Aluminum alloy welded pipe weld ultrasonic flaw detection defect recognition method based on deep learning

Abstract

The present application relates to the technical field of welding detection, and discloses an aluminum alloy welded pipe weld ultrasonic flaw detection defect identification method based on deep learning, which synchronously collects aluminum alloy welded pipe weld ultrasonic echoes through a multi-channel phased array probe, splices a full circumferential echo signal matrix after pretreatment; detects a weld edge and fits a fusion line to generate a mask to extract a region of interest sub-matrix by utilizing acoustic impedance difference; splices a joint feature tensor by performing short-time Fourier transform, wavelet packet decomposition and Hilbert transform on each channel signal; obtains defect sensitive deep features by weighting enhancement through a cascaded attention network; inputs a multi-task network in parallel to output defect type, position and severity results; maps to a three-dimensional cylindrical coordinate system to generate a defect atlas and superimposes the defect atlas on a B-scan image; compares through a three-dimensional threshold matrix, calculates a quality score by weighting a hazard, and automatically outputs a weld quality classification result and a detection report.

Description

Technical Field

[0001] This invention relates to the field of welding inspection technology, and in particular to a method for identifying defects in weld seams of aluminum alloy welded pipes using ultrasonic testing based on deep learning. Background Technology

[0002] Aluminum alloy welded pipes are widely used in aerospace, petrochemical, marine engineering, and rail transportation due to their excellent corrosion resistance, good thermal conductivity, and high specific strength. Weld quality is a key factor determining the service performance and safety of aluminum alloy welded pipes. Internal weld defects such as porosity, slag inclusions, incomplete penetration, lack of fusion, and cracks can significantly reduce the load-bearing capacity and fatigue life of the welded pipe, and may even lead to catastrophic failures. Ultrasonic non-destructive testing technology has become the mainstream method for inspecting weld quality in aluminum alloy welded pipes due to its advantages such as high detection sensitivity, strong penetration, harmlessness to humans, and the ability to perform online testing.

[0003] However, existing technologies have the following drawbacks: Most current deep learning methods are designed for imaging detection data such as X-rays and DR digital X-rays, whose input is two-dimensional image data. Ultrasonic testing, on the other hand, acquires one-dimensional time-series echo signals. These two methods differ fundamentally in data format, information dimension, and noise characteristics. Directly transferring deep learning models from the image domain to the ultrasonic signal domain fails to effectively capture the time-frequency coupled defect features in the ultrasonic echo signal, resulting in low defect identification accuracy and a high confusion rate between defects such as porosity and inclusions. Aluminum alloy welded pipe welds have unique material properties and geometric structures. Aluminum alloys have low acoustic impedance, coarse grains, and significant anisotropy, causing severe attenuation of ultrasonic signals during propagation and susceptibility to grain scattering noise. The curvature of the welded pipe leads to a complex ultrasonic beam propagation path, resulting in severe geometric diffusion and mode conversion interference in the echo signal. Existing methods do not fully consider these special characteristics of aluminum alloy welded pipe welds and lack targeted signal preprocessing and feature extraction strategies, making it difficult to effectively distinguish between defect echoes and false defect signals, resulting in high false negative and false positive rates. Existing methods mostly use single-channel or single-view ultrasonic data for defect identification, failing to fully utilize the spatial complementary information acquired by multi-channel phased array probe arrays, and thus cannot accurately locate the defect in the three-dimensional space of the weld. Furthermore, existing methods typically only output defect type classification results, lacking the ability to comprehensively assess defect severity and overall weld quality, failing to meet the practical needs of quality grading and evaluation for aluminum alloy welded pipes. Existing deep learning models are mostly trained offline; once deployed, their parameters are fixed, making adaptive optimization impossible based on newly emerging defect types or signal characteristic changes during on-site inspection. When welding process parameters are adjusted, base material batches are changed, or the inspection environment changes, the model's generalization ability decreases, and its recognition performance significantly degrades. This necessitates manual data collection and model retraining, resulting in high maintenance costs and poor timeliness. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying defects in the weld seam of aluminum alloy welded pipes using ultrasonic testing based on deep learning. This method solves the problems in the prior art, such as model transfer failure caused by the difference in the form of ultrasonic echo signals and image data, severe signal attenuation and false defect interference caused by the material properties of aluminum alloy and the geometric structure of welded pipes, lack of spatial positioning capability in single-channel single-view detection, and inability of the model to adaptively optimize.

[0005] To achieve the above objectives, this invention provides a method for identifying defects in the weld seams of aluminum alloy welded pipes using ultrasonic testing based on deep learning, comprising the following steps: A circumferential multi-channel phased array probe synchronously acquires ultrasonic echoes from aluminum alloy welded pipe welds. After noise reduction, correction, and gain compensation, the echoes are spliced ​​together in time and space to obtain a full circumferential echo signal matrix. Based on acoustic impedance difference detection of weld edge and fitting of fusion line, weld region is divided by combining groove parameters, generating mask and dot product with original matrix to extract sub-matrix of region of interest; Short-time Fourier transform is performed on the signals of each channel of the submatrix to obtain the time spectrum. Wavelet packet decomposition is used to extract the frequency band energy / entropy, Hilbert transform is used to extract the time domain envelope, and multi-dimensional features are spliced ​​into a joint feature tensor. The joint feature tensor is input into a cascaded network with channel and spatial attention for weighted enhancement, and then connected to the original feature residual to obtain defect-sensitive depth features; This feature is input into a multi-task network that combines classification, location regression, and severity assessment branches in parallel. After training with a joint loss, the network outputs the defect type, location, and severity results. The detection results are mapped to the three-dimensional cylindrical coordinate system of the weld, and a defect map is generated by axial bicubic spline interpolation. The map is then superimposed with the original B-scan image to obtain a comprehensive detection map. The defect information is compared with the three-dimensional judgment threshold, and the comprehensive quality score is calculated by weighting the severity. After matching the level threshold, the weld quality classification result and automatic inspection report are output.

