An ultrasonic flaw detection method for automatic identification and evaluation of weld defects

Abstract

This invention discloses an ultrasonic flaw detection method for automatic identification and evaluation of weld defects, belonging to the field of non-destructive testing technology. The method includes: planning an adaptive scanning path for a phased array ultrasonic probe based on a 3D CAD model of the workpiece to be inspected; acquiring raw ultrasonic signal data through a full-matrix acquisition mode; performing full-focus imaging processing on the raw ultrasonic signal data to generate a high-resolution ultrasonic image, and simultaneously extracting geometric modal information, dynamic modal information, and texture modal information corresponding to the spatial position of the high-resolution ultrasonic image; inputting the high-resolution ultrasonic image into a pre-trained deep neural network for pixel-level semantic segmentation, outputting a defect mask image, and calculating the basic geometric parameters of the defect based on the defect mask image; using the defect contour in the defect mask image as a reference, fusing the corresponding geometric modal information, dynamic modal information, and texture modal information to form a comprehensive defect fingerprint vector.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and in particular to an ultrasonic testing method for automatic identification and evaluation of weld defects. Background Technology

[0002] Non-destructive testing (NDT) technology refers to a class of testing methods that, without damaging or altering the physicochemical properties and performance of the object being tested, utilize physical principles such as sound, light, electricity, magnetism, heat, or radiation to detect, identify, and evaluate defects, damage, organizational state, or performance parameters on the interior and surface of materials, components, or equipment. Its core objective is to achieve early detection and quantitative analysis of potential faults while ensuring structural integrity and operational safety. It is widely used in industrial fields with extremely high safety requirements, such as aerospace, energy and power, rail transportation, pressure vessels, and petrochemicals.

[0003] Traditional ultrasonic flaw detection relies heavily on the experience of the inspectors. The identification, classification and hazard assessment of defects are easily affected by subjective factors, resulting in poor consistency of results and a high rate of missed detections and misjudgments. Moreover, traditional methods mostly use only the grayscale of B-scan images or single A-scan features, failing to systematically integrate the spatial morphology, dynamic response and texture characteristics of defects, which limits the ability to distinguish between small, hidden or similar appearance defects. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an ultrasonic flaw detection method for automatic identification and assessment of weld defects. This solves the problem that traditional ultrasonic flaw detection is highly dependent on the experience of the inspectors. The identification, classification and hazard assessment of defects are easily affected by subjective factors, resulting in poor consistency of results and high rates of missed detection and misjudgment. In addition, traditional methods mostly use only the grayscale of B-scan images or a single A-scan feature, failing to systematically integrate the spatial morphology, dynamic response and texture characteristics of defects, which limits the ability to distinguish between small, hidden or similar appearance defects.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an ultrasonic testing method for automatic identification and evaluation of weld defects, comprising: Based on the 3D CAD model of the workpiece to be inspected, an adaptive scanning path for the phased array ultrasonic probe is planned, and the original ultrasonic signal data is acquired through the full matrix acquisition mode. The original ultrasound signal data is processed by full-focus imaging to generate a high-resolution ultrasound image, and the geometric modal information, dynamic modal information and texture modal information corresponding to the spatial position of the high-resolution ultrasound image are extracted simultaneously. High-resolution ultrasound images are input into a pre-trained deep neural network for pixel-level semantic segmentation, outputting a defect mask image, and the basic geometric parameters of the defect are calculated based on the defect mask image. Based on the defect contour in the defect mask image, the geometric modal information, dynamic modal information and texture modal information at the corresponding position are fused to form a comprehensive defect fingerprint vector. Based on the comprehensive defect fingerprint vector, the defect type is identified and the defect hazard level and maintenance suggestions are output. By fusing a 3D CAD model with a defect geometry reconstructed from a defect mask image and basic geometric parameters, a digital twin containing real internal defects is constructed. Online mechanical simulation is performed on the digital twin based on preset load conditions to calculate the stress intensity factor at the defect tip, the remaining strength of the workpiece, and the fatigue life, generating a structural health assessment report.

[0007] As a preferred embodiment of the ultrasonic flaw detection method for automatic identification and evaluation of weld defects described in this invention, the adaptive scanning path of the phased array ultrasonic probe based on the three-dimensional CAD model of the workpiece to be inspected specifically includes the following steps: Extract the weld centerline from the 3D CAD model and divide the control nodes along the weld length direction with a fixed step size; At each control node, the corresponding base material thickness and bevel geometry parameters are read, and the local incident surface normal vector is calculated. Based on the longitudinal wave velocity in the weld material Longitudinal wave velocity in the probe wedge material Combined with the bevel half angle Solve for the probe deflection angle required to make the ultrasonic beam perpendicularly incident on the weld fusion line. Among them, the deflection angle The following relationship must be satisfied: ; In the formula, This indicates the longitudinal wave propagation velocity in the weld metal material. This indicates the longitudinal wave propagation velocity in the wedge material used in the ultrasonic probe. This represents the angle of a single-sided bevel, which is half of the total bevel angle. This indicates the mechanical deflection angle that the probe needs to be set at this control node; The probe spatial coordinates and deflection angles corresponding to all control nodes are used to generate a continuous and smooth scanning trajectory through cubic spline interpolation; In areas where the radius of curvature of the weld is less than a preset threshold, the density of control nodes is automatically increased to ensure that the sound beam coverage has no blind spots.

