Weld joint ultrasonic phased array detection data intelligent analysis system

Abstract

The invention discloses an intelligent analysis system for ultrasonic phased array detection data of a welding seam, and relates to the technical field of nondestructive testing, and the intelligent analysis system is characterized in that a four-dimensional wave field tensor is constructed by collecting full-waveform ultrasonic data under multi-channel, multi-path and multi-angle conditions; extracting reflected signals which keep time coherence in a discontinuous path, and generating a high-dimensional coherence feature matrix; inputting the features into a self-supervised contrast learning model to obtain a defect semantic embedding vector; recognizing a suspected weak defect area based on the distribution density and the boundary change trend, and performing reverse beam focusing in combination with original data to obtain a defect three-dimensional positioning map; the shape bifurcation index and the boundary stability are calculated through topological analysis, and intelligent discrimination of artifacts and microcrack defects is achieved; the method does not need label data, has high automation and robustness, and is suitable for weak defect identification of complex structure welding seams.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to an intelligent analysis system for ultrasonic phased array testing data of welds. Background Technology

[0002] As a critical connection point in pressure-bearing or high-speed operating structures such as pressure vessels, petrochemical pipelines, nuclear power equipment, and high-speed train bodies, the quality of welds directly affects the safety and service life of the overall structure. Phased-array ultrasonic testing (PAUT) has been widely used in weld inspection due to its advantages of high sensitivity, intuitive imaging, and high testing efficiency.

[0003] However, the complex structure of welds presents challenges to the accurate identification and evaluation of phased array detection data, including redundant background noise, interweaving of various types of defect signals, and unstable reflection paths. This is especially true in confined spaces, complex curved welds, or multi-layered welded components, where defect characteristic signals are extremely weak and easily misinterpreted as structural echoes or artifacts.

[0004] Most common analysis methods currently rely on human experience or rule-based algorithms to interpret detected images, which is not only inefficient but also easily influenced by human subjectivity, resulting in serious problems of missed detections and false positives. At the same time, existing methods cannot fully mine the deep feature information in the original data of multi-channel and multi-angle phased arrays, resulting in weak ability to identify minute defects in complex environments. In particular, in the welds of thick-walled components with a wall thickness greater than 30 mm, it is extremely difficult to reliably identify minute defects such as fine cracks. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent analysis system for ultrasonic phased array detection data of welds to address the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an intelligent analysis system for ultrasonic phased array detection data of welds, comprising:

[0007] Data acquisition module: Acquires the raw dataset D of ultrasonic phased array detection of the weld to be tested;

[0008] Spatiotemporal reconstruction module: Performs spatiotemporal joint reconstruction processing on D to construct a four-dimensional wave field tensor Q with dynamic propagation path reversibility, including time, channel position, incident angle and scan path;

[0009] Feature extraction module: Extracts the reflection group signals in Q that still maintain temporal coherence during discontinuous path propagation, and constructs a high-dimensional coherent feature matrix F, which includes multi-path wavefront coupling degree, multi-channel phase coordination factor and cross-angle spectral entropy ratio;

[0010] Defect representation model training module: Input F into the contrast feature embedding model M trained by self-supervised learning to generate defect semantic embedding vector Z. The model M achieves feature clustering learning by automatically constructing positive and negative sample pairs.

[0011] 3D localization module: Based on the distribution density of Z and the trend of semantic boundary tensor change, identify suspected weak defect regions R, and combine the delay time and energy loss of the corresponding position in Q to perform reverse beam focusing calculation to obtain a 3D fine localization map T of the defect point.

[0012] Evaluation module: Performs in-depth topological analysis on the defect region in T, calculates its shape bifurcation index and structural boundary stability to distinguish between material structure artifacts and real microcrack-like defects, and outputs the final defect analysis results.

[0013] Preferably, the dataset D contains multiple channels, multiple scan paths, and full waveform echo signals at multiple incident angles.

[0014] Preferably, the spatiotemporal joint reconstruction process of D includes:

[0015] The echo signals from each channel are processed uniformly along the time axis.

[0016] Phase alignment is performed on each aligned channel signal;

[0017] Signals from multiple scanning paths are uniformly mapped to an equally spaced three-dimensional grid according to their actual physical coordinates. Sparse regions are filled by weighted averaging of neighboring measurement points, thus constructing a wavefield dataset with spatial continuity.

[0018] Organize time-domain aligned, phase-consistent, and spatially continuous data into a unified four-dimensional wavefield tensor Q.

[0019] Preferably, the extraction of reflection group signals in Q that retain temporal coherence during discontinuous path propagation includes:

[0020] Time-domain windowing is performed on the wavefield slices corresponding to each incident angle in the four-dimensional wavefield tensor Q to extract the main energy segment of the wave packet.

[0021] Calculate the temporal coherence coefficient of wave packets between different scanning paths, and use a correlation threshold of 0.8 as the standard to screen out echo signals with cross-path coherence;

[0022] Based on the phase stability and energy distribution continuity of coherent signals, multi-scale wavelet packet decomposition is used to extract their multipath wavefront coupling characteristics.

[0023] The multipath wavefront coupling features are reorganized according to the channel, path and angle dimensions to construct a high-dimensional coherent feature matrix F.