[0006] The step of "synchronously acquiring ultrasonic echoes from aluminum alloy welded pipe welds using a circumferential multi-channel phased array probe, and then spatiotemporally aligning and stitching them after noise reduction, correction, and gain compensation to obtain a full-circumferential echo signal matrix" includes the following steps: By uniformly arranging N ultrasonic phased array probes along the circumference of the weld seam of the aluminum alloy welded pipe, the multi-channel ultrasonic echo signal of the weld seam is simultaneously triggered and collected to obtain the original multi-channel ultrasonic echo signal sequence. After bandpass filtering and baseline drift correction, the original multi-channel ultrasonic echo signal sequence is used to establish a geometric model of the ultrasonic beam propagation path based on the outer diameter, wall thickness and incident angle of the aluminum alloy welded pipe. The sound path attenuation compensation coefficient is calculated to perform point-by-point gain compensation, and the gain-compensated ultrasonic echo signal is obtained. The multi-channel ultrasonic echo signals after gain compensation are spatiotemporally aligned and spliced ​​according to the circumferential position of the weld to obtain the full circumferential ultrasonic echo signal matrix of the weld.

[0007] The step of "detecting weld edges based on acoustic impedance differences, fitting fusion lines, dividing weld regions by combining bevel parameters, generating a mask and multiplying it with the original matrix to extract the sub-matrix of the region of interest" includes the following steps: The ultrasonic echo signal matrix of the weld circumference is used to extract the edge point set of the weld contour by taking advantage of the abrupt change in echo amplitude caused by the difference in acoustic impedance between the base material area, the heat-affected zone and the weld area. The weld fusion line reference curve is obtained by fitting the edge point set of the weld contour using the edge detection algorithm and the edge detection algorithm. By combining the weld fusion line reference curve with the groove geometry parameters, the heat-affected zone boundary is determined by expanding outward by a preset distance to both sides and the weld center zone boundary is determined by contracting inward by a preset distance, thus generating a multi-region boundary mask for the weld. The ultrasonic echo signal submatrix of the region of interest in the weld is obtained by performing point-by-point multiplication of the multi-region boundary mask of the weld with the ultrasonic echo signal matrix of the entire circumference of the weld.

[0008] The step of “performing short-time Fourier transform on each channel signal of the sub-matrix to obtain the time spectrum, extracting frequency band energy / entropy through wavelet packet decomposition, extracting the time domain envelope through Hilbert transform, and concatenating multi-dimensional features into a joint feature tensor” includes the following steps: The time-frequency spectrum is obtained by performing short-time Fourier transform on each channel signal in the ultrasonic echo signal submatrix of the region of interest of the weld. Multi-scale wavelet packet decomposition is performed on the time-frequency spectrum to extract the energy and energy entropy features of each frequency band. At the same time, Hilbert transform is performed on the original signal to extract the time-domain envelope morphology features. By splicing and fusing the energy distribution features of the time-frequency spectrogram, the energy / entropy features of the wavelet packet, and the temporal morphological features in the channel dimension, a multi-scale time-frequency domain joint feature tensor is obtained.

[0009] The step of "weighting and enhancing the joint feature tensor input into a cascaded network with channel and spatial attention, and connecting it with the original feature residual to obtain defect-sensitive depth features" includes the following steps: The multi-scale time-frequency domain joint feature tensor is input into a cascaded attention network containing channel attention modules and spatial attention modules. Adaptive weighting is applied to the channel and spatial dimensions of the feature tensor to obtain the attention-weighted deep feature tensor. By performing a residual connection between the attention-weighted depth feature tensor and the original multi-scale time-frequency domain joint feature tensor, a defect-sensitive depth feature representation is obtained.

[0010] The step of "inputting the feature into a multi-task network that performs classification, location regression, and severity assessment branches in parallel, training it with a joint loss, and then outputting the defect type, location, and severity results" includes the following steps: The defect-sensitive deep feature representation is input in parallel into the defect type classification branch, the defect location regression branch, and the defect severity assessment branch to construct a multi-task learning network. The multi-task joint loss function is used for end-to-end collaborative training to obtain the multi-task defect recognition network model after training convergence. The defect-sensitive depth feature representation of the weld to be inspected is input into the multi-task defect recognition network model after training and convergence to obtain the defect type probability distribution, defect location coordinate offset and confidence level, and defect severity level score.

[0011] The step of "mapping the detection results to the three-dimensional cylindrical coordinate system of the weld, generating a defect map through axial bicubic spline interpolation, and overlaying it with the original B-scan image to obtain a comprehensive detection map" includes the following steps: The probability distribution of defect type, the coordinate offset and confidence level of defect location, and the score of defect severity are mapped to the three-dimensional cylindrical coordinate system of weld. Bicubic spline interpolation is performed along the weld pipe axis to generate a two-dimensional defect distribution map of weld axis-circumference. The defect distribution map is used to identify defect types with different colors, severity with color intensity, and confidence level with marker size. It is then overlaid and fused with the original ultrasound B-scan image to obtain a comprehensive detection map that includes the original signal and intelligent recognition results.

