Ultrasonic-based non-destructive testing method and system for a welding rectification zone of a steel member

Abstract

The application relates to the technical field of ultrasonic nondestructive testing, and discloses a steel member welding correction region nondestructive testing method and system based on ultrasonic waves. The method comprises the following steps: acquiring an original ultrasonic signal sequence; performing short-time Fourier transform on the original ultrasonic signal sequence to obtain a time-domain enhanced signal sequence; constructing a two-dimensional B-Scan gray image according to the time-domain enhanced signal sequence, and determining a defect region of interest based on gradient amplitude distribution; constructing a total energy functional and performing evolution processing on an initial contour of the defect region of interest by using a greedy iterative algorithm to obtain a refined defect contour; calculating a morphological quantitative index of the refined defect contour and inputting the morphological quantitative index into a pre-trained support vector machine to obtain a defect type label; and combining the defect type and physical size to query a decision matrix to determine a final defect evaluation grade. The method can effectively suppress structural noise.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic nondestructive testing technology, and in particular to a method and system for nondestructive testing of welded correction areas of steel components based on ultrasonic waves. Background Technology

[0002] Currently, steel structures serve as the key framework for modern architecture, bridges, and shipbuilding, and their welding quality directly impacts the overall safety, stability, and service life of infrastructure. With the increasing complexity of steel component manufacturing processes, particularly for welded areas subjected to heat treatment processes such as flame straightening, achieving high-precision non-destructive testing under complex metallographic structures and stress interference, and ensuring the accurate identification and quantitative assessment of minute internal defects, has become a core challenge in this field.

[0003] In existing technologies, traditional non-destructive testing (NDT) schemes primarily rely on manual handheld probes or basic automated scanning equipment for signal acquisition and analysis. These schemes typically utilize acquisition terminals integrated with smart sensors to acquire ultrasonic A-scan signals and determine the presence of defects using preset fixed amplitude thresholds. However, due to the localized high-temperature heating and mechanical deformation experienced in the welded straightening areas of steel components, their internal grain structure is often non-uniform and contains complex residual stresses, leading to strong structural scattering noise during ultrasonic wave propagation. When such high-intensity coherent noise is mixed into the signals acquired by smart sensors, traditional signal processing algorithms struggle to effectively separate defect echoes from structural backscattering, often resulting in the characteristics of microcracks being submerged by noise or the geometric boundaries of pores becoming blurred. This leads to significant deviations in subsequent defect type identification and size quantification, easily causing missed detections or misjudgments.

[0004] Existing technologies suffer from low accuracy in defect detection and unreliable assessment. Summary of the Invention

[0005] This invention provides a non-destructive testing method and system for welded straightening areas of steel components based on ultrasound, in order to solve the technical problems of low defect detection accuracy and unreliable evaluation in the prior art.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a non-destructive testing method for welded straightening areas of steel components based on ultrasound, comprising: The ultrasonic acquisition device is controlled to continuously scan along the welding correction area of ​​the steel component to obtain the original ultrasonic signal sequence, and a short-time Fourier transform is performed on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence. The two-dimensional time-frequency matrix sequence is input into a pre-trained convolutional neural network model to generate a corresponding time-frequency feature mask sequence. The time-frequency feature mask sequence and the two-dimensional time-frequency matrix sequence are then multiplied frame by frame and reconstructed by inverse transformation to obtain a time-domain enhanced signal sequence. Based on the scanning position information of the ultrasound acquisition device, the time-domain enhanced signal sequence is mapped into a two-dimensional B-Scan grayscale image; The gradient magnitude distribution of the two-dimensional B-Scan grayscale image is calculated, the region of interest of the defect is determined based on the local gradient maxima of the gradient magnitude distribution, and the geometric boundary of the region of interest of the defect is extracted as the initial evolution contour. A total energy functional containing an internal smooth energy term and an external gradient energy term is constructed. A greedy iterative algorithm is used to drive the initial evolution profile to shrink towards the high gradient edge to obtain a refined defect profile. Based on the refined defect contour, a morphological quantification index is calculated, and the morphological quantification index is input into a pre-trained support vector machine classifier to output a defect type label. Based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index, a preset decision matrix is ​​queried to determine the final defect assessment level.

