Ultrasonic nondestructive testing method and system for welding seam of steel structure

Abstract

The invention relates to the technical field of nondestructive testing, and discloses an ultrasonic nondestructive testing method and system for a welding seam of a steel structure. The method comprises the following steps: adopting a self-adaptive coupling ultrasonic probe to carry out signal acquisition on the surface of a welding seam of a steel structure to obtain original ultrasonic signal data; performing multispectral adaptive wavelet transform processing on the original ultrasonic signal data to obtain an enhanced ultrasonic signal; performing blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal; performing triangular constraint average interpolation and adaptive extended Kalman filtering fusion processing on the separated weld defect signals to obtain filtered weld defect feature data; and performing regularization Gaussian field weld contour reconstruction and defect positioning processing based on the filtered weld defect feature data to obtain a weld contour model and defect position data. According to the invention, the stability and reliability of weld defect detection are improved, and high-precision detection of weld internal defects is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing, and in particular to a method and system for ultrasonic nondestructive testing of steel structure welds. Background Art

[0002] Traditional weld inspection methods include visual inspection and penetrant testing, but these methods can only detect surface defects and cannot effectively detect hidden problems such as internal pores, cracks, and lack of fusion. Ultrasonic non-destructive testing technology has received widespread attention due to its ability to detect internal defects in welds, but it still faces many challenges in practical applications. The construction site environment is complex, and factors such as electromagnetic interference and vibration noise seriously affect the quality of ultrasonic signals, resulting in a decrease in the reliability of the test results. Secondly, the surface of steel structure welds usually has irregular characteristics, which makes it impossible for traditional ultrasonic probes to maintain good contact with the weld surface, resulting in low energy transmission efficiency, further reducing the detection accuracy.

[0003] Existing ultrasonic weld detection methods have obvious deficiencies in signal processing. Conventional signal noise reduction technologies such as filtering and threshold methods have limited effectiveness when processing ultrasonic signals in complex environments, and it is difficult to effectively distinguish useful signals from noise. Especially when there is overlap between the signal and noise in the frequency domain, traditional methods are prone to cause the loss of useful signals or the introduction of pseudo-features. At the same time, the defect features in the ultrasonic signal are often masked by complex environmental noise, making it difficult for the detection system to accurately identify and locate defects. In addition, for the discrete signal points obtained by ultrasonic detection, the existing methods lack an effective spatial constraint mechanism and cannot fully utilize the correlation information between adjacent detection points, resulting in insufficient accuracy in weld contour reconstruction and inaccurate defect positioning. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method and system for ultrasonic nondestructive testing of steel structure welds. The present invention improves the stability and reliability of weld defect detection and realizes high-precision detection of internal defects in welds.

[0005] To achieve the above object, the present invention provides a method for ultrasonic nondestructive testing of steel structure welds, comprising the following steps: Adopting adaptive coupling ultrasonic probe to collect signals from the weld surface of steel structure and obtain original ultrasonic signal data; Performing multi-spectral adaptive wavelet transform processing on the original ultrasonic signal data to obtain an enhanced ultrasonic signal; performing blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal; Performing triangle constrained average interpolation and adaptive extended Kalman filter fusion processing on the separated weld defect signal to obtain filtered weld defect feature data; Regularized Gaussian field weld contour reconstruction and defect location processing are performed based on the filtered weld defect feature data to obtain a weld contour model and defect location data.

[0006] The present invention also provides a steel structure weld ultrasonic nondestructive testing system, comprising: The signal acquisition module is used to collect signals from the weld surface of the steel structure using an adaptive coupled ultrasonic probe to obtain raw ultrasonic signal data; A transformation processing module, configured to perform multi-spectral adaptive wavelet transformation on the original ultrasonic signal data to obtain an enhanced ultrasonic signal; A blind source separation module is used to perform blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal; An interpolation and fusion module is used to perform triangle constrained average interpolation and adaptive extended Kalman filter fusion processing on the separated weld defect signal to obtain filtered weld defect feature data; The reconstruction and positioning module is used to perform regularized Gaussian field weld contour reconstruction and defect positioning processing based on the filtered weld defect feature data to obtain a weld contour model and defect position data.

[0007] In summary, the adaptively coupled ultrasonic probe can adapt to the irregularities of steel structure weld surfaces, ensuring full contact between the probe and the weld surface, optimizing ultrasonic energy transmission efficiency, and improving the quality of the original ultrasonic signal. A dual-layer signal enhancement mechanism is formed by combining multispectral adaptive wavelet transform processing with blind source separation. This effectively suppresses ambient noise while preserving weld defect characteristics, enhancing defect characteristics in the ultrasonic signal and improving the signal-to-noise ratio. Triangular constrained average interpolation and adaptive extended Kalman filter fusion processing utilize the spatial constraints between adjacent detection points to effectively identify and correct outliers during the detection process, improving the stability and reliability of weld defect detection. The weld contour reconstruction and defect localization method based on a regularized Gaussian field introduces prior knowledge constraints to maintain the local structural stability of the transformed point set, demonstrating excellent accuracy and robustness in contour reconstruction and defect localization of welds with complex geometries. The present ultrasonic nondestructive testing method for steel structure welds can overcome noise interference and the influence of weld surface irregularities in complex environments, achieving high-precision detection of internal weld defects, providing engineers with comprehensive and reliable weld quality assessment results and effectively supporting steel structure safety assessment decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a schematic diagram of the steps of a method for ultrasonic nondestructive testing of steel structure welds in one embodiment of the present invention; Figure 2This is a structural block diagram of an ultrasonic nondestructive testing system for steel structure welds in one embodiment of the present invention.

[0009] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0011] Reference Figure 1 This embodiment provides a method for ultrasonic nondestructive testing of steel structure welds, comprising the following steps: S1, uses an adaptive coupled ultrasonic probe to collect signals from the weld surface of the steel structure to obtain raw ultrasonic signal data; An adaptively coupled ultrasonic probe is placed on the surface of a steel weld, ensuring a tight and uniform interface with the weld surface. The probe utilizes a coupling layer made of flexible polyurethane composite microspheres, effectively adapting to surface irregularities and microscopic convexities and concavities. Its elastic deformation properties ensure good acoustic coupling between the probe and the weld, thus preventing energy loss or signal distortion during ultrasonic signal transmission and reception. Once the probe is secured, a built-in pulse transmission circuit excites the ultrasonic sensor array to synchronously emit a high-frequency, short-pulse ultrasonic signal. The excitation signal frequency is controlled between 2.5MHz and 5MHz, and the signal propagates perpendicularly to the weld surface, entering the steel structure. This allows the ultrasonic wave to longitudinally scan the weld's internal structure as it passes through the weld metal matrix. When the propagating ultrasonic wave encounters discontinuities within the weld, such as material interfaces, pores, cracks, slag inclusions, or lack of fusion, it is reflected due to changes in the acoustic impedance of the medium, generating an ultrasonic echo signal. The return signal is then received in real time by the corresponding receiving unit in the probe array. The received echo signal has a low amplitude. To ensure signal measurability and the dynamic range for subsequent analysis, it undergoes gain adjustment and signal amplification processing via a preamplifier. The preamplifier gain range is designed to be 20 to 60 decibels. The specific gain is dynamically adjusted based on the echo signal strength and the material properties of the weld being inspected, balancing signal-to-noise ratio improvement and signal amplitude stability. The amplified analog signal remains a continuous-time signal. To achieve digital processing, it is input into a high-precision analog-to-digital converter for discretization. The analog-to-digital converter has 14-bit resolution and a sampling frequency set to 30MHz. This fully preserves the subtle time-domain variation characteristics of the ultrasonic signal and avoids high-frequency information loss or sampling aliasing. After the analog-to-digital conversion is complete, the digital signal is transmitted to the back-end processing module via a high-speed data bus (such as USB 3.0 or higher), forming a structurally complete, time-consistent, and spatially matched raw ultrasonic signal data set.