[0012] This invention presents a deep learning-based ultrasonic flaw detection defect identification method for aluminum alloy welded pipes. The method synchronously acquires circumferential ultrasonic echo signals from the weld seam of the aluminum alloy welded pipe using a multi-channel ultrasonic phased array probe. After bandpass filtering for noise reduction, baseline drift correction, and path attenuation gain compensation, the signals are spatiotemporally aligned and stitched together, effectively eliminating signal attenuation differences caused by the anisotropy of the aluminum alloy material and the curvature of the welded pipe, resulting in a circumferential ultrasonic echo signal matrix. The method utilizes acoustic impedance differences to detect weld edges and fits the fusion line. Combined with bevel parameters, the weld region is adaptively divided to generate a mask for extracting the region of interest (ROI) sub-matrix, accurately eliminating false defect interference from the base material region and the heat-affected zone. Short-time Fourier transform, multi-scale wavelet packet decomposition, and Hilbert transform are performed on the signals of each channel of the sub-matrix. The time-frequency domain features, wavelet packet energy entropy features, and time-domain morphological features are stitched and fused into a joint feature tensor, which fully... The defect feature information of time-frequency coupling in ultrasonic echo signals was mined; defect-sensitive depth features were obtained by adaptive weighting enhancement through a cascaded attention network and connecting it with the original feature residuals, and the defect-sensitive channels and spatial locations were adaptively focused; this feature was input into a multi-task network of classification, location regression and severity assessment branches for collaborative training, and the results of defect type, location and severity were output, realizing the integrated output of defect identification, localization and assessment; the detection results were mapped to a three-dimensional cylindrical coordinate system to generate a defect map and superimposed and fused with B-scan images to intuitively display the spatial distribution of defects; finally, by comparing the three-dimensional judgment threshold matrix, calculating the comprehensive quality score by hazard weighting and matching the level threshold, the weld quality grading results and inspection report were automatically output, realizing intelligent, high-precision and fully automated ultrasonic flaw detection defect identification of aluminum alloy welded pipe welds. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0014] Figure 1 This is a flowchart of the ultrasonic flaw detection defect identification method for aluminum alloy welded pipes based on deep learning provided by the present invention.

[0015] Figure 2 This is a flowchart for obtaining the full-circuit echo signal matrix.

[0016] Figure 3 This is a flowchart for extracting a submatrix of the region of interest.

[0017] Figure 4 This is a flowchart of concatenating multi-dimensional features into a joint feature tensor.

[0018] Figure 5 This is a flowchart for obtaining defect-sensitive depth features.

[0019] Figure 6It is a flowchart that outputs the defect type, location, and severity results.

[0020] Figure 7 This is a flowchart of the process of obtaining a comprehensive detection map by overlaying the original B-scan image.

[0021] Figure 8 This is a flowchart showing the output of weld quality grading results and automatic inspection reports after matching the grade threshold. Detailed Implementation

[0022] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0023] Please see Figures 1 to 8 This invention provides a method for identifying defects in the weld seam of aluminum alloy welded pipes using ultrasonic testing based on deep learning, comprising the following steps: The S1 circumferential multi-channel phased array probe synchronously acquires ultrasonic echoes from aluminum alloy welded pipe welds. After noise reduction, correction, and gain compensation, the echoes are spatiotemporally aligned and stitched together to obtain a full circumferential echo signal matrix. S11 uses N ultrasonic phased array probes evenly arranged along the circumference of the weld seam of the aluminum alloy welded pipe to synchronously trigger the acquisition of multi-channel ultrasonic echo signals in the full circumference of the weld seam, and obtains the original multi-channel ultrasonic echo signal sequence. S12 performs bandpass filtering noise reduction and baseline drift correction on the original multi-channel ultrasonic echo signal sequence. Based on the outer diameter, wall thickness and incident angle of the aluminum alloy welded pipe, a geometric model of the ultrasonic beam propagation path is established. The sound path attenuation compensation coefficient is calculated to perform point-by-point gain compensation, and the gain-compensated ultrasonic echo signal is obtained. S13 performs spatiotemporal alignment and channel splicing of the multi-channel ultrasonic echo signals after gain compensation according to the circumferential position of the weld to obtain the full circumferential ultrasonic echo signal matrix of the weld.

[0024] Specifically, N ultrasonic phased array probes are evenly arranged along the circumference of the weld seam of the aluminum alloy welded pipe. N is a positive integer not less than 2, determined according to the outer diameter of the welded pipe and the detection accuracy requirements. Preferably, when the outer diameter of the welded pipe is 100mm to 300mm, N is 8. Eight ultrasonic phased array probes are evenly arranged along the circumference of the weld seam of the aluminum alloy welded pipe. The probe center frequency is 5MHz, the crystal has 64 array elements, and the probes are synchronously triggered to collect multi-channel ultrasonic echo signals of the entire circumference of the weld seam. The collection frequency is 100MHz, and the number of sampling points is 4096, resulting in an 8-channel original multi-channel ultrasonic echo signal sequence. Then, the original multi-channel ultrasonic echo signal sequence was subjected to frequency domain noise reduction using a bandpass filter with a center frequency of 5MHz and a bandwidth of 2MHz. A 5-layer wavelet thresholding method based on the db4 wavelet basis function was used for noise reduction, with the threshold selected using the SureShrink adaptive thresholding rule. Adaptive median filtering was also employed to remove impulse noise. The filter window size was dynamically adjusted based on the signal's local variance, ranging from 3×3 to 7×7. After baseline drift correction, a geometric model of the ultrasonic beam propagation path was established based on the aluminum alloy welded pipe's outer diameter of φ219mm, wall thickness of 8mm, and probe incident angle of 60°. Point-by-point gain compensation was performed on the ultrasonic echo signal amplitude to obtain the gain-compensated ultrasonic echo signal. Finally, the gain-compensated 8-channel ultrasonic echo signal was spatiotemporally aligned and spliced ​​at 45° intervals along the weld circumference to obtain an 8×4096-dimensional weld circumferential ultrasonic echo signal matrix.