[0008] As a preferred embodiment of the ultrasonic flaw detection method for automatic identification and evaluation of weld defects described in this invention, the specific steps for simultaneously extracting geometric modal information, dynamic modal information, and texture modal information are as follows: For high-resolution ultrasound images generated after full-focus imaging processing, the grayscale intensity value of each pixel is directly recorded as geometric modal information. For the defect areas initially identified in the high-resolution ultrasound images, all A-scan signals within the corresponding spatiotemporal range in the original ultrasound signal dataset are traced back. Perform a Hilbert transform on each A-scan signal to obtain an analytical signal, and then take its modulus to obtain the envelope signal; Calculate the time it takes for each envelope signal to rise from 10% peak value to 90% peak value, and record it as the rise time. Calculate the time it takes for each envelope signal to drop from its peak value to 36.8% of its peak value, and denote it as the decay time constant; The dynamic modal information is formed by combining the standard deviation of all rise times with the mean of all decay time constants. A fixed-size window is extracted from the high-resolution ultrasound image with the defect region as the center, and a two-dimensional discrete wavelet transform is performed on the window image. Select the horizontal high-frequency sub-band, vertical high-frequency sub-band, and diagonal high-frequency sub-band obtained from the first layer decomposition; Calculate the energy of each of the three high-frequency subbands. The energy is defined as the average of the sum of squares of all coefficients within the subband. The three energy values ​​and their corresponding gray-level co-occurrence matrix contrast and entropy values ​​are combined to form the texture modal information.

[0009] As a preferred embodiment of the ultrasonic flaw detection method for automatic identification and evaluation of weld defects described in this invention, the step of inputting high-resolution ultrasonic images into a pre-trained deep neural network for pixel-level semantic segmentation specifically includes: Construct the U-Net++ network structure, which includes a 4-level encoder path and a 4-level nested decoder path; Each encoder consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, followed by a 2×2 max pooling layer. Each level of decoder restores spatial resolution through upsampling and performs channel splicing with the outputs of multiple encoders from the same level and shallower levels. The input image is uniformly scaled to 512×512 pixels, and the feature map size is 32×32 after 4 levels of downsampling. The loss function uses a weighted combination of Dice loss and focus loss, where Dice loss is used to improve the segmentation recall of small defects, and focus loss is used to alleviate the gradient sparsity problem caused by background pixels dominating. The network output is a 4-channel probability map, which corresponds to four typical weld defects: cracks, lack of fusion, porosity, and slag inclusions. Apply the argmax operation to the output probability map to generate a single-channel defect mask image; For each connected component in the defect mask image, calculate its area, equivalent diameter, principal axis direction, and perimeter, which serve as the basic geometric parameters of the defect.

[0010] As a preferred embodiment of the ultrasonic flaw detection method for automatic identification and evaluation of weld defects described in this invention, the specific steps of fusing multimodal information to form a comprehensive defect fingerprint vector based on the defect contour in the defect mask image are as follows: Each independent connected component in the defect mask image is used as a processing unit; Calculate the number of pixels covered by the connected component, multiply it by the actual physical area corresponding to a single pixel, and obtain the defect projection area; Derive the equivalent circle diameter from the projected area; The principal axis orientation angle of the defect is calculated using principal component analysis. Extract all A-scan envelope signals corresponding to the connected component in the original ultrasound signal data, calculate the standard deviation of the envelope peak occurrence time, and reflect the spatial dispersion of the internal reflection interface of the defect. Calculate the energy concentration of all envelope signals. The energy concentration is defined as the proportion of energy in the first 30% of the time window to the total energy. Gabor filtering was performed on the local region containing the connected component in the high-resolution ultrasound image, using 4 directions and 3 scales, with the 4 directions being 0°, 45°, 90° and 135°; For each filtered response map, a local binary pattern histogram is calculated, and its statistical features are concatenated into a texture descriptor. Geometric features, dynamic features, and texture descriptors are concatenated in a fixed order to form a high-dimensional comprehensive defect fingerprint vector. The comprehensive defect fingerprint vector is input into a multilayer perceptron classifier, which contains two hidden layers with 128 and 64 neurons in each layer, respectively. The output layer uses the Softmax activation function to output the probability distribution of planar or volumetric defects and maps it to a preset hazard level and maintenance recommendations.

[0011] As a preferred embodiment of the ultrasonic flaw detection method for automatic identification and evaluation of weld defects described in this invention, the defect geometry reconstructed from the defect mask image and basic geometric parameters comprises the following steps: Map each foreground pixel in the defect mask image to a three-dimensional coordinate system; The depth coordinates are determined by the ultrasonic flight time and the material velocity corresponding to the pixel point, that is, the depth is equal to the velocity of sound multiplied by the flight time and then divided by 2; Normal vectors are estimated for all three-dimensional points within the same defective connected domain, and the surface normal vector of each point is calculated using the local plane fitting method. Construct a signed distance field, where the field function value is zero at the defect surface, negative inside, and positive outside; The Poisson surface reconstruction algorithm is invoked to solve for the implicit surface function that satisfies the gradient field approximation of the point cloud normal. Generate a closed triangular mesh model to ensure that the model has a complete topology and no self-intersections; The triangular mesh model is combined with the original 3D CAD model of the workpiece using a Boolean union operation to form a complete structural digital twin containing real internal defects. The digital twin undergoes geometric cleanup, including removing micro-fragments, repairing non-manifold edges, and optimizing surface quality to ensure its suitability for subsequent finite element analysis.