[0012] S2, performing multi-spectral adaptive wavelet transform on the original ultrasonic signal data to obtain an enhanced ultrasonic signal; Specifically, the original ultrasonic signal is input into a wavelet decomposition system. Based on the distribution of local signal features in both time and frequency, the system decomposes the original time-domain signal into multiple scale levels, each representing signal components within a different frequency range. A wavelet function with excellent edge-preserving capabilities is selected to perform a five-level decomposition of the signal, yielding multiple signal components ranging from high-frequency details to low-frequency background. This multi-level processing effectively distinguishes between abrupt signals caused by internal weld defects and continuous interference caused by background noise. Spectral energy analysis is performed on the wavelet decomposition coefficients at different frequency scales, and an energy relationship model between signal and noise is established to identify components that represent valid information and invalid interference. Based on this, a threshold dynamically set based on the noise level is automatically generated for each signal level. These thresholds are used to determine which signal components should be retained or removed. To avoid traditional thresholding methods from accidentally removing useful information while removing noise, an improved wavelet thresholding strategy preserves the amplitude characteristics of the signal's principal components and applies a flexible compensation to signals below the threshold but with certain pulse characteristics. This maximizes the preservation of the spike structure of the weld defect signal and prevents over-smoothing of the signal during processing. After processing all wavelet layers, the layers are reconstructed back into the time domain through an inverse transform operation, forming a new enhanced ultrasonic signal curve. To improve the accuracy of the final signal, waveform optimization is performed at edge points and peak regions. For example, local window smoothing is used to eliminate edge oscillations introduced by the reconstruction process, or a peak recognition mechanism is used to correct the positioning errors of signal mutation points. The result is an ultrasonic signal sequence with clear contours, effective noise suppression, and significantly enhanced defect response characteristics.

[0013] S3, performing blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal; It should be noted that the multi-channel enhanced signals are normalized to eliminate numerical inconsistencies caused by differences in signal strength, acquisition gain deviation, or coupling states across channels. This ensures that all channel signals have a uniform amplitude distribution and energy reference, facilitating unified calculations in subsequent algorithm processing. After normalization, a multi-channel matrix structure containing multiple time-series signals is obtained, with each column representing the normalized waveform data of an independent channel. Cross-correlation analysis is performed on this normalized signal matrix. By calculating the degree of temporal synchronization and waveform similarity between any two channels, a coefficient matrix reflecting the correlation between all channels is constructed. Based on this coefficient matrix, a separation optimization algorithm for signal unmixing is introduced. With the goal of maximizing signal statistical independence, the separation process is guided by the negative entropy criterion defined in information theory. Negative entropy, as a measure of the degree to which a signal deviates from a Gaussian distribution, effectively improves the ability to distinguish non-Gaussian independent source signals in blind source separation. The separation optimization algorithm analyzes the negative entropy distribution characteristics of the signal to preliminarily calculate an initial separation matrix for signal decoupling. To improve separation accuracy and enhance the prominent characteristics of the target defect signal, a sparsity constraint mechanism is introduced based on the initial separation matrix. This constraint mechanism assumes that weld defect signals exhibit significant pulses and local concentrations, exhibiting typical sparsity characteristics, with a small proportion of the overall signal but distinct characteristics. During the separation optimization process, the matrix weight structure is adjusted to favor the preservation of this structural characteristic. A Newton iteration algorithm is used to rapidly optimize the target separation matrix. This iterative process continuously updates the separation matrix's structural parameters and evaluates its convergence in terms of sparsity and independence until the preset accuracy requirements are met or the maximum number of iterations is reached, resulting in a stable and reliable target separation matrix. The target separation matrix is used to linearly transform the standardized multi-channel signal matrix to generate multiple statistically independent signal components. Each independent signal component represents a potential independent component in the original mixed signal, including the weld defect reflection signal, background noise, and other irrelevant interference signals. Each independent signal component is subjected to characteristic analysis and reconstruction evaluation, including observation of temporal pulse characteristics, energy concentration, frequency variation behavior, and peak response morphology. Signal components with distinct defect characteristics are identified by analyzing these indicators. The selected valid signals are reconstructed to form a clear, independent weld echo signal sequence that reflects the physical characteristics of the defect.

[0014] In this embodiment, expectation operations are performed on each channel signal pair in a standardized multi-channel signal matrix. By calculating their joint fluctuation trends within a time window, a correlation coefficient matrix reflecting the mutual correlation between channels is obtained, expressing whether there is a synchronous response phenomenon or potential mixed characteristics of structural coupling in each channel pair. This correlation coefficient matrix is then subjected to eigenvalue decomposition, converting it into a signal feature space composed of eigenvectors. By retaining the main eigencomponents, a dimensionality reduction representation of the signal channel information is achieved, preserving most of the energy distribution and structural differences while reducing the computational complexity of subsequent modeling and optimization. Based on this, a signal probability distribution model is established based on the set of signal eigenvectors. The degree to which the signal deviates from a uniform or Gaussian distribution is described by statistically analyzing the signal distribution behavior in the reduced-dimensional space. To measure signal independence, a reference variable set with the same covariance structure but fully Gaussian characteristics is constructed for calculating the information entropy difference with the real signal. Since Gaussian signals have the characteristic of maximum entropy, this information entropy difference serves as a quantitative indicator of signal non-Gaussianity, thereby forming a negative entropy function that measures the degree of independence between signal sources. A sparsity constraint term is introduced based on the negative entropy function to enhance the algorithm's response to pulse-like defect signals. The sparsity constraint reflects the characteristic of weld defect signals: despite their small proportion in the overall channel matrix, their energy is concentrated and their morphology is prominent. Therefore, the addition of this term effectively improves the algorithm's sensitivity to weld defect detection. The negative entropy term and the sparsity constraint term are combined to form a composite objective function. The optimization calculation is performed using a Newton iteration method. The separation weight parameters are continuously adjusted starting from the initial state. The derivative and curvature of the composite objective function are evaluated during each iteration, and the separation coefficients are modified using the current gradient direction to obtain the optimal set of separation weight coefficients. This set of weight coefficients is then formally reorganized to construct the original separation matrix, which contains the decoupling structure required to transform the mixed channel signal into several independent signal components. During the iterative optimization process, a precision control mechanism is implemented to determine whether the optimization has reached a stable state by comparing the matrix norm difference between the matrix obtained in the current iteration and the result of the previous iteration. When the difference is less than the preset threshold, it indicates that the weight coefficient has converged to a relatively stable structural state. The original separation matrix is used as the final result for orthogonal normalization to ensure that the row vectors of the final separation matrix remain perpendicular to each other and have the same length in a geometric sense, thereby ensuring the linear independence between the signal source components and preventing overfitting or incomplete decoupling problems in the subsequent signal transformation process, thereby generating the initial separation matrix.

[0015] S4, performing triangle constrained average interpolation and adaptive extended Kalman filter fusion processing on the separated weld defect signal to obtain filtered weld defect feature data; Specifically, structured feature extraction is performed on the separated weld defect signals. The most representative characteristic parameters of the signal—namely, peak amplitude, arrival time, waveform width, and signal energy—are selected. These four data items are then constructed as the first eigenvector corresponding to each detection point. All detection points within the detection area are numbered according to a fixed grid rule. Delaunay triangulation is then performed based on their spatial positional relationships, connecting all detection points into a network of triangles that adhere to the maximum and minimum angle principles. This constructs a detection point constraint network with a stable topological structure and spatial constraint logic. The first eigenvectors of the spatial constraint network for multiple detection points are then calculated for difference values. The eigenvalue differences between the two endpoints on the connecting edges within each triangle are calculated one by one. These differences are then compared with the preset characteristic standard deviation values for the normal weld area to determine whether there are any abnormal feature changes. If the eigenvector of a detection point exhibits a significant abrupt change compared to its neighboring points and exceeds the allowable standard deviation range, the point is marked as suspicious, indicating that it is on the edge of a weld defect or that a measurement error exists. On this basis, to improve the spatial continuity and robustness of the data, a triangle-constrained average interpolation strategy is employed for suspicious detection points. Several adjacent points forming a triangle relationship with the suspicious points that have not been marked as abnormal are selected. A weighted average calculation is performed based on the eigenvector information of these reference points to compensate for the eigenvalues of the suspicious points and generate corrected feature data. This corrected feature data is then fed into an extended Kalman filter for state prediction and measurement update. Based on a nonlinear state transition model and an observation model, this filter dynamically adjusts the prediction path based on the interaction between the spatial topology and feature evolution of the detection points, while also performing corrections and updates based on newly acquired measurement data. During the prediction process, feature trends are extended based on the previous state estimate. After each prediction step, the difference between the prediction error and the observation error is evaluated and fed back into the filter state to adjust the state covariance matrix and noise parameter weights, thereby adaptively optimizing each state update. During the filtering iteration, the eigenvectors of all detection points maintain a good balance between spatial structural continuity and temporal logical consistency, ultimately yielding weld defect feature data that has undergone temporal dynamic compensation and spatial structural correction.