[0024] Preferably, the method for obtaining the multipath wavefront coupling degree, multichannel phase coordination factor, and cross-angle spectral entropy ratio includes:

[0025] To obtain the multipath wavefront coupling degree, the normalized cross-correlation function is calculated in units of sampled signals within the time window for the selected cross-path coherent echoes. The peak cross-correlation value and peak delay are calculated between each pair of paths. The peak cross-correlation values ​​of all path pairs are summed according to the path logarithm and normalized by the path logarithm to obtain the multipath wavefront coupling degree.

[0026] To obtain the multi-channel phase coordination factor, the Hilbert transform of each channel signal at the same spatiotemporal point is first performed to obtain the instantaneous phase. The phase lock value (PLV) of the phase difference sequence between any two channels within the time window is calculated, and the average value of all channels relative to the PLV is used as the multi-channel phase coordination factor. The PLV calculation adopts the mode length of the average of the complex phase vectors.

[0027] To obtain the cross-angle spectral entropy ratio, the power spectral density is estimated using the Welch method for the selected time-domain wave packet at each incident angle, the spectral entropy of the power spectrum is calculated, and the spectral entropy within the target frequency band is compared with the spectral entropy of all frequency bands to obtain the spectral entropy ratio for each pair of angles. Finally, the median of the angle-to-spectral entropy ratio is taken as the cross-angle spectral entropy ratio.

[0028] Preferably, the step of inputting F into the contrastive feature embedding model M trained by self-supervised learning to generate the defect semantic embedding vector Z includes:

[0029] The high-dimensional feature matrix F is standardized to generate a multi-view input sample set;

[0030] Positive and negative sample pairs are automatically constructed based on unlabeled data. Positive sample pairs consist of signal segments from the same physical region or with similar feature distributions, while negative sample pairs consist of signals from different spatial regions or with low coherence.

[0031] The sample pairs are input into the contrastive feature embedding model M, which contains a shared encoder. The model M adopts a two-branch convolutional neural network structure and is trained to minimize the contrastive loss through feature cosine similarity metric.

[0032] After the model converges, the encoded F is embedded and mapped to obtain a low-dimensional defect semantic embedding vector Z, which represents the spatiotemporal coherent semantic features of weld defects.

[0033] Preferably, obtaining the three-dimensional fine location map T of the defect point includes:

[0034] Density clustering analysis was performed on the defect semantic embedding vector Z, and an adaptive clustering algorithm based on local reachability density was used to identify abnormally sparse distribution areas in order to preliminarily determine the suspected defect feature clusters.

[0035] Calculate the semantic boundary tensor gradient change rate for each feature cluster boundary region. When the change rate exceeds a preset threshold, the region is determined to have a significant semantic mutation and is marked as a suspected defect region R.

[0036] Retrieve the spatiotemporal coordinate data corresponding to R in Q, extract the delay time and reflection energy distribution, and perform reverse beam focusing calculation on the signals of each channel based on the time delay inversion algorithm;

[0037] By superimposing multiple channels and using energy-weighted averaging, a defect location point cloud in three-dimensional spatial coordinates is constructed, forming a three-dimensional fine location map T of the defect point.

[0038] Preferably, the step of performing deep topological analysis on the defect region in T to calculate its shape bifurcation index and structural boundary stability includes:

[0039] Spatial segmentation is performed on each suspected defect point cloud region in the 3D localization map T, and a density-based clustering algorithm is used to identify independent defect bodies to generate a set of candidate defect regions.

[0040] For each candidate region, the shape bifurcation index is calculated, and the three-dimensional contour change rate is extracted based on the principal curvature analysis of the point cloud. When there are more than three abrupt changes in the direction of the principal curvature axes within a unit volume, it is judged as a complex topological bifurcation structure.

[0041] The candidate defect boundary is sliced ​​at multiple angles, and the boundary fluctuation coefficient of each slice is calculated. If the boundary shape remains stable in multiple angle slices, the structural boundary is considered to have high stability.

[0042] By combining the bifurcation index and boundary stability index, a support vector machine model is used to classify artifacts and microcracks, and the final defect analysis results are output.

[0043] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0044] 1. This invention provides an intelligent analysis method for ultrasonic phased array inspection data of welds. By constructing a dynamically propagating, reversible four-dimensional wavefield tensor and combining it with high-dimensional coherent feature extraction across channels, paths, and angles, it effectively preserves the coherence and physical consistency of weak defect signals in multidimensional space, significantly improving the ability to identify difficult defects such as microcracks in complex weld structures. Compared with traditional detection methods based on image or signal amplitude processing, this invention exhibits higher robustness in defect extraction under strong interference backgrounds and is suitable for intelligent non-destructive testing of thick-walled, irregular curved surfaces, and multi-layer welded structures.

[0045] 2. This invention introduces a semantic embedding model based on self-supervised contrastive learning to achieve low-dimensional clustering representation of high-dimensional ultrasonic features. By combining semantic density variation and boundary tensor analysis with delay-energy inversion focusing and three-dimensional topology recognition, an interpretable three-dimensional map of defect points is finally constructed. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0047] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] For examples, please refer to Figure 1 As shown in this embodiment, an intelligent analysis system for ultrasonic phased array inspection data of weld seams includes:

[0050] Data acquisition module: Acquires the raw dataset D of ultrasonic phased array detection of the weld to be tested;

[0051] Spatiotemporal reconstruction module: Performs spatiotemporal joint reconstruction processing on D to construct a four-dimensional wave field tensor Q with dynamic propagation path reversibility, including time, channel position, incident angle and scan path;

[0052] Feature extraction module: Extracts the reflection group signals in Q that still maintain temporal coherence during discontinuous path propagation, and constructs a high-dimensional coherent feature matrix F, which includes multi-path wavefront coupling degree, multi-channel phase coordination factor and cross-angle spectral entropy ratio;

[0053] Defect representation model training module: Input F into the contrast feature embedding model M trained by self-supervised learning to generate defect semantic embedding vector Z. The model M achieves feature clustering learning by automatically constructing positive and negative sample pairs.