[0012] As a preferred embodiment of the ultrasonic flaw detection method for automatic identification and evaluation of weld defects described in this invention, the specific steps of performing online mechanical simulation on the digital twin based on preset load conditions are as follows: Apply a preset mechanical load, thermal load, or fatigue load spectrum to the digital twin; The material properties are determined using a bilinear elastoplastic constitutive model, with inputs including yield strength, elastic modulus, hardening modulus, and Poisson's ratio. The overall model is divided into tetrahedral finite element meshes, and local mesh refinement is implemented in the defect tip region. The element size in the refined region is no greater than 1 / 10 of the minimum radius of curvature of the defect tip. Transient analysis can be performed using an explicit dynamics solver, or static / quasi-static analysis can be performed using an implicit solver; Stress field data is extracted from the defect leading edge, and the stress intensity factor is calculated along a preset integration path. ,in Calculated using the following formula: ; In the formula, This represents the radial distance measured from the tip of the defect. This represents the normal stress components along the defect propagation direction in polar coordinates. Indicates Type I stress intensity factor; The calculated result Compare with the material fracture toughness KIC to determine whether instability propagation has occurred; If used for fatigue assessment, the damage increment is calculated for each load cycle by combining the material's SN curve and Miner's linear cumulative damage theory. The number of cycles corresponding to when the total damage value reaches 1 is the predicted remaining fatigue life.

[0013] As a preferred embodiment of the ultrasonic flaw detection method for automatic identification and assessment of weld defects described in this invention, the specific steps for generating the structural health assessment report are as follows: Summarize the defect location, type identification results, hazard level, basic geometric parameters, and stress intensity factor. Remaining strength safety margin and predicted remaining fatigue life; The residual strength safety margin is defined as the ratio of the allowable stress to the maximum working stress obtained from the simulation. When the remaining strength safety margin is less than the preset safety factor, or the predicted remaining fatigue life is less than the next planned maintenance cycle, the defect is marked as high risk. In the 3D visualization interface, the defect geometry is displayed on the workpiece model as a semi-transparent red overlay, and the defect number and key parameters are labeled. Generate structured text descriptions, including defect cause analysis, mechanical impact assessment, repair method suggestions, and recommended time for the next inspection; Output a PDF document conforming to ASME or ISO standard format, including a cover, defect list, simulation result screenshots, safety assessment conclusions, and an electronic signature area; The system automatically archives the report to the equipment lifecycle management database and triggers a notification to the maintenance scheduling platform.

[0014] The beneficial effects of this invention are as follows: By integrating adaptive ultrasonic scanning driven by a 3D CAD model, full-focus high-resolution imaging, simultaneous extraction of multimodal features, pixel-level defect segmentation based on deep learning, construction of defect fingerprints based on physical consistency, reconstruction of real defect geometry, and digital twin analysis fused with mechanical simulation into a closed-loop process, the positioning accuracy, identification accuracy, and assessment reliability of weld defect detection are improved. It not only achieves comprehensive quantitative judgment from the presence or absence of defects to defect type, morphology, hazard level, and remaining life, but also avoids the subjectivity and experience dependence of traditional manual interpretation. At the same time, it directly outputs structural safety margin and maintenance decision-making basis through online mechanical simulation, improving the intelligence, automation, and engineering practicality of in-service welded structure health monitoring. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of an ultrasonic testing method for automatic identification and assessment of weld defects in an embodiment. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] Reference Figure 1 This is one embodiment of the present invention, which provides an ultrasonic testing method for automatic identification and evaluation of weld defects, comprising the following steps: S1. Based on the three-dimensional CAD model of the workpiece to be inspected, plan the adaptive scanning path of the phased array ultrasonic probe and acquire the original ultrasonic signal data through the full matrix acquisition mode.

[0020] Furthermore, the weld centerline is extracted from the 3D CAD model, and control nodes are divided along the weld length with fixed step sizes. At each control node, the base metal thickness and groove geometry parameters corresponding to that location are read, and the local incident surface normal vector is calculated. The longitudinal wave velocity in the weld material is then considered. Longitudinal wave velocity in the probe wedge material Combined with the bevel half angle Solve for the probe deflection angle required to make the ultrasonic beam perpendicularly incident on the weld fusion line. Among them, the deflection angle The following relationship must be satisfied: ; In the formula, This indicates the longitudinal wave propagation velocity in the weld metal material. This indicates the longitudinal wave propagation velocity in the wedge material used in the ultrasonic probe. This represents the angle of a single-sided bevel, which is half of the total bevel angle. This indicates the mechanical deflection angle that the probe needs to be set at the control node; a continuous and smooth scanning trajectory is generated by cubic spline interpolation of the probe spatial coordinates and deflection angles corresponding to all control nodes; in areas where the radius of curvature of the weld is less than a preset threshold, the density of control nodes is automatically increased to ensure that the sound beam coverage has no blind spots.

