallow surface defects and internal defects, the correspondence between the mean and the label is reversed. In this way, the original data is divided into two categories: "Defect signal" labels related to defects and "Noise" labels unrelated to defects.

[0062] (7) Defect imaging: After the EM algorithm is executed, the three parameter combinations used to define the two Gaussian distribution models are obtained. , , These two Gaussian distribution models have different means and variances, corresponding to samples related to defects and samples unrelated to defects. Based on the clustering results, the one-dimensional... Convert to 3D The Gaussian distribution samples related to defects are labeled as "Defect signal," while those unrelated to defects are labeled as "Noise." The "Noise" label is assigned a value of 1 as a background element to reduce its impact on defect boundary identification; the "Defect signal" label is normalized within the range [0,1] to generate a high-resolution defect image for the extraction and evaluation of defect size, location, and shape features.

[0063] Example 2

[0064] This invention provides a high-resolution laser-ultrasonic imaging method for detecting defects in rough surface materials, applicable to surface defect detection in metal additive manufacturing. In this method, transmitted surface waves are used as characteristic signals, and the method of this invention is applied to detect and image sample defects.

[0065] like Figure 1 and Figure 2 As shown, the "Noise" tag, which is unrelated to defects, is assigned a value of 1, while the "Defectsignal" tag, which is related to defects, ranges from [0,1]. Figure 1 As can be seen from the figure, this embodiment achieves high-precision imaging of surface defects with diameters of 2.85mm, 1.9mm, 0.95mm, 0.5mm, 0.4mm and 0.3mm.

[0066] In LMD samples with high roughness surfaces ( Figure 2 The method of this invention successfully detected sub-millimeter-sized defects with a diameter of 0.4 mm, significantly improving the detection capability for rough surfaces compared to traditional methods. This further verifies that the method of this invention can effectively identify minute surface defects, overcoming the limitations of traditional ultrasonic imaging methods under rough surface conditions.

[0067] In other embodiments, the method of the present invention is not limited to metal additive manufacturing, but is also applicable to surface defect detection in composite materials or other metal processing.

[0068] Example 3

[0069] This invention provides a high-resolution laser ultrasonic imaging method for defects in rough surface materials, applicable to the detection of shallow (near-surface) and internal defects in metal additive manufacturing. In this method, the defect-scattered echo signal based on time-delay superposition technology is used as a characteristic signal, and the method of this invention is applied to detect and image the sample defects.

[0070] Delay-based superposition (DAS) technology refers to the scattering of ultrasonic waves generated by laser-induced ultrasonic excitation when they encounter defects. The propagation time of the scattered echoes depends on the position of the defect relative to the laser excitation point and the ultrasonic receiving point, as well as the depth of the defect. Specifically, time delay compensation is first applied to multiple received signals, with the compensation amount based on the wave velocity and propagation distance. For each pixel, the corresponding time delay is calculated, and then the compensated signals are accumulated. After time-delay superposition processing, the echo signal from the defect is enhanced, while background noise and incoherent signals are suppressed, thereby improving the signal-to-noise ratio of the defect and facilitating feature extraction.

[0071] This embodiment detected near-surface defects in the SLM sample, such as... Figure 3 As shown, the "Noise" label, which is unrelated to defects, is assigned a value of 0, while the "Defect signal" label, which is related to defects, ranges between [0,1]. It can be seen that the method of this invention can effectively separate defect signals from background interference, achieving high-resolution imaging of near-surface defects with a depth of 0.5 mm and a diameter of 0.5 mm, verifying the reliability and technical advantages of this method under complex working conditions.

[0072] In other embodiments, the method of the present invention is not limited to metal additive manufacturing, but is also applicable to the detection of shallow surface and internal defects in composite materials or other metal processing.

[0073] Therefore, the present invention employs the aforementioned high-resolution laser ultrasonic imaging method for defects in rough surface materials. By relying on the statistical law of the Gaussian mixture distribution between defect and non-defect regions, it automatically distinguishes signals, effectively achieving precise boundary segmentation of deep defects, adaptive enhancement of weak shallow signals, and intelligent suppression of surface noise. It can stably detect multi-scale defects under complex working conditions.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method of laser-ultrasonic high-resolution imaging of defects in a rough-surface material, characterized in that, The method comprises the following steps: S1, traversing scanning the region to be measured by using a laser ultrasonic scanning device, and collecting a time-domain laser ultrasonic A-scan signal; S2, for the collected time-domain laser ultrasonic A-scan signal, according to the corresponding relationship among ultrasonic wave propagation distance, wave speed and wavelength, using windowing processing, preliminarily intercepting a characteristic signal related to a defect; S3, calculating the intensity of the characteristic signal, generating a characteristic signal intensity curve related to a scanning space position code, and performing noise reduction processing on the characteristic signal intensity curve; S4, based on a mixed Gaussian model, modeling the noise-reduced characteristic signal intensity curve into a linear combination of a plurality of Gaussian distributions, introducing a regularization technique, and obtaining a mean vector, a covariance matrix and a mixing weight of each Gaussian distribution through iterative calculation of an expectation maximization clustering algorithm, wherein the regularization technique comprises adding a small positive value on the diagonal line of the covariance matrix, and adaptively adjusting the regularization parameter through a differential evolution algorithm, the adaptive adjustment of the regularization parameter by the differential evolution algorithm comprises adjusting the regularization parameter through mutation operation and crossover operation, and taking the minimization of a negative log-likelihood value as an objective function to obtain an optimal regularization parameter; S5, according to the mean vector and the covariance matrix of each Gaussian distribution, marking the ultrasonic signal intensity and dividing it into a label related to a defect and a label unrelated to a defect, constructing a spatial distribution matrix of the signal intensity, and further generating a high-resolution defect image.

2. A method of laser-ultrasonic high-resolution imaging of defects in a rough surface material according to claim 1, characterized in that, In S2, the characteristic signal related to the defect includes: the characteristic signal of the surface defect is a transmitted surface wave, and the characteristic signal of the shallow surface and internal defect is a defect scattering echo signal based on the delay-and-sum technique.

3. A method of laser-ultrasonic high-resolution imaging of defects in a rough surface material according to claim 1, characterized in that, The expectation maximization clustering algorithm in S4 comprises: S41, randomly setting initial parameters of each Gaussian component: mean vector, covariance matrix and mixing weight; S42, performing iterative optimization through expectation step and maximization step, updating parameter values until the parameter variation is less than a threshold value or the preset iteration number is reached.

4. The method of claim 1, wherein, In S5, the label unrelated to the defect is assigned a value of 0 or 1, and the label related to the defect is normalized in the range of [0, 1].

Images