[0025] S2 detects weld edges and fits fusion lines based on acoustic impedance differences, divides weld regions by combining bevel parameters, generates a mask and multiplies it with the original matrix, and extracts the sub-matrix of the region of interest. S21 utilizes the acoustic impedance differences between the base material region, heat-affected zone, and weld area to extract the weld contour edge point set using an edge detection algorithm, and then obtains the weld fusion line reference curve by fitting the edge point set using the least squares method. S22 combines the weld fusion line reference curve with the groove geometry parameters, expands the heat-affected zone boundary to both sides by a preset distance, and contracts the weld center zone boundary to the inside by a preset distance, thus generating a multi-region boundary mask for the weld. The boundary of the heat-affected zone is determined by expanding outwards to both sides by a preset distance, which is 2-5mm and is determined according to the weld width and the range of the heat-affected zone; the boundary of the weld center zone is determined by contracting inwards by a preset distance, which is 1-3mm and is determined according to the weld weld width. S23 performs point-by-point multiplication of the multi-region boundary mask of the weld with the ultrasonic echo signal matrix of the entire circumference of the weld to obtain the ultrasonic echo signal submatrix of the region of interest of the weld.

[0026] Specifically, firstly, the ultrasonic echo signal matrix of the weld seam in the entire circumference is utilized using the acoustic impedance of the base material region, which is approximately 17.5 × 10⁻⁶. 6kg / m²s, acoustic impedance in the heat-affected zone is approximately 16.8 × 10⁻⁶ kg / m²s. 6 The acoustic impedance of the weld zone is approximately 15.2 × 10⁻⁶ kg / m²s. 6 The abrupt change in echo amplitude caused by the acoustic impedance difference between kg / m²s was investigated. The Canny edge detection algorithm was used to extract the weld contour edge point set. A Gaussian filter kernel with a standard deviation σ=1.4 and a Sobel operator convolution kernel size of 3×3 were used. The high-low threshold ratio was 3:1, with the low threshold set to 10% of the maximum signal amplitude and the high threshold set to 30%. Based on the extracted weld contour edge point set {(x_i, y_i)}, a quadratic curve was fitted using the least squares method as the weld fusion line reference curve, with a fitting error threshold of 0.5mm. Then, combining the weld fusion line reference curve with the V-shaped or U-shaped bevel geometry parameters of the aluminum alloy welded pipe weld, the heat-affected zone boundary was determined by expanding outwards by 3mm on both sides and contracting inwards by 2mm to determine the weld center zone boundary, generating a multi-region boundary mask for the weld. Finally, the multi-region boundary mask of the weld and the ultrasonic echo signal matrix of the weld's full circumference are multiplied point by point to obtain an ultrasonic echo signal submatrix of the region of interest of the weld with a size of 8×2048.

[0027] S3 performs short-time Fourier transform on the signals of each channel of the submatrix to obtain the time spectrum, extracts the frequency band energy / entropy by wavelet packet decomposition, extracts the time domain envelope by Hilbert transform, and splices the multi-dimensional features into a joint feature tensor. S31 performs short-time Fourier transform on each channel signal in the ultrasonic echo signal submatrix of the region of interest in the weld to obtain the time spectrum. Multi-scale wavelet packet decomposition is performed on the time spectrum to extract the energy and energy entropy features of each frequency band. At the same time, Hilbert transform is performed on the original signal to extract the time-domain envelope morphology features. S32 splices and fuses the energy distribution features of the time-frequency spectrum, the energy / entropy features of the wavelet packet, and the temporal morphological features in the channel dimension to obtain a multi-scale time-frequency domain joint feature tensor.

[0028] Specifically, firstly, short-time Fourier transforms are performed on the signals of each channel in the ultrasonic echo signal submatrix of the region of interest in the weld seam. A Hanning window with a window length of 256 points and an overlap rate of 50% is used to obtain the time-frequency spectrum of each channel signal, with a size of 128×128. The time-frequency spectrum is then decomposed into three levels of wavelet packets using the sym8 wavelet basis function. Each level divides the current frequency band in half, resulting in 2³=8 wavelet packet coefficients across the three levels. The energy proportion and energy entropy of the wavelet packet coefficients in each frequency band are calculated, constructing an 8-dimensional wavelet packet energy feature vector [E_1, E_2, ..., E_8] and an 8-dimensional wavelet packet energy entropy feature vector [H_1, H_2, ..., H_8]. These are then concatenated to form a 16-dimensional multi-scale wavelet packet joint feature vector. Simultaneously, a Hilbert transform is performed on the original ultrasonic echo signal to obtain the envelope of the analytic signal. Temporal morphological features such as envelope peak value, envelope width, and envelope area are extracted to form a 3-dimensional temporal morphological feature vector. Then, the energy distribution features of the time-frequency spectrum (128×128), the 16-dimensional joint feature vector of the multi-scale wavelet packet, and the 3-dimensional morphological feature vector of the time domain are spliced ​​and fused in the channel dimension to obtain a multi-scale time-frequency domain joint feature tensor with a size of 128×128×19.