[0054] 3D localization module: Based on the distribution density of Z and the trend of semantic boundary tensor change, identify suspected weak defect regions R, and combine the delay time and energy loss of the corresponding position in Q to perform reverse beam focusing calculation to obtain a 3D fine localization map T of the defect point.

[0055] Evaluation module: Performs in-depth topological analysis on the defect region in T, calculates its shape bifurcation index and structural boundary stability to distinguish between material structure artifacts and real microcrack-like defects, and outputs the final defect analysis results.

[0056] In a preferred embodiment of the present invention, weld inspection data is first acquired to obtain a high-dimensional, multi-channel raw ultrasonic phased array signal dataset to support subsequent spatiotemporal feature reconstruction and intelligent defect identification. Specifically, the acquisition steps include the following:

[0057] The object to be tested is a welded structural component. The weld material can be carbon steel, stainless steel, or other alloy materials, and the structural form can include butt welds, fillet welds, T-welds, etc. To enhance data coverage, the testing area should at least include the weld center, the heat-affected zone, and the near-base material area.

[0058] An ultrasonic phased array probe with a linear or ring array is used, with the number of array elements selectable as 32, 64 or 128, and the operating frequency range is 1 MHz to 10 MHz, depending on the material thickness and the expected defect size.

[0059] The probe performs multi-path scanning along the weld seam, including linear scanning, offset scanning, and staggered path scanning. Each scanning path covers different lateral positions to enhance spatial information redundancy and improve the ability to identify weak signals.

[0060] To improve beam coverage density, multiple incident angles are configured for each scanning path, typically 100°. Several angles distributed within the range (e.g.) (etc.) to achieve angle reuse and directionality enhancement, and acquire echo data from different incident directions.

[0061] The final obtained original dataset D is a four-dimensional dataset, and its data structure can be represented as follows: ;in: This represents the i-th channel. This represents the j-th scanning path. Let represent the k-th incident angle, and s represent the echo signal of each channel in the time domain.

[0062] Each channel acquires complete A-Scan waveform data at each incident angle and along each scanning path, preserving the original reflection intensity, waveform shape, and phase change information, providing a foundation for subsequent in-depth data analysis.

[0063] In one embodiment of the present invention, in order to achieve deep fusion and coherent enhancement of defect echo signals in time, space and angle dimensions, the original ultrasound dataset D needs to be spatiotemporally reconstructed to construct a four-dimensional wavefield tensor Q.

[0064] First, time alignment is performed on the echo signals of all channels in the original data D. The purpose is to synchronize the main echo signals of all channels to a unified reference starting point. The specific operation is as follows:

[0065] The channel with the largest signal amplitude is selected as the reference channel, denoted as channel R. For any channel i to be aligned, the correlation between its echo signal and the reference channel R in time is calculated. The position of maximum correlation of the signal of this channel relative to the reference channel is found, and the signal is shifted forward or backward by the corresponding time point for alignment. The time shift is in units of sampling points, and the time accuracy is determined by the sampling rate (e.g., when the sampling rate is 20 MHz, the minimum step size is 0.05 microseconds). After the above operations are completed for all channels, the main echo of all signals will appear at a unified starting point, eliminating the initial delay difference between array elements.

[0066] Due to manufacturing errors, uneven coupling, or differences in propagation paths between probe elements, the echo signal may exhibit slight phase drift. This step extracts the instantaneous phase and performs relative correction to ensure that all channel signals maintain phase consistency. The specific method is as follows:

[0067] For each aligned channel signal, its analytic signal is constructed as follows: the original real signal and its Hilbert transform are combined into a complex number; the real part of the analytic signal is the original signal, and the imaginary part is the result of its Hilbert transform; the phase angle of each sampling point is calculated using the arctangent function to form the instantaneous phase sequence of the channel; for any channel, its phase sequence is compared with the phase sequence of the reference channel, the phase difference is calculated, and phase shift is performed according to the difference; the phase shift does not change the signal shape, but only adjusts its phase position, thereby ensuring wavefront coherence.

[0068] Because the coverage areas of array elements along different scanning paths may shift during actual detection, it is necessary to map the signal data from all paths to a unified spatial reference coordinate system to achieve spatial continuity. The operation steps are as follows:

[0069] For each scanning path, define its starting coordinates, array direction, and element spacing to form a spatial arrangement matrix corresponding to the path; represent the position of each signal measurement point as a three-dimensional coordinate point (x, y, z), which is automatically calculated based on the array structure; construct an equally spaced three-dimensional grid in the target space, with a grid size typically of 1 mm; for areas in the grid without sampled data, use a weighted average method to complete the signal value; the weight is determined based on the spatial distance between the grid point and the surrounding actual sampling points, with higher weights for closer distances; the distance and weight adopt an exponential decay relationship, typically with a decay radius of 3 mm; for each grid point, use the actual sampling points in its neighborhood for weighted superposition to estimate the signal value of that point, achieving sparse completion.