[0016] In this embodiment, based on a triangular constraint network structure, for each detection point identified as a suspicious point, the set of directly adjacent nodes in the triangular network is traversed. Points also marked as suspicious are filtered out, leaving only unmarked neighboring points determined to be stable and with reliable eigenvalues. This constitutes a valid set of adjacent points that can be used as a reference for interpolation. Based on the spatial coordinates of the suspicious point, the geometric distance between it and each valid adjacent point is calculated. These distances reflect the influence of each reference point on the central point during interpolation. To reflect the spatial attenuation of this influence, the inverse distance principle is used to construct weight coefficients, where closer points have a greater influence, while farther points contribute less to the calculation. This results in a set of spatial weight coefficients. The first eigenvector of each valid adjacent point is multiplied by its corresponding spatial weight coefficient to obtain the weighted feature component value of that point's contribution to the correction process. All these weighted component values are accumulated and normalized so that the sum maintains the relative proportions of the original features without introducing additional offsets, resulting in a corrected feature vector. This correction vector represents a reasonable estimate of the original suspicious point in the current spatial structure, incorporating all credible neighborhood information. To verify the reliability of the correction result and prevent correction errors from propagating to subsequent filtering steps, a Euclidean distance calculation is performed between this corrected feature vector and the second feature vector estimated at the previous moment or stage for the suspicious point to assess the degree of deviation between the two in the feature space. If this distance is within an acceptable range, indicating a high degree of consistency in the interpolation, the generated correction vector is used directly as the final result. Otherwise, the system activates an error control mechanism, dynamically adjusting the interpolation method or reference point composition to ensure that the correction value of each suspicious point not only maintains spatial continuity but also reflects the true signal evolution trend. After completing the Euclidean distance verification, the corrected feature vector replaces the original suspicious point feature data and is written to the corresponding location in the detection network, updating the feature distribution state across the entire detection grid to form the corrected feature data.

[0017] S5, performing regularized Gaussian field weld contour reconstruction and defect location processing based on the filtered weld defect feature data to obtain a weld contour model and defect location data.

[0018] For each weld geometry type being inspected, such as fillet welds, T-welds, or butt welds, a matching standard weld geometry model is selected from a standard model database and used as the initial reference model for reconstruction. This initial model includes contour control points, structural boundaries, and connection logic, and serves as a benchmark framework for subsequent transformations and fitting. Based on the weld defect feature data obtained through Kalman filtering and interpolation in the previous stage, the three-dimensional spatial coordinates and corresponding confidence values of each inspection point are extracted. These data points together constitute a set of contour feature points that are structurally representative and physically reliable for the inspected object in space. This set of contour feature points is mapped to a Gaussian probability distribution using probabilistic modeling. This involves combining multiple Gaussian distribution functions to describe the probability of each inspection point in three-dimensional space and the strength of the defect signal. Each feature point is modeled as a Gaussian function centered at its spatial coordinate, with its confidence value affecting its diffusion. These Gaussian functions are superimposed to form a continuous weld defect probability distribution map, demonstrating the spatial distribution density and intensity of the defect. Based on the initial model and the defect probability distribution map, a comprehensive evaluation energy function consisting of fitting error, transformation continuity, and prior constraints is constructed. The fitting error term measures the geometric difference between the contour feature point set and the standard model; the transformation continuity term ensures that the entire contour does not undergo abrupt changes during the transformation process, ensuring smoothness of the structural curve and boundary coherence; and the prior constraint term, by incorporating prior knowledge of the weld geometry, defect distribution characteristics, and ultrasonic echo physical response as constraints, guides the model to more accurately approximate the actual physical structure during the optimization process. After these three terms are combined into a unified optimization objective function, the L-BFGS quasi-Newton optimization algorithm is used to minimize the energy function. By continuously updating the transformation parameters and evaluating the energy reduction trend, the positions of the model control points, the contour boundary shape, and the weight distribution of the probability function are adjusted in each iteration to obtain an optimal contour transformation parameter set that closely matches the actual weld structure. Based on the optimal transformation parameters, the initial standard model is mapped to the weld spatial shape corresponding to the current inspection data. Combined with the confidence value of each point in the defect probability distribution map, it is determined whether there are continuous regions in the space with a confidence value exceeding a preset threshold. When the defect probability corresponding to a certain spatial area in the transformation model exceeds the threshold, the area is determined to be a suspected defect area. The output includes the weld contour reconstruction result and the defect location data including the spatial location, size, shape and distribution density of each defect.

[0019] In this embodiment, the optimization environment of the comprehensive evaluation energy function is initialized, including setting the initial temperature parameters and cooling rate. The temperature parameter is used to regulate the jumping ability of the search space in each round of iteration, while the cooling rate is used to constrain the convergence rhythm of the optimization process. The two determine the exploration ability of the algorithm in the early stage and the accuracy stability in the later stage. The initial state set is constructed as an identity transformation structure equivalent to the unit mapping, ensuring that the initial state of the algorithm maintains structural alignment with the standard weld model and establishing a stable reference system for subsequent deformation adjustment. Based on the initial state set, the gradient information of the current comprehensive evaluation energy function in the multidimensional parameter space is calculated, that is, its first-order derivative and approximate second-order derivative are obtained respectively, as a basis for measuring the model transformation direction and step size adjustment in the current state. The L-BFGS algorithm adopts a quasi-Newton optimization framework. Without the need to accurately construct the complete Hessian matrix, it dynamically constructs a search vector for correcting the current transformation parameters by accumulating historical gradients and step size information. In each iteration, the current transformation parameters are locally updated based on the gradient direction and energy decay rate to obtain an updated state. This process is repeated continuously, evolving toward the direction corresponding to the minimum value of the energy function. In this way, the transformation parameters are gradually adjusted to the optimal structural state that maximizes the fit of the weld feature point set, maintains contour continuity, and satisfies the a priori structural constraints. This optimization process continues until the difference between the energy function values obtained in two consecutive iterations is less than the system's set energy target. This indicates that the current model morphology has stabilized and the marginal benefit of parameter adjustment is approaching zero. The system then determines that the transformation parameters corresponding to the current state are the final optimal solution. After obtaining this set of optimal transformation parameters, they are applied to a standard weld model to perform structural mapping, resulting in a weld contour reconstructed model that is highly aligned with the actual inspection data. Simultaneously, based on the defect response intensity and confidence information corresponding to each feature point in the transformed model, the defect probability representation of that point in three-dimensional space is calculated. The defect probabilities of all points are then aggregated to form a continuous defect confidence map. A defect confidence threshold is set as the identification criterion, and the entire defect confidence map is traversed and analyzed. When the confidence value in a spatial region continuously exceeds the threshold, the region is identified as a suspected defect area. The spatial region marked as a defect is superimposed on the reconstructed weld contour model, and the output includes multi-dimensional defect location data such as weld geometry, spatial location of the defect area, defect volume extension, and main axis direction.

[0020] In one example, an adaptive coupled ultrasonic probe is used to collect signals from the weld surface of a steel structure to obtain raw ultrasonic signal data, including: An adaptive coupling ultrasonic probe is placed on the surface of the steel structure weld, and a transmitting pulse excitation is performed through the adaptive coupling ultrasonic probe to obtain an ultrasonic signal; Control the ultrasonic signal to propagate into the steel structure in a direction perpendicular to the weld surface, and generate a reflected wave when encountering the internal interface or defect of the weld, and receive the ultrasonic echo signal reflected from the inside of the weld; The ultrasonic echo signal is input into the preamplifier for gain amplification to obtain an amplified ultrasonic signal, and the amplified ultrasonic signal is converted into digital form to obtain the original ultrasonic signal data.