[0007] Secondly, the present invention provides a non-destructive testing system for welded straightening areas of steel components based on ultrasound, comprising: The signal acquisition and transformation module is used to control the ultrasonic acquisition device to continuously scan along the welding correction area of ​​the steel component, acquire the original ultrasonic signal sequence, and perform a short-time Fourier transform on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence. The deep enhancement module is used to input the two-dimensional time-frequency matrix sequence into the pre-trained convolutional neural network model one by one to generate the corresponding time-frequency feature mask sequence, and perform frame-by-frame dot multiplication operation between the time-frequency feature mask sequence and the two-dimensional time-frequency matrix sequence, and perform inverse transformation reconstruction to obtain the time-domain enhanced signal sequence. The image construction module is used to map the time-domain enhanced signal sequence into a two-dimensional B-Scan grayscale image based on the scanning position information of the ultrasound acquisition device. The initial contour extraction module is used to calculate the gradient magnitude distribution of the two-dimensional B-Scan grayscale image, determine the region of interest of the defect based on the local gradient maxima in the gradient magnitude distribution, and extract the geometric boundary of the region of interest of the defect as the initial evolution contour. The contour refinement module is used to construct a total energy functional that includes an internal smooth energy term and an external gradient energy term. It uses a greedy iterative algorithm to drive the initial evolved contour to shrink towards the high gradient edge to obtain a refined defect contour. The intelligent classification module is used to calculate the morphological quantification index based on the refined defect contour, input the morphological quantification index into a pre-trained support vector machine classifier, and output the defect type label. The assessment and decision module is used to query a preset decision matrix based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index to determine the final defect assessment level.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention converts the original ultrasonic signal into a two-dimensional time-frequency matrix and introduces a pre-trained convolutional neural network to generate a time-frequency feature mask for deep denoising and reconstruction of the signal. This technique fully utilizes the powerful feature extraction capability of deep learning for non-stationary signals, and can accurately distinguish between the transient time-frequency aggregation mode of defect echoes and the strong structural scattering noise (coherent noise) unique to the welding correction area. Compared with traditional frequency domain filtering methods, this method significantly suppresses complex background noise while preserving the high-frequency detail features of weak defects to the greatest extent, thereby effectively improving the detection signal-to-noise ratio and sensitivity, and solving the technical problem that small defects are easily missed in strong scattering environments.

[0009] (2) This invention constructs a total energy functional containing internal smoothing terms and external gradient terms, and uses a greedy iterative algorithm to drive the dynamic evolution of the initial contour, thereby achieving refined extraction of defect boundaries. This method overcomes the limitations of traditional threshold segmentation or edge detection operators in accurately locating in low-contrast images, and uses the principle of energy minimization to make the contour adaptively shrink and closely fit the real physical boundary. This not only effectively repairs edge breaks or blurring caused by noise interference, but also achieves sub-pixel level contour positioning accuracy, ensuring the accuracy of subsequent calculations of geometric parameters such as area and perimeter, and providing reliable data support for the quantitative evaluation of defects.

[0010] (3) This invention extracts morphological quantification indicators and combines them with support vector machines for intelligent classification, further introducing spatial distribution density to correct the risk coefficient, thus establishing a comprehensive evaluation system for the entire chain. On the one hand, this scheme accurately distinguishes between high-risk planar defects (such as cracks) and low-risk volumetric defects (such as pores) using shape compactness factors; on the other hand, it quantifies the mutual interference and cumulative destructive effects of defect clusters through spatial density analysis, avoiding the underestimation of risk caused by isolated evaluation of densely distributed micro-defects. The final evaluation conclusion, which includes clear handling suggestions (such as rework and monitoring), significantly improves the engineering practical value and decision-making reliability of the detection results. Attached Figure Description

[0011] Figure 1This is a schematic diagram of the non-destructive testing method for the welded straightening area of ​​steel components based on ultrasound, provided in the first embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the ultrasonic-based non-destructive testing system for the welded correction area of ​​steel components provided in the second embodiment of the present invention. Detailed Implementation

[0012] 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, and 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.