[0070] After completing the time-domain, phase-domain, and spatial completion processes described above, all signals are organized into a unified data volume, namely the four-dimensional wavefield tensor Q. Its structure is defined as follows: the first dimension is the time series, representing each signal sampling point; the second dimension is the channel position, corresponding to the array element number; the third dimension is the scan path number; and the fourth dimension is the incident angle number (or the setting sequence number corresponding to the incident angle).

[0071] Any element Q(t,c,p,a) in the tensor represents the signal amplitude at time t, channel c, path p, and incident angle a.

[0072] In a preferred embodiment of the present invention, in order to further extract physically consistent reflection features from the four-dimensional wavefield tensor Q and improve the ability to identify weak defect signals, it is necessary to perform multi-dimensional coherence analysis on the echo information in Q across paths, channels, and angles, and then construct a high-dimensional coherence feature matrix F. This process includes two main stages: first, extracting coherent reflection group signals; and second, calculating feature parameters and constructing F.

[0073] The four-dimensional wavefield tensor Q is sliced ​​according to each incident angle dimension to obtain multiple three-dimensional wavefield data blocks (including channel, path, and time dimensions). For each wavefield slice, the main energy segment of the echo is extracted based on a preset time window (e.g., the time of wave crest appearance ± 5 microseconds). The size of the time window can be calculated based on the sound wave propagation time to ensure that the complete defect echo is included.

[0074] In the windowed signal, normalized cross-correlation is performed on signals with the same channel and incident angle under any two scanning paths. The maximum value of the cross-correlation function is taken as the temporal coherence coefficient, representing the similarity of signals between the two paths. If this coefficient is greater than a set threshold (0.8), the echo is considered to have physical coherence between the paths, and the signal is retained for subsequent feature calculations.

[0075] For the selected coherent signals, a multi-scale wavelet packet decomposition algorithm (such as one based on the Daubechies wavelet basis) is used to extract their local energy distribution and wavefront structure in different frequency bands. By comparing the similarity of coefficients at the same scale under each path, the wavefront coupling relationship between multi-path signals is characterized, providing basic data for constructing a high-dimensional feature matrix.

[0076] The coherent signals described above are rearranged in a three-dimensional order of channel, path, and angle to form a dataset with a clearly defined structure. This dataset is used to calculate three key characteristic parameters: multipath wavefront coupling, multichannel phase coordination factor, and cross-angle spectral entropy ratio.

[0077] Multipath wavefront coupling calculation method: For all selected cross-path coherent signals, calculate their normalized cross-correlation function on a per-sample-point basis within each time window. For any path pair, the maximum value of its cross-correlation function is used as the wavefront coupling index for that path pair. Average the cross-correlation peaks of all path pairs and normalize them by the total number of path pairs to obtain a global multipath wavefront coupling scalar. To improve effectiveness, a path pair is only included in the statistics if the time delay corresponding to the cross-correlation peak is less than a preset threshold Δt (e.g., 0.5 microseconds).

[0078] For example, this invention selects echo signal segments with cross-path temporal coherence that have been screened in the previous stage, denoted as the signal set. Where i = 1, 2, ..., N, represents different scanning paths; N is the total number of scanning paths, each For echo signals within a uniform time window, the sampling length is L (e.g., L = 512 sampling points); the time window width is set according to the main energy segment of the wave packet, and is generally within the range of ±5 microseconds.

[0079] For each pair of paths (i,j), calculate its corresponding signal. and Normalized cross-correlation function Defined as: ; Where: t is the sampling time, and τ is the time delay variable, with a value range of... For example, it can be set to ±10 sampling points; the cross-correlation function has been normalized to its maximum value, and its range is... ;

[0080] For each path pair (i,j), record the maximum value of the cross-correlation function. and its corresponding peak latency The path is included in the coupling statistics only if the following conditions are met: Among them, peak delay threshold This is a preset value used to filter signal pairs with inconsistent propagation paths or severe distortion. It is generally set to 5 sampling points, corresponding to approximately 0.5 microseconds (in a system with a sampling rate of 10 MHz).

[0081] For the set of all path pairs that satisfy the delay condition The average value of their cross-correlation peaks is calculated as the multipath wavefront coupling degree. The formula is as follows: ; in, The total number of path pairs that satisfy the conditions.

[0082] The final MPWCI index ranges from 0 to 1, with a higher value indicating more consistent propagation and stronger coupling of the wavefront across multiple paths. Please refer to Table 1:

[0083] Table 1. Statistical relationship between multipath wavefront coupling degree (MPWCI) and defect reliability, for example: Serial Number Defect types MPWCI average Is it a real defect? Model consensus rate (%) 1 Material artifacts 0.42 no 91.2 2 Unstructured echo 0.38 no 88.6 3 Microcrack defects 0.81 yes 95.4 4 Porous defects 0.76 yes 92.3

[0084] As shown in Table 1, the multipath wavefront coupling degree is highly correlated with the actual defects, and the reliability of defects with MPWCI>0.75 is significantly higher, supporting the practical identification value of the feature parameters of this invention.

[0085] Multi-channel phase coordination factor calculation method: At each spatiotemporal point, a Hilbert transform is applied to the signals of all channels to obtain instantaneous phase information. For each pair of channels, the phase-locked value (PLV) of their phase difference is calculated within a time window. This value is defined as: for any two channels i and j, the phase difference within the time window is calculated. The expression is: In the formula, Let be the instantaneous phase of the i-th channel signal. Let be the instantaneous phase of the j-th channel signal; the average of the phase difference vector expressed in complex form is the magnitude of PLV, ranging from 0 to 1, with a value closer to 1 indicating stronger phase coordination; finally, the average PLV of all channel pairs is taken as the multi-channel phase coordination factor at that point. In this embodiment, if the average PLV value is greater than 0.7, it is considered that there is significant inter-channel phase coordination.