[0021] In this example, an adaptive coupling ultrasonic probe is stably positioned on the surface of a steel structure weld, ensuring an efficient acoustic energy conduction path between the probe and the weld. The probe incorporates an adaptive coupling layer made of a blend of flexible polyurethane and highly elastic microspheres. This layer forms a uniform bond with the weld surface. Even with weld excess height, weld pool irregularities, residual flash, or other subtle geometric perturbations, the adaptive coupling layer actively adapts to the surface contour by deforming, effectively minimizing energy attenuation caused by air pockets and acoustic impedance discontinuities. This structure maintains a stable acoustic coupling quality under the application of a certain external pressure, ensuring effective excitation and clear reception of subsequent ultrasonic signals. Once the probe is positioned, a built-in excitation circuit applies narrow pulsed electrical signals to the probe's multi-channel piezoelectric transducer array. These pulses, stimulated by the piezoelectric material, are rapidly converted into longitudinal ultrasonic signals with stable frequency and short waveforms, which are then emitted along the probe's acoustic beam axis toward the weld surface. Because the internal channel layout of the probe array adopts a matrix structure and the acoustic beam propagation path is precisely calibrated, all transmitted signals are controlled to propagate in a direction perpendicular to the weld surface. This vertical propagation path can minimize the impact of the beam incidence angle error on the reflection path. After entering the metal structure, it can directly penetrate the weld volume vertically, forming a symmetrical echo in the defect-free state and a strong discontinuous reflection wave in the presence of defects, thereby enhancing the response sensitivity to abnormal interfaces within the weld. After entering the weld metal matrix, the ultrasonic signal propagates along the original path to the metallurgical structure area inside the weld. If the weld is a continuous and homogeneous material, the signal will continue to propagate and eventually return after reflecting from the bottom surface. However, if the signal encounters an area with a sudden change in acoustic impedance such as cracks, slag inclusions, pores, or lack of fusion, it will produce one or more irregular reflections at the defect interface. These reflected waves will return along the path and return to the probe receiving channel with different time delays. The receiving end, another component of the probe array, monitors the echo signal from within the metal in real time and feeds it back to the signal processing module in the form of an analog voltage signal. Because the reflected wave signal from some defects is low in amplitude and contaminated with ambient noise, a preamplifier module performs signal gain processing. Designed between the ultrasonic probe and the analog-to-digital converter, the preamplifier boosts the weak analog voltage signal received to an appropriate amplitude range, enhancing the signal's dynamic range and resolution. To accommodate signal amplitude variations caused by varying weld materials, structural thicknesses, and excitation energy, the preamplifier's gain is adjustable, ranging from 20 to 60 decibels. Dynamic signal amplitude correction is achieved through automatic or manual control. When the signal amplitude is insufficient, the gain is automatically increased to prevent the signal from sinking below the noise floor. When the signal amplitude is excessive and approaches saturation, the system automatically adjusts the gain to prevent signal distortion, ensuring that the output signals of all channels remain within the effective operating range of the analog processor.After the analog signal has been amplified by the preamplifier, it enters the analog-to-digital converter for digitization under conditions of stable amplitude and structural integrity. As a key link in signal discretization, the analog-to-digital converter periodically samples the continuous-time signal according to the set sampling frequency and sampling accuracy, and converts the analog voltage at each time point into the corresponding digital output. In this process, the sampling frequency is higher than twice the maximum frequency component in the signal to avoid information folding and spectrum aliasing. The value range is above 30MHz, and the sampling accuracy is directly related to the ability to restore details in subsequent signal analysis and is required to reach 14 bits or more. After sampling is completed, all channel signals are combined into a structured multi-channel raw data frame and sent to the back-end data processing unit in real time via a high-speed data interface such as USB3.0 or Ethernet transmission channel.

[0022] In one example, the original ultrasonic signal data is subjected to multispectral adaptive wavelet transform processing to obtain an enhanced ultrasonic signal, including: Perform wavelet decomposition on the original ultrasonic signal data to obtain wavelet decomposition coefficients at different frequency scales; The spectrum energy analysis is performed on the wavelet decomposition coefficients of different frequency scales, and a quadratic polynomial fitting algorithm is used to establish a spectrum energy and noise correlation model to obtain the standard deviation of noise in each layer. Based on the standard deviation of noise and coefficient length of each layer, the adaptive threshold corresponding to each decomposition layer is calculated, and nonlinear compensation is performed on the wavelet decomposition coefficient according to the adaptive threshold to obtain the denoised wavelet coefficient; The denoised wavelet coefficients are reconstructed by inverse wavelet transform to obtain the ultrasonic signal after preliminary noise reduction; The ultrasonic signal after preliminary noise reduction is processed by peak feature preservation and pseudo-Gibbs phenomenon suppression to obtain an enhanced ultrasonic signal.

[0023] In this example, the raw ultrasonic signal data is input into a wavelet decomposition module, where the signal is decomposed at multiple levels using wavelet basis functions to extract local variation characteristics at different frequency scales. Wavelet functions with symmetry and peak retention are selected, such as the Symlet-type function with compact support. A wavelet decomposition structure with five or more layers is established. Through the cascade of multiple low-pass and high-pass filter banks, the original signal is decomposed into a low-frequency approximation signal and multiple high-frequency detail signals. As the decomposition level increases, the low-frequency component gradually captures the global trend of the signal, while the high-frequency component retains the signal's mutation information and spike echoes. This decomposition method allows for a preliminary separation of background noise and defect echoes in the ultrasonic wave in the frequency domain. After wavelet decomposition, the wavelet coefficients obtained at each frequency scale are subjected to spectral energy analysis to extract the intensity characteristics of the information contained therein. By calculating the energy density of each wavelet coefficient layer, the concentration of energy distribution in the signal layer is assessed and compared with known noise distribution characteristics. On this basis, a quadratic polynomial fitting algorithm is introduced to model the correlation between spectral energy and noise. By training and optimizing the fitting curves of energy distribution and noise interference in multiple frequency bands, a mathematical model describing noise behavior is formed. This model estimates the noise interference intensity across the entire frequency domain, thereby deriving the noise standard deviation corresponding to each layer of wavelet decomposition coefficients. Based on the noise standard deviation of each layer and the length of the current layer coefficients (i.e., the number of data points included), a set of dynamic adaptive thresholds is constructed. These thresholds vary with the noise intensity and sample size of different frequency layers. To avoid the inadvertent deletion of high-frequency information or the residual low-frequency noise caused by the use of a uniform static threshold, an adaptive threshold mechanism applies a differentiated suppression strategy to each wavelet coefficient sequence in a layered and customized manner. During denoising, a flexible compensation strategy that introduces a nonlinear compensation mechanism is adopted to preserve the main peak shape of the signal while reducing the impact of low-amplitude fluctuations on signal clarity. For each wavelet coefficient, if its amplitude is significantly higher than the threshold of this layer, the coefficient is partially retained, and the degree of retention varies with the amplitude of its excess threshold. For coefficients slightly above or below the threshold, they are smoothed and down-weighted through nonlinear compression processing, thereby suppressing noise while maximizing the integrity of the signal edge information and echo structure. After this processing stage, the set of wavelet coefficients is transformed from a state containing the original signal and a large number of interference items to a denoised state that emphasizes significant features and suppresses background interference. The processed wavelet coefficients of each layer are re-input into the inverse wavelet transform module. Through inverse filtering and merging reconstruction operations, the process of regression from the frequency domain to the time domain is completed while maintaining the local characteristics of the signal, forming a preliminary denoised ultrasonic signal.Due to the inevitable filter window effect during wavelet decomposition and reconstruction, "pseudo-Gibbs phenomena"—unphysical oscillations or edge blurring—occur near signal edges or breakpoints. Based on the initial denoised signal, two post-processing operations, peak feature preservation and pseudo-Gibbs phenomenon suppression, are performed. In the peak feature preservation phase, a short-duration pulse location algorithm is introduced to perform local window detection on all significant rising edges, steep peaks, and short-period, high-amplitude signals in the ultrasonic signal. These signals are then identified within the time series and protected against spikes. Signals in these areas are then restored to maximum strength during the denoising process, ensuring that reflections from minor defects are not mistakenly identified as noise and subsequently deleted. In the pseudo-Gibbs suppression phase, the abnormal relationship between the rate of change of the signal boundary and the oscillation amplitude is detected to locate the locations of reconstructed oscillations. These segments are then compensated using smooth interpolation and edge correction algorithms, ensuring that the signal maintains its sharpness while achieving greater physical plausibility and temporal continuity. The result is an enhanced ultrasonic signal.

[0024] In one example, blind source separation is performed on the enhanced ultrasonic signal to obtain a separated weld defect signal, including: Performing standardization processing on the enhanced ultrasonic signal to obtain a standardized multi-channel signal matrix; Perform cross-correlation calculation on the standardized multi-channel signal matrix to obtain a correlation coefficient matrix, and input the correlation coefficient matrix into the separation optimization algorithm to perform separation weight calculation by maximizing negative entropy to obtain the initial separation matrix; Apply sparsity constraints to the initial separation matrix and use Newton iteration method to optimize and solve it to obtain the target separation matrix; Use the target separation matrix to perform linear transformation on the standardized multi-channel signal matrix to obtain multiple independent signal components; Feature analysis and feature reconstruction are performed on multiple independent signal components to obtain separated weld defect signals.