[0021] It should be noted that by combining the three-dimensional CAD model to accurately extract the geometric features of the weld and dynamically adjusting the probe deflection angle and spatial trajectory, it is possible to ensure that the ultrasonic beam always covers the key areas of the weld with optimal incident conditions, effectively improving the detection sensitivity of high-risk areas such as fusion lines, effectively avoiding missed detections caused by beam deviation or obstruction, and adaptively densifying the control nodes in areas with complex curvature, ensuring data acquisition quality with no blind spots and high consistency throughout the weld.

[0022] In particular, the core of this scanning path planning method lies in deeply coupling the prior geometric information of the workpiece with the physical laws of ultrasonic wave propagation. By predicting the interaction between the sound beam and the weld structure in the digital model stage, the probe movement is no longer a simple movement along the surface, but an intelligent behavior with material-structure-acoustic collaborative optimization. This feedforward control strategy based on physical mechanisms fundamentally solves the problem of signal attenuation or blind spots caused by incident angle mismatch in traditional fixed-angle scanning in complex bevels or welds with varying thicknesses, laying an irreplaceable foundation for high-quality raw data acquisition.

[0023] S2. Perform full-focus imaging processing on the original ultrasound signal data to generate a high-resolution ultrasound image, and simultaneously extract the geometric modal information, dynamic modal information and texture modal information corresponding to the spatial position of the high-resolution ultrasound image.

[0024] Furthermore, for the high-resolution ultrasound image generated after full-focus imaging processing, the gray intensity value of each pixel is directly recorded as geometric modal information; for the defect area initially identified in the high-resolution ultrasound image, all A-scan signals in the corresponding spatiotemporal range in the original ultrasound signal dataset are traced back; Hilbert transform is performed on each A-scan signal to obtain the analytical signal, and then its modulus is taken to obtain the envelope signal.

[0025] The time taken for each envelope signal to rise from 10% peak value to 90% peak value is calculated and denoted as the rise time. The time taken for each envelope signal to fall from peak value to 36.8% peak value is calculated and denoted as the decay time constant. The standard deviation of all rise times and the mean of all decay time constants are combined to form dynamic modal information. A fixed-size window is extracted from the high-resolution ultrasound image with the defect area as the center, and a two-dimensional discrete wavelet transform is performed on the window image. The horizontal high-frequency sub-band, vertical high-frequency sub-band, and diagonal high-frequency sub-band obtained from the first layer decomposition are selected. The energy of the three high-frequency sub-bands is calculated respectively, and the energy is defined as the average of the sum of squares of all coefficients in the sub-band. The three energy values ​​and their corresponding gray-level co-occurrence matrix contrast and entropy values ​​are combined to form texture modal information.

[0026] It should be noted that extracting geometric, dynamic, and textural modal information simultaneously from high-resolution images and raw time-domain signals not only preserves the spatial morphological characteristics of defects but also integrates their reflection dynamics response and microstructural texture characteristics, thereby constructing a multi-dimensional and highly complementary feature system. This effectively enhances the robustness and discrimination ability of subsequent defect identification against noise interference, edge blurring, and minor anomalies.

[0027] In particular, this step breaks through the limitations of traditional ultrasonic testing that relies solely on image grayscale or single temporal features. By constructing geometric, dynamic, and textural information channels in parallel during imaging, it achieves a leap from simply seeing to truly understanding. The dynamic mode captures the transient response characteristics of defects to ultrasonic pulses, reflecting the complexity of their internal interfaces; the textural mode reveals the statistical laws of microscopic scattering structures. Together with the geometric morphology, these two modes form a multi-perspective characterization of the physical essence of defects, enabling subsequent identification to no longer rely on manual experience thresholds, but possessing inherent interpretability and generalization ability.

[0028] S3. Input the high-resolution ultrasound image into the pre-trained deep neural network, perform pixel-level semantic segmentation, output the defect mask image, and calculate the basic geometric parameters of the defect based on the defect mask image.

[0029] Furthermore, a U-Net++ network structure is constructed, which includes a 4-level encoder path and a 4-level nested decoder path. Each encoder consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, followed by a 2×2 max pooling layer. Each decoder restores spatial resolution through upsampling operations and performs channel concatenation with the outputs of multiple encoders from the same level and shallower levels. The input images are uniformly scaled to 512×512 pixels, and the feature map size is 32×32 after 4 levels of downsampling. The loss function is a weighted combination of Dice loss and focus loss, where Dice loss is used to improve the segmentation recall of small-area defects, and focus loss is used to alleviate the gradient sparsity problem caused by background pixels dominating. The network output is a 4-channel probability map, corresponding to four typical weld defects: cracks, lack of fusion, porosity, and slag inclusions. The argmax operation is applied to the output probability map to generate a single-channel defect mask image. Based on each connected component in the defect mask image, its area, equivalent diameter, principal axis direction, and perimeter are calculated as the basic geometric parameters of the defect.