[0013] Reference Figure 1 The first embodiment of the present invention provides a non-destructive testing method for the welded straightening area of ​​steel components based on ultrasound, comprising the following steps: S11, control the ultrasonic acquisition device to continuously scan along the welding correction area of ​​the steel component to acquire the original ultrasonic signal sequence, and perform a short-time Fourier transform on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence. S12, the two-dimensional time-frequency matrix sequence is input into the pre-trained convolutional neural network model to generate the corresponding time-frequency feature mask sequence, and the time-frequency feature mask sequence and the two-dimensional time-frequency matrix sequence are multiplied frame by frame and reconstructed by inverse transformation to obtain the time-domain enhanced signal sequence; S13, based on the scanning position information of the ultrasound acquisition device, the time-domain enhanced signal sequence is mapped into a two-dimensional B-Scan grayscale image; S14, calculate the gradient magnitude distribution of the two-dimensional B-Scan grayscale image, determine the region of interest of the defect based on the local gradient maxima of the gradient magnitude distribution, and extract the geometric boundary of the region of interest of the defect as the initial evolution contour. S15, construct a total energy functional that includes an internal smooth energy term and an external gradient energy term, and use a greedy iterative algorithm to drive the initial evolution profile to shrink towards the high gradient edge to obtain a refined defect profile; S16, Calculate the morphological quantification index based on the refined defect contour, input the morphological quantification index into the pre-trained support vector machine classifier, and output the defect type label. S17. Based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index, query the preset decision matrix to determine the final defect assessment level.

[0014] In step S11, the ultrasonic acquisition device is controlled to continuously scan along the welding correction area of ​​the steel component to acquire the original ultrasonic signal sequence. A short-time Fourier transform is then performed on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence, including: The ultrasound acquisition device is controlled to acquire the original A-scan signal sequence at a preset sampling rate; The Hanning window is selected as the window function, and a preset window length and overlap rate are set. A sliding window short-time Fourier transform is performed on each frame of the original A-scan signal sequence. The amplitude spectrum of the transformation result is calculated to generate the two-dimensional time-frequency matrix sequence, wherein the horizontal axis of the two-dimensional time-frequency matrix represents time, the vertical axis represents frequency, and the matrix element values ​​represent frequency energy distribution.

[0015] It should be noted that the acquisition of the raw A-scan signal sequence by the ultrasound acquisition device is achieved using a position synchronization triggering mechanism. The system monitors the physical displacement of the probe in real time through an incremental rotary encoder connected to the scanner. Whenever the probe moves through a preset spatial step (e.g., 0.5 mm), the encoder sends a TTL trigger pulse to the ultrasound acquisition card, triggering the acquisition card to record a segment of length... The time-domain voltage signal is denoted as a single-frame A-scan signal. ,in The continuous scanning process generates a series of A-scan signals ordered by spatial location, which constitutes the original ultrasonic signal sequence.

[0016] It should be noted that the use of the Hanning window and the execution of a sliding window short-time Fourier transform (STFT) are intended to extract local non-stationary features of the signal. The Hanning window function... The calculation formula is: in, The preset window length is given. For each frame with a length of... A-scan signal The calculation process for STFT is as follows: in, For time window indexing, The step size for window sliding (determined by the overlap ratio). For frequency indexing. Calculate the complex number result. modulus This yields the two-dimensional time-frequency matrix. Each column of this matrix corresponds to a time window, and each row corresponds to a frequency component. The values ​​of the elements in the matrix directly reflect the energy intensity at that frequency at that moment.

[0017] It is worth noting that the preset sampling rate was determined based on a comprehensive consideration of the ultrasonic probe's center frequency and the Nyquist sampling theorem. The effective bandwidth upper limit of the probe's echo signal was determined through spectral analysis of historical detection data. Select A preset sampling rate of 5 to 10 times (e.g., 100MHz) is used to ensure complete capture of high-frequency harmonic components caused by minute defects. The preset window length is determined based on statistical analysis of historical defect echo durations. The system statistically analyzes the average pulse width of typical crack defect echoes from historical data. And select those that can cover The number of sampling points for the duration is used as the preset window length (e.g., 64 points) to achieve an optimal balance between time resolution and frequency resolution. The preset overlap rate is determined based on the energy recovery criterion, and is typically selected as 75% (i.e., the step size). This is to minimize signal energy loss caused by window function edge attenuation.