[0025] In this example, a multi-channel enhanced signal set is normalized to construct a standardized multi-channel signal matrix. In this matrix, each column represents the time series signal acquired by a channel, and each row corresponds to the synchronized response of the signals across all channels at a specific sampling time. Because channels vary in sensor sensitivity, coupling pressure, geometric position, and local reflectance characteristics, normalization is used to adjust each channel signal to a uniform amplitude space with zero mean and unit variance, eliminating amplitude offsets and scale differences between channels. Cross-correlation analysis is performed on the standardized multi-channel signal matrix. By calculating the time series cross-correlation coefficients between any two channels pair by pair, a symmetric correlation coefficient matrix with element values ranging from -1 to 1 is constructed. This matrix reflects the consistency of the responses across the time dimension between the channels. The main diagonal elements are always 1, while larger off-diagonal elements indicate stronger signal synchronization between the corresponding channel pairs. By analyzing the structural characteristics of the correlation coefficient matrix, a preliminary identification of channel sets with high levels of mixing within the signal, as well as potential common source components or interference overlap patterns, can be made. To separate mixed signals and extract independent source terms, the correlation coefficient matrix is input into the blind source separation optimization module. With the goal of maximizing signal independence, the signal mixing process is inversely solved using a negentropy maximization strategy. Negentropy is an important measure of signal non-Gaussianity; higher values indicate a signal deviates more from a Gaussian distribution and is more discernible. In the blind source separation process, negentropy is introduced as the optimization objective function. Using a separation optimization algorithm, the current signal structure is transformed and weighted to obtain an initial separation matrix. Each row of this matrix corresponds to an unmixing transformation vector, which is used to linearly project the original signal into a mutually independent coordinate space. A sparsity constraint mechanism is introduced based on the initial separation matrix. Sparsity requires that the separated signal components remain near zero at most time points, exhibiting significant peaks only at a few locations. By incorporating the sparsity constraint into the optimization objective function, the algorithm is guided to enhance its sensitivity to the pulse echo signal while maintaining independence. During the optimization process, the Newton iteration method is used as a solution tool. By repeatedly iteratively optimizing the current separation matrix, the negative entropy term and the sparsity term in the objective function converge to the optimal value at the same time, obtaining a set of target separation matrices with stable structure, good sparsity, and strong unmixing ability. The target separation matrix is applied to the standardized multi-channel signal matrix, and matrix linear transformation operations are performed. Coordinate axis rotation and scale reconstruction are performed in the signal space, thereby mapping the original mixed signal into a set of statistically independent signal components with significantly different response characteristics. Each independent component theoretically corresponds to a potential source signal component, which includes the real weld defect echo, structural reflection interference, environmental background noise, and system inherent noise. Since the transformed signal dimension is consistent with the original number of channels, these components are subjected to feature analysis and reconstruction screening to determine which components contain the physical response of the weld defect.During the feature analysis phase, each independent signal component is individually evaluated for its time-domain waveform structure, frequency-domain energy distribution, peak amplitude, pulse duration, relative position, and synchronization. Components with high amplitude, short duration, and abrupt structures concentrated within the expected echo window are identified as defect-related signals, while signals with long periods, low amplitude, or an average spectral distribution are considered noise interference. Based on this, the valid signal components are reconstructed into weld defect echo response curves based on the feature classification results. Useless or interfering components are removed, resulting in a highly pure, well-defined, and structurally relevant defect signal sequence.

[0026] In one example, a cross-correlation calculation is performed on a standardized multi-channel signal matrix to obtain a correlation coefficient matrix, and the correlation coefficient matrix is input into a separation optimization algorithm to perform separation weight calculation by maximizing negative entropy to obtain an initial separation matrix, including: Performing expectation operations on each channel signal pair in the standardized multi-channel signal matrix to obtain a correlation coefficient matrix; Perform eigenvalue decomposition on the correlation coefficient matrix to obtain the signal feature representation after dimensionality reduction; A signal probability distribution model is constructed based on the signal feature representation, and a Gaussian random variable with the same variance as the signal feature representation is generated as a reference model. The information entropy difference between the signal probability distribution model and the reference model is calculated to generate a negative entropy function. Add sparse constraint terms to the negative entropy function, construct a composite objective function, and perform Newton iterative optimization calculation based on the composite objective function to obtain the separation weight coefficient; Perform formal reorganization on the separation weight coefficients to obtain the original separation matrix, and check whether the iteration termination condition is satisfied that the matrix norm difference between two adjacent iterations is less than the preset target value; If it is less than , the original separation matrix is orthogonally normalized to obtain the initial separation matrix.

[0027] In this example, an expectation operation is performed on each channel signal pair in a standardized multi-channel signal matrix. By collaboratively analyzing sample point pairs across all channels, the statistical correlation is calculated. The expectation operation reflects the joint variation trends between channels in the temporal dimension, and the correlation coefficient matrix formed by summarizing the calculation results serves as a global characterization of the degree of coupling between the signals. This correlation coefficient matrix is a symmetric real number matrix, where the main diagonal represents the perfect correlation between each channel and itself, with a value of 1. The off-diagonal lines represent the correlation between two different channels, with values closer to 1 indicating greater signal mixing and values closer to 0 indicating greater independence. The correlation coefficient matrix is then subjected to eigenvalue decomposition, splitting it into a set of eigenvectors and corresponding eigenvalues. The magnitude of the eigenvalue reflects the dominance of the corresponding direction in the signal space, while the eigenvector represents the distribution structure of the main directions in the signal space. By retaining the principal components with larger eigenvalues and discarding the dimensions corresponding to smaller eigenvalues, a dimensionality reduction reconstruction of the high-dimensional channel space is effectively achieved, preserving the most representative information features of the original signal while reducing computational complexity. After the dimensionality reduction operation, a probability distribution model is constructed based on the resulting signal feature representation. This model characterizes the density distribution of data samples across different signal dimensions, thereby quantitatively expressing the non-Gaussianity, centralization, and skewness of the signal source. Furthermore, to standardize the distribution characteristics of the actual signal, a set of random variables with the same variance as the actual signal but a Gaussian distribution is constructed as a reference model. In theory, a signal with a Gaussian distribution has the highest information entropy. The more the probability distribution of the actual signal deviates from the Gaussian distribution, the higher its separability and recognition value. Based on this principle, a negative entropy function is constructed by calculating the difference in information entropy between the signal's true probability distribution model and the reference Gaussian model. A larger value of this function indicates a sparser and more non-Gaussian structure, resulting in a more effective separation. To enhance the recognizability of weld defect signals and suppress unstructured noise, a sparsity constraint term is introduced on the basis of the constructed negative entropy function. This sparsity term characterizes the discontinuity of the target signal in the temporal dimension. Specifically, the defect signal should remain low or near-zero amplitude for most of the time, exhibiting a high-amplitude, impulsive response only at specific moments. The sparsity constraint term is integrated with the negative entropy function to construct a composite objective function that simultaneously considers the statistical independence of the signal and the sparsity of the temporal structure, and has stronger physical significance and application adaptability. After the objective function is constructed, the Newton iterative optimization method is used to solve it. The Newton iterative method continuously calculates the derivative information of the objective function under the current separation weight parameters and uses the gradient and Hessian approximation to predict the optimal direction, thereby correcting the weight parameters in each iteration and gradually approaching the optimal solution. During the iteration process, the system continuously tracks whether the value of the composite objective function tends to be stable, while monitoring the change amplitude of the separation weight update to ensure that each round of iteration can bring about actual decoupling improvement.When the system detects that the matrix norm difference between the separation weight coefficients obtained in the current iteration and the previous iteration is less than the preset minimum change threshold, or when the objective function tends to be stable without obvious decline after several rounds of iterations, it is determined that the current separation weight has reached the global optimum or an acceptable local optimum. The separation weight coefficients are formally reorganized to convert their structure into a linear transformation structure that can be directly applied to the multi-channel signal matrix, and the original separation matrix is formed. In order to ensure the numerical stability and geometric reversibility of the matrix in the subsequent transformation process, it is orthogonalized and normalized so that each set of separation vectors is mutually orthogonal and has a unit norm. This operation improves the controllability of the separation matrix in the mathematical space and prevents signal distortion caused by scale expansion or inconsistent rotation during the linear superposition process, ensuring that the signal structure after unmixing is clear and the amplitude is stable. The initial separation matrix is obtained.

[0028] In one example, the separated weld defect signals are subjected to triangle constrained average interpolation and adaptive extended Kalman filter fusion processing to obtain filtered weld defect feature data, including: Extracting the peak amplitude, arrival time, waveform width and energy value from the separated weld defect signal, and forming the peak amplitude, arrival time, waveform width and energy value into first eigenvectors of multiple detection points; Performing Delaunay triangulation on the first eigenvectors of the multiple detection points to generate a spatial constraint network of the multiple detection points; Performing difference calculation on the first eigenvectors of the spatial constraint network of multiple detection points to obtain difference values, and comparing the difference values with the standard deviation of the normal area of the weld to obtain suspicious points among the multiple detection points; Perform triangle constrained average interpolation on the suspicious point and use the first eigenvectors of the detection points adjacent to the suspicious point that are not marked as suspicious points to perform weighted average calculation to obtain corrected feature data; The corrected characteristic data are input into the extended Kalman filter for state prediction and measurement update to obtain filtered weld defect characteristic data.