[0030] It should be noted that the semantic segmentation network with deep supervision and dense jump connection structure can fully integrate multi-scale contextual information, effectively suppress background interference while maintaining accurate defect boundary localization. It has a high recall rate, especially for crack-type defects with small area but high hazard, thus providing an accurate and complete defect profile basis for subsequent quantitative analysis.

[0031] In particular, the network architecture adopted is not a simple application of a general segmentation model, but a structural adaptation made for the unique challenges of weld defects, such as small scale, blurred boundaries, and class imbalance. Its nested skip connection mechanism can effectively transmit shallow details and deep semantics, and suppress background noise interference while preserving micron-level crack edges. The design of the composite loss function guides the model to focus on the most critical small high-risk defects in engineering from the perspective of optimization objectives, ensuring that the segmentation results are not only visually accurate, but also meet the actual safety assessment needs, providing reliable input for subsequent quantitative analysis.

[0032] S4. Based on the defect contour in the defect mask image, fuse the geometric modal information, dynamic modal information and texture modal information at the corresponding positions to form a comprehensive defect fingerprint vector, and determine the defect type based on the comprehensive defect fingerprint vector, output the defect hazard level and maintenance suggestions.

[0033] Furthermore, each independent connected component in the defect mask image is used as a processing unit; the number of pixels covered by the connected component is calculated, multiplied by the actual physical area corresponding to a single pixel, to obtain the defect projection area; the equivalent circular diameter is derived from the projection area; the principal axis direction angle of the defect is calculated through principal component analysis; all A-scan envelope signals corresponding to the connected component in the original ultrasonic signal data are extracted, and the standard deviation of the envelope peak occurrence time is calculated to reflect the spatial dispersion of the internal reflection interface of the defect.

[0034] Calculate the energy concentration of all envelope signals, defined as the proportion of energy in the first 30% of the time window to the total energy; perform Gabor filtering on the local region where the connected component is located in the high-resolution ultrasound image, selecting 4 directions and 3 scales, where the 4 directions are 0°, 45°, 90° and 135°; calculate the local binary pattern histogram for each filtered response map, and concatenate its statistical features into a texture descriptor; concatenate the geometric features, dynamic features and texture descriptor in a fixed order to form a high-dimensional comprehensive defect fingerprint vector.

[0035] The comprehensive defect fingerprint vector is input into a multilayer perceptron classifier, which contains two hidden layers with 128 and 64 neurons in each layer, respectively. The output layer uses the Softmax activation function to output the probability distribution of planar or volumetric defects and maps it to a preset hazard level and maintenance recommendations.

[0036] It should be noted that by spatially aligning and deeply fusing multimodal features based on the defect contour, the resulting comprehensive fingerprint vector can fully characterize the physical nature and structural behavior of the defect. This allows the classification model to not only rely on the appearance but also reflect its internal scattering characteristics and energy distribution patterns, thereby improving the accuracy of distinguishing defects with similar appearances but different causes or hazard levels.

[0037] In particular, the construction logic of the comprehensive defect fingerprint vector embodies the feature fusion idea centered on defects. All modal information is strictly aligned to the spatial range of the same physical entity, avoiding misjudgment caused by cross-regional feature aliasing. This fusion is not a simple splicing, but extracts the most discriminative statistics from three orthogonal dimensions: spatial distribution, time response, and frequency domain energy, forming a compact representation with strong distinguishing power for defect types. The classifier driven by this can identify defects that are similar in appearance but have different physical causes, thus outputting a hazard level judgment that is more in line with engineering practice.

[0038] S5. The 3D CAD model is fused with the defect geometry reconstructed from the defect mask image and basic geometric parameters to construct a digital twin containing the real internal defects.

[0039] Furthermore, each foreground pixel in the defect mask image is mapped to a three-dimensional spatial coordinate system; the depth coordinate is determined by the ultrasonic flight time and the material velocity corresponding to that pixel, that is, the depth is equal to the velocity of sound multiplied by the flight time and then divided by 2; the normal vector of all three-dimensional points in the same defect connected domain is estimated, and the surface normal vector of each point is calculated by the local plane fitting method.

[0040] A signed distance field is constructed, with the field function value being zero at the defect surface, negative inside, and positive outside. The Poisson surface reconstruction algorithm is called to solve for the implicit surface function that satisfies the gradient field approximation of the point cloud normal. A closed triangular mesh model is generated, ensuring that the model is topologically complete and free of self-intersections. This triangular mesh model is then combined with the original workpiece 3D CAD model using a Boolean union operation to form a complete structural digital twin containing the actual internal defects. The digital twin is then geometrically cleaned, including removing small fragments, repairing non-manifold edges, and optimizing the surface quality to ensure its suitability for subsequent finite element analysis.

[0041] It should be noted that the defect information in the two-dimensional image is accurately mapped and reconstructed into a three-dimensional solid model with real spatial topology and geometric continuity, and seamlessly integrated with the original workpiece CAD model to construct a physically consistent and structurally complete digital twin. This provides a reliable geometric basis for subsequent high-fidelity mechanical simulation and avoids the stress concentration distortion problem caused by traditional simplified modeling.