[0086] Method for calculating the cross-angle spectral entropy ratio: For the echo signal within a selected time window at each incident angle, the Welch method is used to calculate its power spectral density (PSD). The steps are as follows: Divide the signal into segments (e.g., each segment has a length of 128 sampling points and an overlap rate of 50%); apply a window function (e.g., Hanning window) to each segment; perform a fast Fourier transform (FFT) on each segment; and average the power spectra of all segments to obtain the power spectral density (PSD) at that angle.

[0087] After normalizing the PSD, its spectral entropy value is calculated according to the Shannon entropy definition. The expression is: ; In the formula, f is the frequency index, representing each discrete frequency component in the power spectrum; The normalized power spectral density is calculated by dividing the power spectral density at a selected angle by the power spectral density at all angles.

[0088] For each pair of incident angle signals, calculate its dominant frequency band (e.g., The ratio of the spectral entropy of (MHz) to the spectral entropy of the entire frequency band is used to obtain the spectral entropy ratio of each pair of angles. The median value of all angle pairs is used as the cross-angle spectral entropy ratio of the data block.

[0089] The above three feature results are organized into a feature vector and then concatenated according to the channel, path and angle dimensions to construct the final high-dimensional coherent feature matrix F.

[0090] In a preferred embodiment of the present invention, in order to realize the deep semantic representation of defect features in the ultrasonic phased array detection data of weld, the high-dimensional coherent feature matrix F that has been extracted needs to be input into a contrastive feature embedding model M trained based on a self-supervised contrastive learning mechanism to generate a low-dimensional defect semantic embedding vector Z.

[0091] To enhance the model's robustness to different sampling scenarios and signal perturbations, the high-dimensional coherent feature matrix F is first normalized and perturbation-enhanced. Normalization employs the Z-score method, standardizing each feature dimension based on its mean and standard deviation. Perturbation enhancement includes:

[0092] Gaussian noise superposition: Add zero-mean, zero-variance Gaussian white noise to each feature dimension of F;

[0093] Feature Dropout: Randomly masking some feature dimensions with a set probability (e.g., 10%) to simulate the situation of missing features.

[0094] Through the above operations, a multi-view augmented sample set F′ is constructed and used as the training input for the model.

[0095] Since the training process does not rely on manual annotation, this invention designs an automatic sample pair construction mechanism to achieve unsupervised feature comparison learning. The steps are as follows:

[0096] Positive sample pair definition: Signal segments from the same spatial region or with similar coherent feature distributions are extracted from F′ and used as positive sample pairs input to the model;

[0097] Negative sample pair definition: Select feature segments from different physical regions (e.g., distance greater than 10 mm) or with low spectral coherence (below 0.5) as negative sample pairs.

[0098] The sample pair set in the training set is constructed using the above method. (label), where label takes the value 1 to represent a positive pair and 0 to represent a negative pair.

[0099] Model M employs a two-branch convolutional neural network structure with shared weights. Each branch is responsible for feature encoding of the input vector F, outputting a 128-dimensional embedding vector. The network structure includes: an input layer that receives coherent feature vectors of dimension N from F′; an encoding layer containing three one-dimensional convolutional layers with kernel sizes of 3, 3, and 1, and channel numbers of 64, 128, and 128 respectively; a pooling and normalization layer using max pooling and batch normalization operations; and an output layer that outputs a 128-dimensional embedding vector through a fully connected layer. Training uses a contrastive loss function, which uses Euclidean distance or cosine similarity as the criterion: for positive sample pairs, it minimizes the distance between the two vectors; for negative sample pairs, it maximizes the distance while keeping it greater than a set safety margin (e.g., 1.0). Training uses the Adam optimizer with a learning rate of 0.001, and the number of training epochs is typically 50 to 100 depending on convergence.

[0100] After model training is complete, all high-dimensional feature matrices F to be analyzed are input into model M, and after encoding, a low-dimensional embedding vector Z is output. This vector is a 128-dimensional real number array, representing the clustering features of the current sample in the semantic space.

[0101] In a preferred embodiment of the present invention, to achieve accurate extraction and precise localization of weak defects, after obtaining the defect semantic embedding vector Z, it is necessary to perform cluster analysis and semantic boundary extraction on its distribution characteristics in the embedding space, and combine the time delay and energy information of the corresponding physical data in Q, and use a reverse beam focusing algorithm to perform three-dimensional localization of the defect reflection point. Specifically, the following steps are included:

[0102] Clustering is performed on the defect semantic embedding vectors Z of all samples using a density-based spatial clustering algorithm (DBSCAN). This algorithm clusters based on the number density of neighboring samples at each point, and can identify irregularly shaped and unknown cluster structures, making it particularly suitable for small anomaly clusters in sparse embedding spaces.

[0103] In DBSCAN, the core parameters are set as follows: the neighborhood radius ε is 0.3; the minimum number of sample points MinPts is 5; after the algorithm is executed, the embedding vectors marked as "boundary points" or "outliers" in the cluster labels will be regarded as potential weak defect signals; the clusters of these signals initially identified constitute a suspected defect candidate set, and proceed to the next step of semantic boundary analysis.