[0029] S4 inputs the joint feature tensor into a cascaded network with channel and spatial attention, weights and enhances it, and connects it with the original feature residual to obtain defect-sensitive depth features; S41 inputs the multi-scale time-frequency domain joint feature tensor into a cascaded attention network containing channel attention modules and spatial attention modules, and performs adaptive weighting enhancement on the channel dimension and spatial dimension of the feature tensor to obtain the attention-weighted deep feature tensor. S42 performs a residual connection between the attention-weighted depth feature tensor and the original multi-scale time-frequency domain joint feature tensor to obtain a defect-sensitive depth feature representation.

[0030] Specifically, firstly, a multi-scale time-frequency domain joint feature tensor of size 128×128×19 is input into a cascaded attention network containing a channel attention module and a spatial attention module. The channel attention module contains two parallel global pooling branches: a global average pooling branch performs average pooling on the input feature tensor F∈R^(H×W×C) along the spatial dimension to generate a 1×1×19-dimensional channel descriptor z_avg, and a global max pooling branch performs max pooling along the spatial dimension to generate a 1×1×19-dimensional channel descriptor z_max. The two descriptors share a two-layer multilayer perceptron. The number of neurons in the first layer is determined by reducing the dimensionality of the input channel number C according to the reduction rate r (r is 16), and the activation function is ReLU. The number of neurons in the second layer is restored to the input channel number C. The output is summed element-wise and then activated by the Sigmoid activation function to generate a channel weight vector M_c∈R^(1×1×C), which is adaptively weighted according to the channel dimension of the feature tensor. The spatial attention module performs max pooling and average pooling along the channel dimension on the channel-weighted feature tensor, generating two 128×128×1-dimensional feature maps. These are then concatenated along the channel dimension to form a 128×128×2-dimensional feature map. This map is then passed through a 7×7 convolutional layer with a stride of 1, padding of 3, and an output channel of 1, and activated by a sigmoid activation function to generate a 128×128-dimensional spatial attention map M_s∈R^(128×128). The spatial dimension of the feature tensor is adaptively weighted to obtain the attention-weighted depth feature tensor. Finally, the attention-weighted depth feature tensor is joined element-wise with the original multi-scale time-frequency domain joint feature tensor to output a defect-sensitive depth feature representation of size 128×128×19.

[0031] S5 inputs this feature into a multi-task network that combines classification, location regression, and severity assessment branches. After training with the joint loss, it outputs the defect type, location, and severity results. S51 uses the defect-sensitive deep feature representation to input the defect type classification branch, defect location regression branch and defect severity assessment branch in parallel to construct a multi-task learning network. It then uses a multi-task joint loss function for end-to-end collaborative training to obtain a multi-task defect recognition network model after training convergence. S52 inputs the defect-sensitive depth feature representation of the weld to be inspected into the multi-task defect recognition network model after training and convergence, and obtains the defect type probability distribution, defect location coordinate offset and confidence level, and defect severity level score.

[0032] Specifically, firstly, a multi-task learning network is constructed by inputting a 128×128×19 dimension defect-sensitive depth feature into three branches in parallel. The defect type classification branch employs a three-layer perceptron structure with 512, 256, and 128 neurons in the hidden layers, using ReLU activation, and a 5-dimensional Softmax output layer. It outputs the probability distribution of weld defects belonging to five types: porosity, slag inclusion, incomplete penetration, lack of fusion, or crack. The defect location regression branch uses a fully convolutional network structure with three convolutional layers having kernel sizes of 3×3, 3×3, and 1×1, and channels of 256, 128, and 4, respectively. It outputs the coordinate offset and confidence score of the defect in the weld depth and circumferential directions. The defect severity assessment branch uses an attention-weighted time-series aggregation network with a bidirectional LSTM hidden layer of 128 dimensions and 4 attention heads. It integrates the defect echo amplitude, duration, and diffusion range to output a defect severity rating in the range of 0 to 1. An end-to-end collaborative training method using a multi-task joint loss function and the Adam optimizer was employed. After 200 epochs of training, a converged multi-task defect recognition network model was obtained. Then, the defect-sensitive depth feature representation of the weld to be detected was input into the converged multi-task defect recognition network model to obtain the defect type probability distribution, defect location coordinate offset and confidence level, and defect severity level score.

[0033] S6 maps the detection results to the three-dimensional cylindrical coordinate system of the weld, generates a defect map by axial bicubic spline interpolation, and superimposes it with the original B-scan image to obtain a comprehensive detection map. S61 maps the probability distribution of defect types, the coordinate offset and confidence level of defect location, and the severity level score of defect to the three-dimensional cylindrical coordinate system of the weld, and performs bicubic spline interpolation and splicing along the weld pipe axis to generate a two-dimensional defect distribution map of the weld axial-circumferential direction. S62 uses different colors to identify defect types, color intensity to indicate severity, and marker size to indicate confidence level in the defect distribution map. It then overlays and fuses the map with the original ultrasound B-scan image to obtain a comprehensive detection map that includes the original signal and intelligent recognition results.