[0029] In this example, key parameters with physical significance are extracted from the separated weld defect signals, including peak amplitude, arrival time, waveform width, and overall energy. Peak amplitude represents the intensity of the defect echo, arrival time reflects the depth of the ultrasonic signal propagation path, waveform width reveals the thickness of the reflection interface or the effect of multiple superposition, and energy represents the integrated energy intensity of the signal in the time domain. These parameters together constitute the first eigenvector of each detection point in the spatial structure, and the set of eigenvectors corresponding to all detection points constitutes a structural mapping of the weld surface characteristic response. After mapping the three-dimensional spatial positions of all detection points and their corresponding first eigenvectors into geometric space, the Delaunay triangulation algorithm is used to perform spatial topological organization on these detection points. This partitions the point set into non-overlapping triangular regions, ensuring that no points other than the triangle vertices are contained within the triangles. This constructs a spatial constraint network that exhibits spatial connectivity, local optimality, and boundary stability. This network structure reflects the geometric adjacency between weld detection points and provides the neighborhood support required for subsequent interpolation and outlier identification, ensuring that the spatial feature evolution process has a clear regional scope and connectivity logic. After establishing the spatial constraint network, the difference between the eigenvectors of each detection point and its neighbors is calculated. The first eigenvector of the current point is subtracted from the eigenvectors of all connected neighboring points, and the magnitude of the sum of these differences in four-dimensional feature space is calculated. This difference represents the degree of inconsistency between the defect characteristics of the point and its neighbors. If the difference is significantly higher than the standard deviation of the neighboring points in the normal weld region, it indicates an abnormal trend or discontinuity at the point and is marked as a suspicious point. The standard deviation of the normal region is obtained by selecting a set of defect-free feature points within a continuous segment of the weld and calculating the difference distribution of their neighborhood characteristics. Three times the standard deviation is set as the anomaly threshold to ensure that only detection points with truly abrupt changes in characteristics are classified as abnormal. For detection locations identified as suspicious points, they are corrected using triangular constrained average interpolation. All neighboring points that form a triangular relationship with the suspicious point are found, and valid neighboring points that have not been marked as abnormal are selected from these points. The eigenvectors of these points are then used to construct an interpolation set. The weight coefficient is calculated based on the Euclidean distance between the suspicious point and each valid adjacent point. The feature vectors of these adjacent points are multiplied and weighted together using an inverse distance weighting method to obtain a smooth, continuous, and spatially consistent corrected feature vector. The feature vectors of all inspection points are then input into the Extended Kalman Filter module for dynamic correction. As a tool for processing nonlinear state transition systems, the Extended Kalman Filter effectively handles random perturbations and noise introduced into defect characteristics in space and time.In this framework, the eigenvector of each detection point is considered a state variable. The filter infers the current state based on the state prediction model of the previous moment and performs observation updates based on the corrected data measured in the current sampling period. This in turn adjusts the state estimate and updates the error covariance matrix. In the state prediction phase, the smooth evolution trend of the weld structure in space is used to construct a state transition model, so that the eigenvalues show a slow-changing trend in adjacent regions. In the observation update phase, the difference between the measured data and the predicted value, i.e., the amount of new information, is used to update the prediction deviation and adjust the observation weight based on the prediction credibility. In this process, the Kalman gain dependent variable is adjusted according to the dynamic changes in the error covariance, so that the system automatically increases the observation weight at abnormal points and takes the prediction as the main factor in stable areas, thereby achieving stable repair and dynamic optimization of the overall feature data. The state estimate output by the filter is used as the final weld defect feature result for each detection point.

[0030] In one example, triangle constrained average interpolation is performed on the suspicious point and weighted average calculation is performed using the first feature vectors of the detection points adjacent to the suspicious point that are not marked as suspicious points to obtain corrected feature data, including: Based on the triangle constraint network, all detection points adjacent to the suspicious point are identified, and detection points that are not marked as suspicious points are filtered out to obtain valid adjacent points; Calculate the spatial distance from the suspicious point to each valid adjacent point, and construct the inverse distance weight based on the spatial distance to obtain the spatial weight coefficient; Multiply the first eigenvectors of the valid adjacent points by the corresponding spatial weight coefficients to obtain weighted eigenvalues, and perform normalized summation on the weighted eigenvalues to obtain the corrected eigenvectors of the suspicious points. The Euclidean distance is calculated between the second eigenvector and the corrected eigenvector of the suspicious point to obtain the Euclidean distance, and the corrected eigenvector is used as the final correction value according to the Euclidean distance; The final correction value is used to replace the second eigenvector of the original suspicious point, and the data of the corresponding position in the detection point network is updated to obtain the corrected feature data.

[0031] In this example, the Delaunay triangulation results are used to form a spatial topological network to identify all directly adjacent test points connected to any spatial node marked as a suspicious point. These connections are naturally formed through the triangulated boundaries, ensuring complete spatial geometric logic and local continuity. Test points not marked as suspicious are screened out, and points with similar characteristic deviations or abnormal fluctuations are eliminated. This leaves only valid neighboring points that exhibit stable characteristics within the normal weld region, which serve as the reference set for subsequent interpolation and repair calculations. After obtaining valid neighboring points, the three-dimensional spatial distance is calculated based on the geometric coordinate difference between the suspicious point and each valid neighbor. This distance, an important measure of similarity and influence, determines the weighted contribution of each neighbor to the interpolation of the suspicious point's features. To accurately reflect the spatial attenuation characteristic, where closer distances increase influence, an inverse distance function is used as the weighting factor. That is, smaller neighbors have a greater influence on the interpolation results, while more distant points have a smaller influence. This prevents distant points from interfering with local repairs. All inverse distance weight coefficients are normalized to form a set of spatial weight coefficients. The first eigenvector of each valid adjacent point is multiplied by its corresponding spatial weight coefficient item by item to form a weighted eigenvalue. All components are accumulated and normalized and summed to form a corrected eigenvector that integrates spatial continuity and local signal trends. This corrected vector is the optimal approximate response prediction result of the current suspicious point under spatial topology, and it is also an intermediate state that is credible in a physical sense. In order to ensure that the correction result does not deviate seriously from the original data and to prevent the introduction of unreasonable new errors in the interpolation process, before applying the corrected eigenvector, it is calculated by Euclidean distance with the second eigenvector estimated in the previous iteration or the previous stage of the suspicious point to measure the straight-line distance between the two in the multidimensional feature space. This Euclidean distance is considered a measure of consistency between the correction vector and the historical features. When the distance is below the set threshold, it indicates that the current correction result is close to the original value and falls within the scope of reasonable compensation. Therefore, the correction vector is directly used as the final correction value for the update operation. When the Euclidean distance is too large, it indicates that factors such as neighborhood imbalance, feature mutation, or external interference have caused the correction result to deviate from the physical logic. In this case, the correction amplitude is limited, the reference point is re-evaluated, or contextual time constraints are introduced to improve the credibility of the correction. The final correction value is used to replace the second eigenvector of the original suspicious point, so that the feature state of the current point is restored to a spatially consistent state within the credible interval, and the data update is written to the corresponding position of the entire detection point network. As all suspicious points are corrected one by one and the network is updated synchronously, the original feature data set with fluctuations, missing or discontinuities is transformed into a high-quality feature data network that is structurally coherent, numerically stable, and physically reasonable, thus obtaining the corrected feature data.

[0032] In one example, regularized Gaussian field weld contour reconstruction and defect location processing are performed based on filtered weld defect feature data to obtain a weld contour model and defect location data, including: Select a standard weld model that matches the weld type to be inspected as the initial model, and obtain the benchmark data of the weld profile based on the initial model; Extracting three-dimensional coordinate points and confidence values from the filtered weld defect feature data to obtain a contour feature point set; The contour feature point set is converted into a Gaussian probability distribution form, and multiple Gaussian functions are used to describe the distribution of weld defects to obtain a defect probability distribution map; A comprehensive evaluation energy function including fitting error, transformation continuity and prior constraints is constructed based on the initial model and defect probability distribution diagram; The L-BFGS algorithm is used to minimize the comprehensive evaluation energy function to obtain the optimal transformation parameters. Based on the optimal transformation parameters, the area where the defect confidence exceeds the confidence threshold is identified to obtain the weld contour model and defect location data.