[0104] To further confirm the spatial clustering and boundary abruptness of the defect signals, this invention calculates the semantic boundary tensor change rate for the boundary region of each candidate cluster.

[0105] The tensor rate of change is defined as: the magnitude of the vector gradient change between a boundary point and its neighboring interior points in the embedding space, calculated using the L2 norm;

[0106] If the average tensor gradient magnitude of a boundary point in its local neighborhood is greater than 0.15 (a set threshold), it indicates that there is a significant semantic jump, reflecting abnormal information.

[0107] These boundary regions are further labeled as suspected weak defect regions R, and are mapped and traced using the original Q coordinates corresponding to their embedding vectors.

[0108] Extract the channel, angle, path, and time index corresponding to region R from Q, retrieve its reflected signal delay time and energy intensity information, and establish an inversion model for backfocusing. This process includes:

[0109] Construct a sound propagation path function S(i,j,k) to represent the sound wave propagation distance from the defect point to the i-th channel, j-th path, and k-th angle;

[0110] Using the known speed of sound of a material (e.g., 5900 meters per second for carbon steel), the time delay is inversely calculated into the spatial propagation distance;

[0111] Perform consistency matching on time delay data from multiple channels and angles, and remove abnormal channels with a time delay deviation of more than 1 microsecond;

[0112] The delay-and-sum beamforming algorithm is used to synchronously superimpose the echo signals of all valid paths, and the three-dimensional spatial coordinates of the reflection point are reconstructed based on energy weighting.

[0113] All suspected reflection points obtained after the above focusing process are aggregated according to their spatial coordinates (x, y, z) to generate a 3D point cloud map T. The specific processing steps are as follows: the spatial coordinates of each reflection point are obtained by jointly calculating the channel position, path angle, and inversion propagation distance; each point is assigned an intensity weight, and its focusing energy weighted average is taken; the point cloud is sparsely filled and filtered and smoothed using voxel resampling (e.g., 1 mm³ resolution) to remove isolated reflection points; the final output map T is the 3D fine-grained localization map of the defect point, containing the spatial location, boundary morphology, and energy distribution information of the defect. This map T will be used as input data for subsequent topology analysis and defect determination. As shown in Table 2, the impact of different feature combinations on the model embedding effect is as follows:

[0114] Table 2 shows the impact of different feature combinations on the model embedding performance (comparison of embedding vector separation), for example: Feature input combination Embedded vector clustering contour coefficients Inter-class distance ratio (average) Defect identification accuracy (%) Using only amplitude image features 0.42 1.3 82.7 Amplitude + Phase Coordination Factor (MCPCI) 0.57 1.8 87.1 Amplitude + Phase + MPWCI + CASER (This invention) 0.71 2.4 93.6

[0115] As shown in Table 2, the high-dimensional feature combination proposed in this invention can significantly improve the clustering separability of the semantic embedding model and enhance the model's ability to distinguish defect types under unlabeled conditions.

[0116] In a preferred embodiment of the present invention, to distinguish between genuine microcrack-like defects in the weld and artifact reflections caused by material structure complexity or signal scattering, it is necessary to perform topological structure modeling and boundary stability assessment on the suspected defect areas in the three-dimensional positioning image T, and then perform intelligent defect classification based on topological indices. The steps are as follows:

[0117] First, spatial clustering is performed on the reflection point cloud in the 3D fine-grained localization map T to identify multiple independent defect regions. This step uses a density-based spatial clustering algorithm (DBSCAN), which groups high-density regions into the same cluster by judging the local point density relationship. The cluster radius ε is set to 2 mm, and the minimum number of points MinPts is set to 10. Each set of points obtained after clustering constitutes a candidate defect region for subsequent topology analysis. This clustering method has the advantages of not requiring a preset number of clusters and automatically removing discrete noise points, and is suitable for defect structures of different scales.

[0118] To analyze the spatial topological complexity of the defect region, principal curvature analysis is performed on each candidate region: First, surface normal vectors are constructed based on the local neighborhood of the 3D point cloud; then, a principal curvature extraction algorithm (such as a PCA-based surface fitting method) is used to calculate the number of principal curvature directions per unit volume; if a local region has more than three significantly different principal curvature axis directions (with an angle greater than 30 degrees), the region is defined as having a bifurcation structure; the proportion of voxels exhibiting bifurcation in the entire point cloud region is statistically analyzed and defined as the shape bifurcation index, with a value ranging from 0 to 1. A larger value indicates a more complex structure and more branches. For example, if 40% of the voxels in a defect region exhibit multiple principal axis abrupt changes, the bifurcation index for that region is 0.4.

[0119] To analyze the stability of the defect structure's boundary and further evaluate the consistency of its reflection behavior, multi-angle cross-sectional slicing was performed on each defect region: the 3D point cloud region was sliced ​​in horizontal, vertical, and oblique directions, generating a 2D cross-section every 1 mm; the boundary profile was extracted for each slice, and the boundary fluctuation coefficient, i.e., the degree of variation of the boundary shape profile at multiple angles, was calculated; the fluctuation coefficient is defined as the standard deviation of the distance from the boundary point to its center point in each slice divided by the average distance, and the lower the value, the more stable the boundary; if the boundary fluctuation coefficient of more than half of the slices is below 0.2, the defect region boundary is considered to have good stability. This index is used to exclude non-realistic defects caused by stray echoes or surface structures.