[0034] Specifically, firstly, the probability distribution of defect types, the coordinate offset and confidence level of defect locations, and the severity rating of defects are mapped to a three-dimensional cylindrical coordinate system (r, θ, z) with the welded pipe axis as the Z-axis, the weld start face as the Z=0 plane, and the 0° circumferential position of the welded pipe as the angle zero point. The defect depth is mapped to a radial coordinate r ranging from 0 to 8 mm, the circumferential position is mapped to an angular coordinate θ ranging from 0° to 360°, and the axial position is mapped to an axial coordinate z ranging from 0 to 1000 mm. Bicubic spline interpolation is then performed along the welded pipe axis with a step size of 1 mm to generate a 1000×360-dimensional two-dimensional defect distribution map of the weld axial-circumferential direction. Then, the defect distribution map is marked with red for porosity, yellow for slag inclusion, blue for incomplete penetration, green for lack of fusion, and black for cracks. The severity is indicated by the color depth: light colors are mild (0-0.3), medium colors are moderate (0.3-0.6) and severe (0.6-0.8), and dark colors are critical (0.8-1.0). The confidence level is indicated by the diameter of the markers, which ranges from 2 to 10 pixels, with 2 pixels representing low confidence and 10 pixels representing high confidence. The map is then superimposed and fused with the original ultrasonic B-scan image with a transparency of α=0.6 to obtain a comprehensive detection map that includes the original signal and the intelligent recognition results.

[0035] S7 compares defect information with three-dimensional judgment thresholds, calculates a comprehensive quality score by weighting according to hazard, and outputs weld quality grading results and automatic inspection reports after matching the grade thresholds.

[0036] S71 sets a three-dimensional judgment threshold matrix of defect type, severity and location according to the industry standard for aluminum alloy welded pipe weld seam. The defect detection results are compared with the three-dimensional judgment threshold matrix for each defect. For defects that exceed the standard, a weighted comprehensive calculation is performed based on the hazard weight of defect type, severity and location sensitivity to obtain the comprehensive quality score of weld seam. S72 compares the comprehensive quality score of the weld with the preset quality level threshold, and divides the weld quality into four levels: excellent, qualified, rework, and scrap, to obtain the weld quality grading results and automatically generated weld quality inspection report.

[0037] Specifically, firstly, a three-dimensional threshold matrix for defect type, severity, and location is established based on industry standards and process requirements for aluminum alloy welded pipe welds. In terms of defect type, the allowable thresholds are: porosity diameter ≤ 2.0 mm, slag inclusion length ≤ 3.0 mm, incomplete penetration depth ≤ 10% of wall thickness (0.8 mm), lack of fusion length ≤ 2.0 mm, and cracks are not allowed. In terms of severity, minor defects (0-0.3) are allowed but must be recorded; general defects (0.3-0.6) are allowed individually; severe defects (0.6-0.8) should not exceed two in total; and critical defects (0.8-1.0) are directly deemed excessive. In terms of location, the allowable threshold for defects in the root zone (0-2 mm) is reduced by 20%; the allowable threshold for defects in the capping layer zone (6-8 mm) is increased by 10%; and the standard threshold is used for defects in the filling zone (2-6 mm). The probability distribution of defect types, the severity rating of defects, and the coordinates of defect locations are compared with a three-dimensional judgment threshold matrix for each defect. For defects exceeding the standard, a weighted comprehensive calculation is performed based on the hazard weights of defect types (porosity 0.6, slag inclusion 0.7, incomplete penetration 0.9, lack of fusion 0.85, cracks 1.0), severity, and location sensitivity to obtain a comprehensive weld quality score. Then, the comprehensive weld quality score is compared with preset quality level thresholds. A comprehensive quality score ≥90 is judged as excellent, 70≤score<90 as acceptable, 50≤score<70 as requiring rework, and a score <50 as scrap, resulting in a weld quality grading result and an automatically generated weld quality inspection report.

[0038] Beneficial effects: This invention arranges multi-channel ultrasonic phased array probes along the circumference of the weld seam of an aluminum alloy welded pipe and simultaneously triggers acquisition. After bandpass filtering for noise reduction, baseline drift correction, and gain compensation based on the sound path attenuation geometric model, spatiotemporal alignment and splicing are performed to obtain a full-circumference ultrasonic echo signal matrix of the weld seam. This effectively eliminates the serious problem of ultrasonic signal attenuation caused by the low acoustic impedance, coarse grains, and significant anisotropy of aluminum alloy material, as well as the problems of complex sound beam propagation path, geometric diffusion of echo signals, and mode conversion interference caused by the curvature of the welded pipe. This significantly improves the distinguishability between defect echoes and false defect signals, and greatly reduces the false alarm rate and the missed detection rate.

[0039] This invention detects weld edges and fits a fusion line reference curve by utilizing the acoustic impedance differences between the base metal area, heat-affected zone, and weld area. It adaptively divides the weld area by combining groove geometry parameters, generates a mask, and extracts the region of interest sub-matrix by point-by-point product of the original matrix. This accurately locates the effective detection area of ​​the weld and eliminates the interference of false defect signals in the base metal area and heat-affected zone. This allows subsequent feature extraction and defect identification to be performed only in the core area of ​​the weld, improving computational efficiency and recognition accuracy.