[0033] In this example, an appropriate standard weld model is selected as the initial model based on the specific weld type to be inspected. The standard model is chosen based on the weld geometry (e.g., fillet weld, butt weld, or T-weld), welding method, and material. Selecting a standard model that matches the weld type to be inspected provides an accurate geometric reference for subsequent analysis and helps the system infer and reconstruct the actual weld contour based on the model in subsequent steps. Using the standard weld model as the initial reference model, baseline data for the weld contour is obtained based on the model's geometric parameters. This baseline data includes the weld's shape, dimensions, welding path, and weld continuity, forming an idealized representation of the weld type. Key information is extracted from the filtered weld defect feature data, including 3D coordinate points and their associated confidence values. The 3D coordinate points reflect the spatial location of the defect at each location within the weld, while the confidence value indicates the reliability and strength of the defect detected at that location. These 3D coordinate points and their corresponding confidence values are combined to form a set of contour feature points, representing the spatial location and intensity distribution of the weld defect. These feature point sets encompass all potential defect regions within the weld. The contour feature point set is converted into a Gaussian probability distribution, mapping the actual detected point set from discrete space to a continuous probability space to more accurately describe the distribution of defects within the weld. Because weld defects manifest as local discontinuities or sudden changes, multiple Gaussian functions are used to model the defect distribution. Each Gaussian function corresponds to a defect region. Combining these Gaussian functions generates a defect probability distribution map that describes the intensity and spatial distribution of defects, effectively capturing the spatial clustering of defects within the weld. Based on the initial model and the defect probability distribution map, a comprehensive evaluation energy function is constructed, including fitting error, transformation continuity, and prior constraints. The fitting error component measures the discrepancy between the initial model and the actual defect distribution, ensuring that the selected standard model is adaptable to variations in actual inspection data. The transformation continuity term ensures that the weld contour remains smooth and continuous after spatial transformation. Prior constraints leverage known physical properties of the weld, such as weld thickness, surface quality, and material properties, to constrain the model optimization process and avoid overfitting and unreasonable deformation. The L-BFGS algorithm is used to minimize the comprehensive evaluation energy function to obtain the optimal transformation parameters. The L-BFGS algorithm is an efficient quasi-Newton method for solving large-scale optimization problems. The algorithm iteratively updates the optimization parameters, each time using the gradient information and historical data from the previous calculation to approximate the optimal solution. By continuously adjusting the transformation parameters, the L-BFGS algorithm minimizes the error between the generated weld contour model and the actual inspection data while maintaining the continuity and physical consistency of the model. After several iterations, a set of optimal transformation parameters is obtained, which represents the best match between the standard weld model and the actual weld defect data.Based on the optimal transformation parameters, the defect confidence level is analyzed and all areas with defect confidence levels exceeding a preset threshold are marked as defective. By analyzing the defect signal strength in the defect probability distribution map, the defect confidence levels in these areas are determined to be sufficiently high, confirming these areas as actual defect locations. The output includes a weld contour model and defect location data.

[0034] In one example, the L-BFGS algorithm is used to minimize the comprehensive evaluation energy function to obtain the optimal transformation parameters. Based on the optimal transformation parameters, the area where the defect confidence exceeds the confidence threshold is identified to obtain the weld contour model and defect location data, including: The initial temperature parameters and cooling rate of the comprehensive evaluation energy function are set using the L-BFGS algorithm, and the initial temperature parameters and cooling rate are set as identity transformation to obtain the initial state set; Calculating the first-order derivative and the second-order derivative of the comprehensive evaluation energy function through the initial state set, and performing transformation parameter replacement according to the first-order derivative and the second-order derivative to obtain the transformation parameters of the current iteration; Based on the transformation parameters of the current iteration, the iterative calculation is repeated until the change in the energy function value between two adjacent iterations is less than the energy target value, and the optimal transformation parameters are obtained; The defect confidence is calculated based on the optimal transformation parameters, and the area where the defect confidence exceeds the confidence threshold is marked as the defect area, and the weld contour model and defect location data are obtained.

[0035] In this example, the L-BFGS optimization environment is initialized, and a reasonable set of initial parameters is set for the comprehensive evaluation energy function, including the initial temperature parameter and cooling rate. These two parameters are derived from the annealing concept, but are reinterpreted in the L-BFGS optimization framework as metavariables for adjusting the search step size and controlling the gradient update scale. To ensure that the initial state does not impose any external bias on the model, the initial temperature parameter and cooling rate are both set to the form of identity transformation, that is, the variables are initialized to the unit matrix or zero offset vector, which is geometrically equivalent to keeping the current original structural form of the model unchanged, thus forming the initial state set. This initial state set is both a set of control points and a transformation reference frame without any prior deformation, providing a stable starting point for the subsequent energy optimization process. Using this initial state set as input, a derivative calculation operation is performed on the comprehensive evaluation energy function. The first-order derivative of the energy function with respect to the current transformation parameter, known as the gradient, is calculated, reflecting the direction and rate of change of the energy function at the current parameter. Its approximate second-order derivative, known as the Hessian matrix or its low-rank approximation, is calculated to obtain information about the local curvature of the energy function, thereby assessing the steepest direction and optimal step size of the function's descent path under the current parameter state. The advantage of the L-BFGS algorithm lies in its lack of explicit construction of a complete second-order derivative matrix. Instead, it constructs the update direction by differencing historical gradients and state variables, effectively reducing memory consumption and improving iteration efficiency. Using this derivative information, the current set of model control points is transformed and replaced with a nonlinear mapping in feature space. This essentially adjusts the weld contour model's 3D shape to better align with the actual defect feature point set. After each parameter replacement, the value of the comprehensive evaluation energy function is recalculated based on the updated model state and compared with the energy value recorded in the previous iteration to determine convergence progress. If the energy difference between two iterations exceeds the preset energy target threshold, the current transformation has not yet reached its optimal state, and the system will proceed to the next round of derivative calculation and parameter update based on the current parameters. If the energy difference falls below the threshold, the model transformation has approached the minimum of the energy function, and the system terminates the iteration, deeming the current parameter state as the optimal transformation. The entire optimization process uses the monotonic decrease of the energy function as the convergence criterion. By guiding the model transformation path to continuously approximate the distribution of the actual defect point set, it gradually eliminates fitting errors, optimizes transformation continuity, and satisfies spatial constraints. After obtaining the optimal transformation parameters, they are applied to the defect probability distribution map to reconstruct the transformed weld contour model. Based on this model, the defect confidence level at each spatial location is recalculated. This confidence level is a statistically significant defect probability calculated by combining the optimal transformation mapping with the Gaussian field distribution. Each point within the weld region is assigned a confidence level in the three-dimensional space after the model mapping, and these values constitute the three-dimensional probability field of the weld defect.The system then sets a confidence threshold as the classification boundary, traverses the entire contour model, and marks all areas with confidence exceeding this threshold. The marking process combines spatial connectivity analysis with a gradient assessment mechanism to ensure that each area marked as a defect not only meets the conditions in terms of statistical value, but also constitutes a coherent and stable defect area in terms of spatial topology, avoiding the mislabeling of isolated noise or occasional anomalies. Through this process, the final complete output result of the weld contour model and defect location data is formed. The contour model contains the geometric shape of the standard model under the optimal transformation, and the defect location data records the spatial coordinates, defect volume, principal axis direction, aspect ratio, and confidence weight of all high-confidence areas.

[0036] Reference Figure 2 This embodiment provides a steel structure weld ultrasonic nondestructive testing system, including: The signal acquisition module 1 is used to collect signals from the weld surface of the steel structure using an adaptive coupled ultrasonic probe to obtain raw ultrasonic signal data; Transformation processing module 2, used for performing multi-spectral adaptive wavelet transformation on the original ultrasonic signal data to obtain an enhanced ultrasonic signal; Blind source separation module 3, used to perform blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal; Interpolation fusion module 4 is used to perform triangle constrained average interpolation and adaptive extended Kalman filter fusion processing on the separated weld defect signals to obtain filtered weld defect feature data; The reconstruction and positioning module 5 is used to perform regularized Gaussian field weld contour reconstruction and defect positioning processing based on the filtered weld defect feature data to obtain a weld contour model and defect position data.

[0037] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0038] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, system, article, or method comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, system, article, or method. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, system, article, or method comprising the element.

[0039] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for ultrasonic nondestructive testing of steel structure welds, characterized in that: The following steps are involved: Adopting adaptive coupled ultrasonic probe to collect signals from the weld surface of steel structure and obtain original ultrasonic signal data; Performing multi-spectral adaptive wavelet transform processing on the original ultrasonic signal data to obtain an enhanced ultrasonic signal; performing blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal; Performing triangle constrained average interpolation and adaptive extended Kalman filter fusion processing on the separated weld defect signal to obtain filtered weld defect feature data; Regularized Gaussian field weld contour reconstruction and defect location processing are performed based on the filtered weld defect feature data to obtain a weld contour model and defect location data.

2. The ultrasonic nondestructive testing method for steel structure welds according to claim 1, characterized in that: The adaptive coupled ultrasonic probe is used to collect signals from the weld surface of the steel structure to obtain raw ultrasonic signal data, including: Placing the adaptive coupling ultrasonic probe on the surface of the steel structure weld, and transmitting pulse excitation through the adaptive coupling ultrasonic probe to obtain an ultrasonic signal; Controlling the ultrasonic signal to propagate into the steel structure in a direction perpendicular to the weld surface, and generating a reflected wave when encountering an internal interface or defect in the weld, and receiving an ultrasonic echo signal reflected from the inside of the weld; The ultrasonic echo signal is input into a preamplifier for gain amplification to obtain an amplified ultrasonic signal, and the amplified ultrasonic signal is converted into digital form to obtain original ultrasonic signal data.