[0042] In particular, the 3D reconstruction of the defect geometry does not merely pursue visual realism, but aims to meet the stringent requirements of mechanical simulation for geometric continuity and topological correctness. By combining time-of-flight physical constraints and surface reconstruction algorithms, it ensures that the generated defect model has real physical meaning in the depth direction, maintains a smooth and closed surface morphology, and is seamlessly integrated with the parent material model. This high-fidelity digital twin eliminates the idealized assumptions about the defect shape in traditional manual modeling, enabling subsequent stress field calculations to truly reflect the geometric singularities of the actual defect tip, effectively improving the engineering credibility of the simulation results.

[0043] S6. Perform online mechanical simulation on the digital twin based on preset load conditions, calculate the stress intensity factor at the defect tip, the remaining strength of the workpiece, and the fatigue life, and generate a structural health assessment report.

[0044] Furthermore, preset mechanical loads, thermal loads, or fatigue load spectra are applied to the digital twin; the material properties adopt a bilinear elastoplastic constitutive model, with input yield strength, elastic modulus, hardening modulus, and Poisson's ratio; the overall model is divided into tetrahedral finite element meshes, and local mesh refinement is implemented in the defect tip region, with the element size in the refined region not exceeding 1 / 10 of the minimum radius of curvature at the defect tip; transient analysis is performed using an explicit dynamic solver, or static / quasi-static analysis is performed using an implicit solver.

[0045] Stress field data is extracted from the defect leading edge, and the stress intensity factor is calculated along a preset integration path. ,in Calculated using the following formula: ; In the formula, This represents the radial distance measured from the tip of the defect. This represents the normal stress components along the defect propagation direction in polar coordinates. Represents the Type I stress intensity factor; the calculated value will be used to determine the stress intensity factor. Compared with the material fracture toughness KIC, it is determined whether instability propagation has occurred; if used for fatigue assessment, the material SN curve and Miner's linear cumulative damage theory are combined to calculate the damage increment for each load cycle, and the number of cycles corresponding to when the total damage value reaches 1 is the predicted remaining fatigue life.

[0046] Summarize the defect location, type identification results, hazard level, basic geometric parameters, and stress intensity factor. The remaining strength safety margin and predicted remaining fatigue life are defined as the ratio of allowable stress to the maximum working stress obtained from simulation. When the remaining strength safety margin is less than the preset safety factor, or the predicted remaining fatigue life is less than the next planned maintenance cycle, the defect is marked as high risk.

[0047] In the 3D visualization interface, the defect geometry is displayed on the workpiece model in a semi-transparent red overlay, with the defect number and key parameters labeled; a structured text description is generated, including defect cause analysis, mechanical impact assessment, maintenance method suggestions, and recommended time for the next inspection; a PDF document conforming to ASME or ISO standard format is output, including a cover, defect list, simulation result screenshots, safety assessment conclusions, and an electronic signature area; the system automatically archives the report to the equipment lifecycle management database and triggers a notification to the maintenance scheduling platform.

[0048] It should be noted that online mechanical simulation based on real defect geometry can accurately quantify the impact of defects on structural load-bearing capacity, directly output engineering-interpretable safety margins and life prediction indicators, and, combined with visualization and standardized report automatic generation mechanisms, seamlessly transform detection results into actionable operation and maintenance decisions, greatly improving the closed-loop efficiency and reliability from problem discovery to problem resolution.

[0049] In particular, this step advances nondestructive testing from qualitative discovery to quantitative prediction. Its core lies in establishing a direct mapping link from test data to structural performance indicators. By applying actual working condition loads to the real defect geometry and using refined meshes and material constitutive models for simulation, the resulting stress intensity factor and fatigue life are no longer theoretical estimates, but personalized evaluation results based on the measured defect morphology. This integrated process of detection-modeling-simulation-decision-making transforms maintenance recommendations from experience-driven to data- and physics-driven, improving the scientific rigor and foresight of in-service safety assessments of major equipment.

[0050] In summary, this invention improves the positioning accuracy, identification accuracy, and assessment reliability of weld defect detection by integrating a closed-loop process through three-dimensional CAD model-driven adaptive ultrasonic scanning, full-focus high-resolution imaging, simultaneous extraction of multimodal features, deep learning pixel-level defect segmentation, physical consistency-based defect fingerprint construction, real defect geometry reconstruction, and digital twin analysis incorporating mechanical simulation. It not only achieves comprehensive quantitative judgment from the presence or absence of defects to defect type, morphology, hazard level, and remaining lifespan, but also avoids the subjectivity and experience dependence of traditional manual interpretation. Furthermore, it directly outputs structural safety margins and maintenance decision-making basis through online mechanical simulation, enhancing the intelligence, automation, and engineering practicality of in-service welded structure health monitoring.