[0040] This invention obtains the time-frequency spectrum by performing short-time Fourier transform on the signals of each channel of the submatrix, extracts the frequency band energy and energy entropy features by performing multi-scale wavelet packet decomposition, and extracts the time-domain envelope morphological features by performing Hilbert transform. The time-frequency domain features, wavelet packet energy and entropy features, and time-domain morphological features are spliced ​​and fused into a joint feature tensor in the channel dimension. This fully explores the defective feature information of time-frequency coupling in ultrasonic echo signals, overcomes the defect of existing deep learning methods that directly transfer image domain models to the one-dimensional ultrasonic time-series signal domain and cannot effectively capture time-frequency coupling features, and significantly improves the recognition accuracy of easily confused defects such as pores and inclusions.

[0041] This invention adaptively weights and enhances the joint feature tensor input to a cascaded attention network containing channel attention and spatial attention modules, and connects it with the original feature residual to obtain defect-sensitive depth features. It adaptively focuses on the defect-sensitive feature channels and spatial locations, improving the distinguishability of defect features. This allows the network to automatically learn and strengthen the time-frequency feature components that are key to defect identification, while suppressing interference from irrelevant noise features.

[0042] This invention achieves collaborative output of defect type identification, three-dimensional spatial location positioning, and severity assessment by inputting defect-sensitive depth features in parallel into a multi-task learning network with three branches: defect type classification, location regression, and severity assessment. After end-to-end collaborative training with joint loss, the network outputs defect type, location, and severity results. This overcomes the shortcomings of existing methods that only output a single classification result and lack spatial positioning and quality assessment capabilities, thus meeting the actual needs of quality grading and evaluation of aluminum alloy welded pipe welds.

[0043] This invention maps the detection results to the three-dimensional cylindrical coordinate system of the weld, generates a defect map by bicubic spline interpolation along the axial direction, and superimposes it with the original B-scan image to obtain a comprehensive detection map. This visually displays the distribution of defects in the three-dimensional space of the weld. Different colors are used to identify the defect type, color depth indicates severity, and marker size indicates confidence level. This allows inspectors to quickly and accurately grasp the overall quality of the weld, improving the readability and engineering applicability of the detection results.

[0044] This invention compares defect information with a three-dimensional threshold matrix of defect type, severity, and location, calculates a comprehensive quality score by weighting the severity, and outputs weld quality grading results and automatic inspection reports after matching the grade thresholds. This achieves intelligent grading and evaluation of weld quality, dividing the comprehensive quality score into four levels: excellent, qualified, rework, and scrap. It automatically generates an inspection report containing a defect list, location distribution map, quality score, and disposal suggestions, significantly improving inspection efficiency and standardization.

[0045] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments and equivalent changes made in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A method for identifying defects in weld seams of aluminum alloy welded pipes using ultrasonic testing based on deep learning, characterized in that, Includes the following steps: A circumferential multi-channel phased array probe synchronously acquires ultrasonic echoes from aluminum alloy welded pipe welds. After noise reduction, correction, and gain compensation, the echoes are spliced ​​together in time and space to obtain a full circumferential echo signal matrix. Based on acoustic impedance difference detection of weld edge and fitting of fusion line, weld region is divided by combining groove parameters, generating mask and dot product with original matrix to extract sub-matrix of region of interest; Short-time Fourier transform is performed on the signals of each channel of the submatrix to obtain the time spectrum. Wavelet packet decomposition is used to extract the frequency band energy / entropy, Hilbert transform is used to extract the time domain envelope, and multi-dimensional features are spliced ​​into a joint feature tensor. The joint feature tensor is input into a cascaded network with channel and spatial attention for weighted enhancement, and then connected to the original feature residual to obtain defect-sensitive depth features; This feature is input into a multi-task network that combines classification, location regression, and severity assessment branches in parallel. After training with a joint loss, the network outputs the defect type, location, and severity results. The detection results are mapped to the three-dimensional cylindrical coordinate system of the weld, and a defect map is generated by axial bicubic spline interpolation. The map is then superimposed with the original B-scan image to obtain a comprehensive detection map. The defect information is compared with the three-dimensional judgment threshold, and the comprehensive quality score is calculated by weighting the severity. After matching the level threshold, the weld quality classification result and automatic inspection report are output.

2. The method for identifying defects in ultrasonic testing of aluminum alloy welded pipe welds based on deep learning as described in claim 1, characterized in that, The process of "synchronously acquiring ultrasonic echoes from aluminum alloy welded pipe welds using a circumferential multi-channel phased array probe, followed by noise reduction, correction, gain compensation, and spatiotemporal alignment and stitching to obtain a full-circumferential echo signal matrix" includes the following steps: By uniformly arranging N ultrasonic phased array probes along the circumference of the weld seam of the aluminum alloy welded pipe, the multi-channel ultrasonic echo signal of the weld seam is simultaneously triggered and collected to obtain the original multi-channel ultrasonic echo signal sequence. After bandpass filtering and baseline drift correction, the original multi-channel ultrasonic echo signal sequence is used to establish a geometric model of the ultrasonic beam propagation path based on the outer diameter, wall thickness and incident angle of the aluminum alloy welded pipe. The sound path attenuation compensation coefficient is calculated to perform point-by-point gain compensation, and the gain-compensated ultrasonic echo signal is obtained. The multi-channel ultrasonic echo signals after gain compensation are spatiotemporally aligned and spliced ​​according to the circumferential position of the weld to obtain the full circumferential ultrasonic echo signal matrix of the weld.