[0018] For example, in a single testing task, an ultrasonic probe with a center frequency of 5MHz is used, and the preset sampling rate is set to 50MHz. The length of one frame of A-scan signal acquired is... Point. Set the preset window length. Points, with a preset overlap rate of 75% (i.e., step size). (Point). Using the STFT formula described above to process this signal, the time axis dimension of the generated two-dimensional time-frequency matrix is... The frequency axis dimension is (If a single-sided spectrum is taken, it would be 33). Assuming that at the index of 5MHz in the 30th time window (corresponding to the echo arrival time), the calculated amplitude spectrum value is 0.85, then this matrix element... The value is 0.85, representing the concentration of the dominant frequency energy of the echo at that moment. The system repeats this process for all A-scan signals in the sequence, and finally outputs a two-dimensional time-frequency matrix sequence containing thousands of such matrices.

[0019] In step S12, the two-dimensional time-frequency matrix sequence is input into a pre-trained convolutional neural network model to generate a corresponding time-frequency feature mask sequence. The time-frequency feature mask sequence and the two-dimensional time-frequency matrix sequence are then multiplied frame-by-frame, and an inverse transform is performed to reconstruct the time-domain enhanced signal sequence, including: The deep time-frequency clustering pattern in the two-dimensional time-frequency matrix sequence is extracted using the convolutional neural network model, and the time-frequency feature mask sequence with values ​​between 0 and 1 is output, wherein the mask value of the region corresponding to the defect echo is close to 1, and the mask value of the noise region is close to 0. The time-frequency feature mask sequence is multiplied element-wise with the two-dimensional time-frequency matrix sequence to suppress broadband noise energy with low mask values, resulting in a weighted time-frequency matrix sequence. Perform a short-time inverse Fourier transform on the weighted time-frequency matrix sequence to restore the phase information and obtain the time-domain enhanced signal sequence with improved signal-to-noise ratio.

[0020] It should be noted that the generation of the time-frequency feature mask sequence using the convolutional neural network model is implemented using a U-Net network based on an encoder-decoder architecture. The U-Net network is a preferred implementation of the convolutional neural network model, specifically designed to solve pixel-level image segmentation and mask generation problems. For each two-dimensional time-frequency matrix in the sequence... (size is) The network first performs multi-layer convolution and max-pooling operations through the encoder path to extract the local texture and global distribution features of the defect echo in the time-frequency domain, obtaining a deep abstract feature map. Then, it performs upsampling and feature concatenation (Skip Connection) through the decoder path to restore the spatial resolution of the feature map. In the network's output layer, a sigmoid activation function is used to map the output value of each pixel to... Generate a soft mask matrix with the same size as the input matrix within the interval. The formula for calculating the Sigmoid function is: in, For the last layer of the network in time index and frequency index The linear output value at that point. This generates... This refers to a single-frame mask in the time-frequency feature mask sequence.

[0021] It should be noted that the element-wise multiplication of the time-frequency feature mask sequence with the two-dimensional time-frequency matrix sequence, followed by the inverse short-time Fourier transform (ISTFT), is performed in accordance with the principle of phase preservation and overlap-add. Specifically, the system first calls the original phase information preserved during the STFT calculation in step S11. Then, using a soft mask matrix For the original amplitude spectrum Weighted calculations are performed to compute the enhanced complex spectrum. : Finally, for The discrete reconstruction formula for performing the ISTFT operation is as follows: in, This is the synthesis window function (usually the same as the analysis window). The step size is [value]. This operation synthesizes the denoised waveform in the time domain, i.e., the time-domain enhanced signal sequence.