[0051] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An ultrasonic inspection method for automatic identification and evaluation of weld defects, characterized in that: include: Based on the 3D CAD model of the workpiece to be inspected, an adaptive scanning path for the phased array ultrasonic probe is planned, and the original ultrasonic signal data is acquired through the full matrix acquisition mode. The original ultrasound signal data is processed by full-focus imaging to generate a high-resolution ultrasound image, and the geometric modal information, dynamic modal information and texture modal information corresponding to the spatial position of the high-resolution ultrasound image are extracted simultaneously. High-resolution ultrasound images are input into a pre-trained deep neural network for pixel-level semantic segmentation, outputting a defect mask image, and the basic geometric parameters of the defect are calculated based on the defect mask image. Based on the defect contour in the defect mask image, the geometric modal information, dynamic modal information and texture modal information at the corresponding position are fused to form a comprehensive defect fingerprint vector. Based on the comprehensive defect fingerprint vector, the defect type is identified and the defect hazard level and maintenance suggestions are output. By fusing a 3D CAD model with a defect geometry reconstructed from a defect mask image and basic geometric parameters, a digital twin containing real internal defects is constructed. Online mechanical simulation is performed on the digital twin based on preset load conditions to calculate the stress intensity factor at the defect tip, the remaining strength of the workpiece, and the fatigue life, generating a structural health assessment report.

2. The ultrasonic inspection method for automatic identification and evaluation of weld defects as claimed in claim 1, wherein: The adaptive scanning path of the phased array ultrasonic probe is planned based on the three-dimensional CAD model of the workpiece to be inspected. The specific steps are as follows: Extract the weld centerline from the 3D CAD model and divide the control nodes along the weld length direction with a fixed step size; At each control node, the corresponding base material thickness and bevel geometry parameters are read, and the local incident surface normal vector is calculated. According to the longitudinal wave speed in the weld material and the longitudinal wave speed in the probe wedge material combining the bevel half angle solving for the probe deflection angle required to have the ultrasonic beam normal to the weld fusion line where the deflection angle satisfies the following relationship: ; In the formula, This indicates the longitudinal wave propagation velocity in the weld metal material. This indicates the longitudinal wave propagation velocity in the wedge material used in the ultrasonic probe. This represents the angle of a single-sided bevel, which is half of the total bevel angle. This indicates the mechanical deflection angle that the probe needs to be set at this control node; The probe spatial coordinates and deflection angles corresponding to all control nodes are used to generate a continuous and smooth scanning trajectory through cubic spline interpolation; In areas where the radius of curvature of the weld is less than a preset threshold, the density of control nodes is automatically increased to ensure that the sound beam coverage has no blind spots.

3. The ultrasonic testing method for automatic identification and evaluation of weld defects as described in claim 2, characterized in that: The specific steps for simultaneously extracting geometric modal information, dynamic modal information, and texture modal information are as follows: For high-resolution ultrasound images generated after full-focus imaging processing, the grayscale intensity value of each pixel is directly recorded as geometric modal information. For the defect areas initially identified in the high-resolution ultrasound images, all A-scan signals within the corresponding spatiotemporal range in the original ultrasound signal dataset are traced back. Perform a Hilbert transform on each A-scan signal to obtain an analytical signal, and then take its modulus to obtain the envelope signal; Calculate the time it takes for each envelope signal to rise from 10% peak value to 90% peak value, and record it as the rise time. Calculate the time it takes for each envelope signal to drop from its peak value to 36.8% of its peak value, and denote it as the decay time constant; The dynamic modal information is formed by combining the standard deviation of all rise times with the mean of all decay time constants. A fixed-size window is extracted from the high-resolution ultrasound image with the defect region as the center, and a two-dimensional discrete wavelet transform is performed on the window image. Select the horizontal high-frequency sub-band, vertical high-frequency sub-band, and diagonal high-frequency sub-band obtained from the first layer decomposition; Calculate the energy of each of the three high-frequency subbands. The energy is defined as the average of the sum of squares of all coefficients within the subband. The three energy values ​​and their corresponding gray-level co-occurrence matrix contrast and entropy values ​​are combined to form the texture modal information.

4. The ultrasonic testing method for automatic identification and evaluation of weld defects as described in claim 3, characterized in that: The specific steps for inputting high-resolution ultrasound images into a pre-trained deep neural network for pixel-level semantic segmentation are as follows: Construct the U-Net++ network structure, which includes a 4-level encoder path and a 4-level nested decoder path; Each encoder consists of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function, followed by a 2×2 max pooling layer. Each level of decoder restores spatial resolution through upsampling and performs channel splicing with the outputs of multiple encoders from the same level and shallower levels. The input image is uniformly scaled to 512×512 pixels, and the feature map size is 32×32 after 4 levels of downsampling. The loss function uses a weighted combination of Dice loss and focus loss, where Dice loss is used to improve the segmentation recall of small defects, and focus loss is used to alleviate the gradient sparsity problem caused by background pixels dominating. The network output is a 4-channel probability map, which corresponds to four typical weld defects: cracks, lack of fusion, porosity, and slag inclusions. Apply the argmax operation to the output probability map to generate a single-channel defect mask image; For each connected component in the defect mask image, calculate its area, equivalent diameter, principal axis direction, and perimeter, which serve as the basic geometric parameters of the defect.