3. The method for identifying defects in weld seams of aluminum alloy welded pipes based on deep learning as described in claim 1, characterized in that, The process of "detecting weld edges based on acoustic impedance differences, fitting fusion lines, dividing weld regions by combining bevel parameters, generating a mask and multiplying it with the original matrix to extract the sub-matrix of the region of interest" includes the following steps: The ultrasonic echo signal matrix of the weld circumference is used to extract the edge point set of the weld contour by taking advantage of the abrupt change in echo amplitude caused by the difference in acoustic impedance between the base material area, the heat-affected zone and the weld area. The weld fusion line reference curve is obtained by fitting the edge point set of the weld contour using the edge detection algorithm and the edge detection algorithm. By combining the weld fusion line reference curve with the groove geometry parameters, the heat-affected zone boundary is determined by expanding outward by a preset distance to both sides and the weld center zone boundary is determined by contracting inward by a preset distance, thus generating a multi-region boundary mask for the weld. The ultrasonic echo signal submatrix of the region of interest in the weld is obtained by performing point-by-point multiplication of the multi-region boundary mask of the weld with the ultrasonic echo signal matrix of the entire circumference of the weld.

4. The method for ultrasonic flaw detection and defect identification of aluminum alloy welded pipe welds based on deep learning as described in claim 1, characterized in that, The process of "performing short-time Fourier transform on each channel signal of the sub-matrix to obtain the time spectrum, wavelet packet decomposition to extract frequency band energy / entropy, Hilbert transform to extract the time domain envelope, and concatenating multi-dimensional features into a joint feature tensor" includes the following steps: The time-frequency spectrum is obtained by performing short-time Fourier transform on each channel signal in the ultrasonic echo signal submatrix of the region of interest of the weld. Multi-scale wavelet packet decomposition is performed on the time-frequency spectrum to extract the energy and energy entropy features of each frequency band. At the same time, Hilbert transform is performed on the original signal to extract the time-domain envelope morphology features. By splicing and fusing the energy distribution features of the time-frequency spectrogram, the energy / entropy features of the wavelet packet, and the temporal morphological features in the channel dimension, a multi-scale time-frequency domain joint feature tensor is obtained.

5. The method for identifying defects in weld seams of aluminum alloy welded pipes based on deep learning as described in claim 1, characterized in that, The process of "weighting and enhancing the joint feature tensor input into a cascaded network with channel and spatial attention, and connecting it with the original feature residuals to obtain defect-sensitive depth features" includes the following steps: The multi-scale time-frequency domain joint feature tensor is input into a cascaded attention network containing channel attention modules and spatial attention modules. Adaptive weighting is applied to the channel and spatial dimensions of the feature tensor to obtain the attention-weighted deep feature tensor. By performing a residual connection between the attention-weighted depth feature tensor and the original multi-scale time-frequency domain joint feature tensor, a defect-sensitive depth feature representation is obtained.

6. The method for identifying ultrasonic flaws in aluminum alloy welded pipe welds based on deep learning as described in claim 1, characterized in that, The process of "inputting the feature into a multi-task network that performs classification, location regression, and severity assessment branches in parallel, training it with a joint loss, and then outputting the defect type, location, and severity results" includes the following steps: The defect-sensitive deep feature representation is input in parallel into the defect type classification branch, the defect location regression branch, and the defect severity assessment branch to construct a multi-task learning network. The multi-task joint loss function is used for end-to-end collaborative training to obtain the multi-task defect recognition network model after training convergence. The defect-sensitive depth feature representation of the weld to be inspected is input into the multi-task defect recognition network model after training and convergence to obtain the defect type probability distribution, defect location coordinate offset and confidence level, and defect severity level score.

7. The method for identifying defects in weld seams of aluminum alloy welded pipes based on deep learning as described in claim 1, characterized in that, The process of "mapping the detection results to the three-dimensional cylindrical coordinate system of the weld, generating a defect map through axial bicubic spline interpolation, and overlaying it with the original B-scan image to obtain a comprehensive detection map" includes the following steps: The probability distribution of defect type, the coordinate offset and confidence level of defect location, and the score of defect severity are mapped to the three-dimensional cylindrical coordinate system of weld. Bicubic spline interpolation is performed along the weld pipe axis to generate a two-dimensional defect distribution map of weld axis-circumference. The defect distribution map is used to identify defect types with different colors, severity with color intensity, and confidence level with marker size. It is then overlaid and fused with the original ultrasound B-scan image to obtain a comprehensive detection map that includes the original signal and intelligent recognition results.

8. The method for identifying defects in ultrasonic testing of aluminum alloy welded pipe welds based on deep learning as described in claim 1, characterized in that, The process of "comparing defect information with three-dimensional judgment thresholds, calculating a comprehensive quality score based on hazard weighting, and outputting weld quality grading results and automatic inspection reports after matching grade thresholds" includes the following steps: Based on the industry standard for aluminum alloy welded pipe weld seams, a three-dimensional judgment threshold matrix of defect type, severity, and location is set. The defect detection results are compared with the three-dimensional judgment threshold matrix one by one. For defects exceeding the standard, a weighted comprehensive calculation is performed based on the hazard weight of defect type, severity, and location sensitivity to obtain the comprehensive quality score of the weld seam. The overall quality score of the weld is compared with the preset quality level threshold, and the weld quality is divided into four levels: excellent, qualified, rework and scrap. The weld quality classification results and automatically generated weld quality inspection reports are obtained.

Images