[0120] The two structural indices mentioned above are input into a trained Support Vector Machine (SVM) classification model to distinguish between real microcracks and material artifacts. The model uses the shape bifurcation index and structural boundary stability as two-dimensional input vectors; the SVM employs a radial basis function (RBF) kernel function and is trained under supervision using historically labeled defect samples; the classification results are output as two-class labels: "real microcracks" or "structural artifacts," along with confidence scores; the final defect analysis results include defect type, location, and structural feature descriptions.

[0121] Example 2: To verify the effectiveness of the intelligent analysis method for weld ultrasonic phased array detection data described in this invention in the identification of complex defects in practice, a weld test block containing multiple types of typical defects was selected for experimentation. Seven representative reflection sources were set up on the test block, covering typical defects such as microcracks, porosity, inclusions, and incomplete penetration, as well as structural artifact reflection interference, to evaluate the identification performance of this invention under high noise and high interference backgrounds.

[0122] The detection system and parameter configuration are as follows:

[0123] Testing equipment: OmniScan X3 phased array system;

[0124] Probe model: 5L64-A31, frequency 5 MHz, 64 elements, element spacing 0.6 mm;

[0125] Scanning method: Linear scan + angular sector scan Step length );

[0126] Sampling frequency: 20 MHz, sampling points: 512 points / channel;

[0127] Data structure: Form a four-dimensional wave field tensor Q, with dimensions of time × channel × path × incident angle;

[0128] Wave packet window: ±5µs, spectral analysis range 1–8 MHz.

[0129] Table 3. Defect / Interference Distribution Table, for example: Defect number Defect types Preset feature description D1 Microcrack (0.6 mm) The crack is shallow and highly sensitive to angles. D2 pores Aperture approximately 0.8 mm, weak echo. D3 Unfused The defects extend linearly, and the wavefront is complex. D4 Material inclusions Highly random and unstable signal D5 Grain boundary interference Strong interference from multiple angles, unstructured reflections D6 Weld root artifacts High-frequency reflection intensity is high but discontinuous D7 Microcrack (1.2 mm) The signal is weak but highly coherent.

[0130] As shown in Table 3, consistent with the original description, the process includes: constructing a four-dimensional tensor Q (time, channel, path, angle); extracting three types of features to construct a high-dimensional matrix F: multi-path wavefront coupling degree (MPWCI), multi-channel phase synergy factor (MCPCI), and cross-angle spectral entropy ratio (CASER); inputting F into a self-supervised semantic embedding model M to generate a defect semantic embedding vector Z; generating a localization map through clustering and back-focusing, further calculating topological structure indicators, and determining the defect type.

[0131] Table 4. Results Statistics Table, for example: Defect number Actual type MPWCI MCPCI CASER Embedded classification results Determining if it is a real defect 3D positioning error (mm) D1 microcracks 0.83 0.74 0.21 Microcrack-like defects yes 0.6 D2 pores 0.76 0.68 0.25 Porous defects yes 0.9 D3 Unfused 0.79 0.71 0.18 Incomplete fusion defects yes 1.1 D4 Material inclusions 0.43 0.39 0.46 Material artifacts no - D5 Grain boundary reflection 0.49 0.44 0.52 Material artifacts no - D6 Root structure artifacts 0.52 0.48 0.41 Non-defect area no - D7 microcracks 0.81 0.72 0.23 Microcrack-like defects yes 0.7

[0132] As shown in Table 4, the identification accuracy is as follows: 6 / 7 defects were accurately identified (accuracy of approximately 85.7%, 100% after artifact removal); the artifact removal rate is as follows: 3 / 3 of material or structural artifacts were not misidentified as defects (false positive rate of 0); the mean positioning error is approximately 0.83 mm, which meets the engineering requirements (≤1.5 mm); and the clustering vector separation (embedded visualization results) is significantly better than the method without using high-dimensional features.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. An intelligent analysis system for ultrasonic phased array inspection data of welds, characterized in that: include: Data acquisition module: Acquires the raw dataset D of ultrasonic phased array detection of the weld to be tested; Spatiotemporal reconstruction module: Performs spatiotemporal joint reconstruction processing on D to construct a four-dimensional wave field tensor Q with dynamic propagation path reversibility, including time, channel position, incident angle and scan path; Feature extraction module: Extracts the reflection group signals in Q that still maintain temporal coherence during discontinuous path propagation, and constructs a high-dimensional coherent feature matrix F, which includes multi-path wavefront coupling degree, multi-channel phase coordination factor and cross-angle spectral entropy ratio; Defect representation model training module: Input F into the contrast feature embedding model M trained by self-supervised learning to generate defect semantic embedding vector Z. The model M achieves feature clustering learning by automatically constructing positive and negative sample pairs. 3D localization module: Based on the distribution density of Z and the trend of semantic boundary tensor change, identify suspected weak defect regions R, and combine the delay time and energy loss of the corresponding position in Q to perform reverse beam focusing calculation to obtain a 3D fine localization map T of the defect point. Evaluation module: Performs in-depth topological analysis on the defect region in T, calculates its shape bifurcation index and structural boundary stability to distinguish between material structure artifacts and real microcrack-like defects, and outputs the final defect analysis results.

2. The intelligent analysis system for ultrasonic phased array inspection data of welds according to claim 1, characterized in that: The dataset D contains multiple channels, multiple scan paths, and full waveform echo signals at multiple incident angles.