5. The ultrasonic testing method for automatic identification and evaluation of weld defects as described in claim 4, characterized in that: The specific steps for fusing multimodal information to form a comprehensive defect fingerprint vector based on the defect contour in the defect mask image are as follows: Each independent connected component in the defect mask image is used as a processing unit; Calculate the number of pixels covered by the connected component, multiply it by the actual physical area corresponding to a single pixel, and obtain the defect projection area; Derive the equivalent circle diameter from the projected area; The principal axis orientation angle of the defect is calculated using principal component analysis. Extract all A-scan envelope signals corresponding to the connected component in the original ultrasound signal data, calculate the standard deviation of the envelope peak occurrence time, and reflect the spatial dispersion of the internal reflection interface of the defect. Calculate the energy concentration of all envelope signals. The energy concentration is defined as the proportion of energy in the first 30% of the time window to the total energy. Gabor filtering was performed on the local region containing the connected component in the high-resolution ultrasound image, using 4 directions and 3 scales, with the 4 directions being 0°, 45°, 90° and 135°; For each filtered response map, a local binary pattern histogram is calculated, and its statistical features are concatenated into a texture descriptor. Geometric features, dynamic features, and texture descriptors are concatenated in a fixed order to form a high-dimensional comprehensive defect fingerprint vector. The comprehensive defect fingerprint vector is input into a multilayer perceptron classifier, which contains two hidden layers with 128 and 64 neurons in each layer, respectively. The output layer uses the Softmax activation function to output the probability distribution of planar or volumetric defects and maps it to a preset hazard level and maintenance recommendations.

6. The ultrasonic testing method for automatic identification and evaluation of weld defects as described in claim 5, characterized in that: The specific steps for reconstructing the defect geometry from the defect mask image and basic geometric parameters are as follows: Map each foreground pixel in the defect mask image to a three-dimensional coordinate system; The depth coordinates are determined by the ultrasonic flight time and the material velocity corresponding to the pixel point, that is, the depth is equal to the velocity of sound multiplied by the flight time and then divided by 2; Normal vectors are estimated for all three-dimensional points within the same defective connected domain, and the surface normal vector of each point is calculated using the local plane fitting method. Construct a signed distance field, where the field function value is zero at the defect surface, negative inside, and positive outside; The Poisson surface reconstruction algorithm is invoked to solve for the implicit surface function that satisfies the gradient field approximation of the point cloud normal. Generate a closed triangular mesh model to ensure that the model has a complete topology and no self-intersections; The triangular mesh model is combined with the original 3D CAD model of the workpiece using a Boolean union operation to form a complete structural digital twin containing real internal defects. The digital twin undergoes geometric cleanup, including removing micro-fragments, repairing non-manifold edges, and optimizing surface quality to ensure its suitability for subsequent finite element analysis.

7. The ultrasonic testing method for automatic identification and evaluation of weld defects as described in claim 6, characterized in that: The specific steps for performing online mechanical simulation of the digital twin based on preset load conditions are as follows: Apply a preset mechanical load, thermal load, or fatigue load spectrum to the digital twin; The material properties are determined using a bilinear elastoplastic constitutive model, with inputs including yield strength, elastic modulus, hardening modulus, and Poisson's ratio. The overall model is divided into tetrahedral finite element meshes, and local mesh refinement is implemented in the defect tip region. The element size in the refined region is no greater than 1 / 10 of the minimum radius of curvature of the defect tip. Transient analysis can be performed using an explicit dynamics solver, or static / quasi-static analysis can be performed using an implicit solver; Stress field data is extracted from the defect leading edge, and the stress intensity factor is calculated along a preset integration path. ,in Calculated using the following formula: ; In the formula, This represents the radial distance measured from the tip of the defect. This represents the normal stress components along the defect propagation direction in polar coordinates. Indicates Type I stress intensity factor; The calculated result Compare with the material fracture toughness KIC to determine whether instability propagation has occurred; If used for fatigue assessment, the damage increment is calculated for each load cycle by combining the material's SN curve and Miner's linear cumulative damage theory. The number of cycles corresponding to when the total damage value reaches 1 is the predicted remaining fatigue life.

8. The ultrasonic testing method for automatic identification and evaluation of weld defects as described in claim 7, characterized in that: The specific steps for generating the structural health assessment report are as follows: Summarize the defect location, type identification results, hazard level, basic geometric parameters, and stress intensity factor. Remaining strength safety margin and predicted remaining fatigue life; The residual strength safety margin is defined as the ratio of the allowable stress to the maximum working stress obtained from the simulation. When the remaining strength safety margin is less than the preset safety factor, or the predicted remaining fatigue life is less than the next planned maintenance cycle, the defect is marked as high risk. In the 3D visualization interface, the defect geometry is displayed on the workpiece model as a semi-transparent red overlay, and the defect number and key parameters are labeled. Generate structured text descriptions, including defect cause analysis, mechanical impact assessment, repair method suggestions, and recommended time for the next inspection; Output a PDF document conforming to ASME or ISO standard format, including a cover, defect list, simulation result screenshots, safety assessment conclusions, and an electronic signature area; The system automatically archives the report to the equipment lifecycle management database and triggers a notification to the maintenance scheduling platform.

Images