3. The ultrasonic nondestructive testing method for steel structure welds according to claim 1, characterized in that: The performing multi-spectral adaptive wavelet transform processing on the original ultrasonic signal data to obtain an enhanced ultrasonic signal includes: Performing wavelet decomposition on the original ultrasonic signal data to obtain wavelet decomposition coefficients of different frequency scales; The spectrum energy analysis is performed on the wavelet decomposition coefficients of different frequency scales, and a quadratic polynomial fitting algorithm is used to establish a spectrum energy and noise correlation model to obtain the standard deviation of noise in each layer. Based on the noise standard deviation and coefficient length of each layer, the adaptive threshold corresponding to each decomposition layer is calculated, and nonlinear compensation is performed on the wavelet decomposition coefficient according to the adaptive threshold to obtain the denoised wavelet coefficient; Performing inverse wavelet transform on the denoised wavelet coefficients to reconstruct the denoised ultrasonic signal to obtain a preliminarily denoised ultrasonic signal; The ultrasonic signal after the preliminary noise reduction is subjected to peak feature preservation and pseudo-Gibbs phenomenon suppression processing to obtain an enhanced ultrasonic signal.

4. The ultrasonic nondestructive testing method for steel structure welds according to claim 1, characterized in that: The step of performing blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal includes: performing standardization processing on the enhanced ultrasonic signal to obtain a standardized multi-channel signal matrix; Performing cross-correlation calculation on the standardized multi-channel signal matrix to obtain a correlation coefficient matrix, and inputting the correlation coefficient matrix into a separation optimization algorithm to perform separation weight calculation for maximizing negative entropy to obtain an initial separation matrix; Applying sparsity constraints to the initial separation matrix and optimizing it using Newton iteration method to obtain a target separation matrix; Performing a linear transformation on the standardized multi-channel signal matrix using the target separation matrix to obtain a plurality of mutually independent signal components; Feature analysis and feature reconstruction are performed on the multiple independent signal components to obtain separated weld defect signals.

5. The ultrasonic nondestructive testing method for steel structure welds according to claim 4, characterized in that: The cross-correlation calculation is performed on the standardized multi-channel signal matrix to obtain a correlation coefficient matrix, and the correlation coefficient matrix is input into the separation optimization algorithm to perform separation weight calculation for maximizing negative entropy to obtain an initial separation matrix, including: performing an expectation operation on each channel signal pair in the standardized multi-channel signal matrix to obtain a correlation coefficient matrix; Performing eigenvalue decomposition on the correlation coefficient matrix to obtain a signal feature representation after dimensionality reduction; Constructing a signal probability distribution model based on the signal feature representation, generating a Gaussian random variable with the same variance as the signal feature representation as a reference model, calculating the information entropy difference between the signal probability distribution model and the reference model, and generating a negative entropy function; Adding a sparse constraint term to the negative entropy function to construct a composite objective function, and performing Newton iterative optimization calculation based on the composite objective function to obtain a separation weight coefficient; Performing formal reorganization on the separation weight coefficients to obtain an original separation matrix, and checking whether an iteration termination condition is satisfied that the matrix norm difference between two adjacent iterations is less than a preset target value; If it is less than , the original separation matrix is orthogonally normalized to obtain an initial separation matrix.

6. The ultrasonic nondestructive testing method for steel structure welds according to claim 1, characterized in that: The separated weld defect signal is subjected to triangle constrained average interpolation and adaptive extended Kalman filter fusion processing to obtain filtered weld defect feature data, including: Extracting peak amplitude, arrival time, waveform width, and energy value from the separated weld defect signal, and combining the peak amplitude, arrival time, waveform width, and energy value into first feature vectors of multiple detection points; Performing Delaunay triangulation on the first eigenvectors of multiple detection points to generate a spatial constraint network of the multiple detection points; performing difference calculation on first eigenvectors of the spatial constraint network of the plurality of detection points to obtain a difference value, and comparing the difference value with a standard deviation of a normal area of the weld to obtain a suspicious point among the plurality of detection points; Performing triangle constrained average interpolation on the suspicious point and performing weighted average calculation using the first eigenvectors of detection points adjacent to the suspicious point that are not marked as suspicious points to obtain corrected feature data; The corrected characteristic data is input into an extended Kalman filter for state prediction and measurement update to obtain filtered weld defect characteristic data.

7. The ultrasonic nondestructive testing method for steel structure welds according to claim 6, characterized in that: The performing triangle constrained average interpolation on the suspicious point and performing weighted average calculation using first feature vectors of detection points adjacent to the suspicious point that are not marked as suspicious points to obtain corrected feature data includes: Based on the triangle constraint network, all detection points adjacent to the suspicious point are identified, and detection points that are not marked as suspicious points are screened out to obtain valid adjacent points; Calculating the spatial distance from the suspicious point to each valid adjacent point, and constructing an inverse distance weight based on the spatial distance to obtain a spatial weight coefficient; Multiplying the first eigenvectors of the valid adjacent points by the corresponding spatial weight coefficients to obtain weighted eigenvalues, and performing normalized summation on the weighted eigenvalues to obtain a corrected eigenvector of the suspicious point; Performing Euclidean distance calculation on the second eigenvector of the suspicious point and the corrected eigenvector to obtain a Euclidean distance, and using the corrected eigenvector as a final correction value based on the Euclidean distance; The final correction value is used to replace the second eigenvector of the original suspicious point, and the data of the corresponding position in the detection point network is updated to obtain the corrected feature data.

8. The ultrasonic nondestructive testing method for steel structure welds according to claim 1, characterized in that: The regularized Gaussian field weld profile reconstruction and defect location processing are performed based on the filtered weld defect feature data to obtain a weld profile model and defect location data, including: Select a standard weld model that matches the weld type to be inspected as the initial model, and obtain the benchmark data of the weld profile based on the initial model; Extracting three-dimensional coordinate points and confidence values from the filtered weld defect feature data to obtain a contour feature point set; Converting the contour feature point set into a Gaussian probability distribution form, and using multiple Gaussian functions to describe the distribution of weld defects to obtain a defect probability distribution map; Constructing a comprehensive evaluation energy function including fitting error, transformation continuity and prior constraints according to the initial model and the defect probability distribution graph; The L-BFGS algorithm is used to minimize the comprehensive evaluation energy function to obtain the optimal transformation parameters, and the area where the defect confidence exceeds the confidence threshold is identified based on the optimal transformation parameters to obtain the weld contour model and defect location data.

9. The ultrasonic nondestructive testing method for steel structure welds according to claim 8, characterized in that: The L-BFGS algorithm is used to minimize the comprehensive evaluation energy function to obtain the optimal transformation parameters, and the area where the defect confidence exceeds the confidence threshold is identified according to the optimal transformation parameters to obtain the weld contour model and defect location data, including: Using the L-BFGS algorithm to set the initial temperature parameter and cooling rate of the comprehensive evaluation energy function, and setting the initial temperature parameter and cooling rate as identity transformation to obtain an initial state set; Calculating the first-order derivative and the second-order derivative of the comprehensive evaluation energy function through the initial state set, and performing transformation parameter replacement according to the first-order derivative and the second-order derivative to obtain the transformation parameters of the current iteration; Based on the transformation parameters of the current iteration, iterative calculation is repeatedly performed until the change in the energy function value between two adjacent iterations is less than the energy target value, thereby obtaining the optimal transformation parameters; The defect confidence is calculated based on the optimal transformation parameters, and the area where the defect confidence exceeds the confidence threshold is marked as a defect area, thereby obtaining a weld contour model and defect position data.

10. An ultrasonic nondestructive testing system for steel structure welds, characterized in that: The method for ultrasonic nondestructive testing of steel structure welds according to any one of claims 1 to 9 is implemented, wherein the ultrasonic nondestructive testing system for steel structure welds comprises: The signal acquisition module is used to collect signals from the weld surface of the steel structure using an adaptive coupled ultrasonic probe to obtain raw ultrasonic signal data; A transformation processing module, configured to perform multi-spectral adaptive wavelet transformation on the original ultrasonic signal data to obtain an enhanced ultrasonic signal; A blind source separation module is used to perform blind source separation on the enhanced ultrasonic signal to obtain a separated weld defect signal; An interpolation and fusion module is used to perform triangle constrained average interpolation and adaptive extended Kalman filter fusion processing on the separated weld defect signal to obtain filtered weld defect feature data; The reconstruction and positioning module is used to perform regularized Gaussian field weld contour reconstruction and defect positioning processing based on the filtered weld defect feature data to obtain a weld contour model and defect position data.

Images