3. The intelligent analysis system for ultrasonic phased array inspection data of welds according to claim 1, characterized in that: The spatiotemporal joint reconstruction process of D includes: The echo signals from each channel are processed uniformly along the time axis. Phase alignment is performed on each aligned channel signal; Signals from multiple scanning paths are uniformly mapped to an equally spaced three-dimensional grid according to their actual physical coordinates. Sparse regions are filled by weighted averaging of neighboring measurement points, thus constructing a wavefield dataset with spatial continuity. Organize time-domain aligned, phase-consistent, and spatially continuous data into a unified four-dimensional wavefield tensor Q.

4. The intelligent analysis system for ultrasonic phased array inspection data of welds according to claim 3, characterized in that: The extraction of reflection group signals in Q that retain temporal coherence during discontinuous path propagation includes: Time-domain windowing is performed on the wavefield slices corresponding to each incident angle in the four-dimensional wavefield tensor Q to extract the main energy segment of the wave packet. Calculate the temporal coherence coefficient of wave packets between different scanning paths, and use a correlation threshold of 0.8 as the standard to screen out echo signals with cross-path coherence; Based on the phase stability and energy distribution continuity of coherent signals, multi-scale wavelet packet decomposition is used to extract their multipath wavefront coupling characteristics. The multipath wavefront coupling features are reorganized according to the channel, path and angle dimensions to construct a high-dimensional coherent feature matrix F.

5. The intelligent analysis system for ultrasonic phased array inspection data of welds according to claim 4, characterized in that: The methods for obtaining the multipath wavefront coupling degree, multichannel phase coordination factor, and cross-angle spectral entropy ratio include: To obtain the multipath wavefront coupling degree, the normalized cross-correlation function is calculated in units of sampled signals within the time window for the selected cross-path coherent echoes. The peak cross-correlation value and peak delay are calculated between each pair of paths. The peak cross-correlation values ​​of all path pairs are summed according to the path logarithm and normalized by the path logarithm to obtain the multipath wavefront coupling degree. To obtain the multi-channel phase coordination factor, the Hilbert transform of each channel signal at the same spatiotemporal point is first performed to obtain the instantaneous phase. The phase lock value (PLV) of the phase difference sequence between any two channels within the time window is calculated, and the average value of all channels relative to the PLV is used as the multi-channel phase coordination factor. The PLV calculation adopts the mode length of the average of the complex phase vectors. To obtain the cross-angle spectral entropy ratio, the power spectral density is estimated using the Welch method for the selected time-domain wave packet at each incident angle, the spectral entropy of the power spectrum is calculated, and the spectral entropy within the target frequency band is compared with the spectral entropy of all frequency bands to obtain the spectral entropy ratio for each pair of angles. Finally, the median of the angle-to-spectral entropy ratio is taken as the cross-angle spectral entropy ratio.

6. The intelligent analysis system for ultrasonic phased array inspection data of welds according to claim 1, characterized in that: The step of inputting F into the contrastive feature embedding model M trained by self-supervised learning to generate the defect semantic embedding vector Z includes: The high-dimensional feature matrix F is standardized to generate a multi-view input sample set; Positive and negative sample pairs are automatically constructed based on unlabeled data. Positive sample pairs consist of signal segments from the same physical region or with similar feature distributions, while negative sample pairs consist of signals from different spatial regions or with low coherence. The sample pairs are input into the contrastive feature embedding model M, which contains a shared encoder. The model M adopts a two-branch convolutional neural network structure and is trained to minimize the contrastive loss through feature cosine similarity metric. After the model converges, the encoded F is embedded and mapped to obtain a low-dimensional defect semantic embedding vector Z, which represents the spatiotemporal coherent semantic features of weld defects.

7. The intelligent analysis system for ultrasonic phased array inspection data of welds according to claim 6, characterized in that: The obtained three-dimensional fine localization map T of the defect point includes: Density clustering analysis was performed on the defect semantic embedding vector Z, and an adaptive clustering algorithm based on local reachability density was used to identify abnormally sparse distribution areas in order to preliminarily determine the suspected defect feature clusters. Calculate the semantic boundary tensor gradient change rate for each feature cluster boundary region. When the change rate exceeds a preset threshold, the region is determined to have a significant semantic mutation and is marked as a suspected defect region R. Retrieve the spatiotemporal coordinate data corresponding to R in Q, extract the delay time and reflection energy distribution, and perform reverse beam focusing calculation on the signals of each channel based on the time delay inversion algorithm; By superimposing multiple channels and using energy-weighted averaging, a defect location point cloud in three-dimensional spatial coordinates is constructed, forming a three-dimensional fine location map T of the defect point.

8. The intelligent analysis system for ultrasonic phased array inspection data of welds according to claim 7, characterized in that: The process of performing deep topological analysis on the defect region in T, calculating its shape bifurcation index and structural boundary stability, includes: Spatial segmentation is performed on each suspected defect point cloud region in the 3D localization map T, and a density-based clustering algorithm is used to identify independent defect bodies to generate a set of candidate defect regions. For each candidate region, the shape bifurcation index is calculated, and the three-dimensional contour change rate is extracted based on the principal curvature analysis of the point cloud. When there are more than three abrupt changes in the direction of the principal curvature axes within a unit volume, it is judged as a complex topological bifurcation structure. The candidate defect boundary is sliced ​​at multiple angles, and the boundary fluctuation coefficient of each slice is calculated. If the boundary shape remains stable in multiple angle slices, the structural boundary is considered to have high stability. By combining the bifurcation index and boundary stability index, a support vector machine model is used to classify artifacts and microcracks, and the final defect analysis results are output.

Images