[0022] It is worth noting that the construction of the pre-trained convolutional neural network model is achieved using a hybrid data augmentation strategy based on physical model simulation and real texture transfer to ensure the completeness and objectivity of the training data. First, to obtain absolutely pure and comprehensive positive samples (defect echoes), the system uses the Finite Difference Time Domain (FDTD) algorithm to construct a three-dimensional acoustic field simulation model of the steel component. Various defect geometries such as cracks, pores, and lack of fusion are set in the model, and the propagation and reflection process of ultrasonic waves in the medium is simulated. By setting non-reflective boundary conditions to eliminate boundary echo interference, a noise-free pure defect echo signal library is calculated. Secondly, in order to obtain negative samples (background noise) that conform to actual working conditions, the system selects a defect-free parent material region with the same material and grain size as the steel component under test, performs long-term continuous scanning, and collects a real structural scattering noise signal library. Subsequently, a pairwise noisy-clean training sample set is constructed using the following linear superposition formula. : in, and The dynamic gain coefficient is used to randomly adjust the ratio of the two factors to ensure that the signal-to-noise ratio (SNR) of the generated signal is uniformly distributed within a certain range. to The range is designed to cover various operating conditions, from weak signals to strong signals. To simulate the Gaussian white noise component of electronic thermal noise. Finally, an ideal binary mask. The generation is automatically determined based on the actual physical location of the simulation model, and the pure signal is calculated. The Hilbert envelope is used to mark time index intervals where the envelope amplitude exceeds a preset energy threshold (e.g., 5% of the peak value) as 1 (foreground), and the remaining regions as 0 (background). Using this automatically generated labeled data to train the network completely eliminates the subjective errors of manual annotation.

[0023] For example, for an input two-dimensional time-frequency matrix, in coordinates At this point, the original amplitude This location corresponds to the actual center of the defect echo. The pre-trained convolutional neural network model identifies its time-frequency clustering characteristics and outputs a mask value. After dot product, the weighted amplitude is The defect energy is almost completely preserved. And in the coordinates... At this point, the original amplitude This location contains random scattering noise. The convolutional neural network model outputs the mask value. After dot product, the weighted amplitude is The noise energy was attenuated by 95%. After ISTFT reconstruction, the signal-to-noise ratio of the frame signal was improved from the original 6dB to 22dB.

[0024] In step S13, based on the scanning position information of the ultrasound acquisition device, the time-domain enhanced signal sequence is mapped into a two-dimensional B-Scan grayscale image, including: Perform a Hilbert transform on each frame of the time-domain enhanced signal sequence and calculate the signal envelope amplitude; The signal envelope amplitude is mapped to a gray value of a preset bit depth to obtain a one-dimensional gray vector sequence; Based on the scanning position information, the one-dimensional grayscale vector sequence is spatially arranged and interpolated in the scanning direction to generate the two-dimensional B-Scan grayscale image.

[0025] It should be noted that for each frame of time-domain enhanced signal Performing the Hilbert Transform aims to extract the instantaneous energy characteristics of the signal and eliminate the interference of high-frequency carrier oscillations in the radio frequency waveform on the imaging visual effect. This operation first constructs an analytic signal. : in, for The Hilbert transform is defined as: Then, the magnitude of the analytic signal is calculated to obtain the signal envelope amplitude, which reflects the strength of the echo energy. : It should be noted that for discrete sequences, the Hilbert transform can be implemented using a Fast Fourier Transform (FFT); mapping the signal envelope amplitude to grayscale values ​​of a preset bit depth is achieved using a linear dynamic range compression algorithm. This algorithm first traverses the entire time-domain enhanced signal sequence and calculates the global maximum envelope amplitude in all frames. and global minimum envelope magnitude Set the preset bit depth to (For example, 8-bit, i.e., 256 levels of grayscale). For the envelope value at any given time... Its corresponding grayscale value The calculation formula is: This mapping quantizes floating-point voltage amplitudes into integer gray levels, forming a dataset composed of multiple one-dimensional gray vectors.

[0026] It should be noted that the generation of the two-dimensional B-Scan grayscale image based on the scan position information is achieved using spatial coordinate remapping and bilinear interpolation techniques. The system constructs a two-dimensional matrix. The vertical axis represents the sound path depth (corresponding to the sampling point index), and the horizontal axis represents the scanning distance (corresponding to the scanning position information). The one-dimensional grayscale vector of each frame is then calculated based on its corresponding encoder physical coordinates. The values ​​are filled into the corresponding columns of the matrix. Since the physical scan step size may not perfectly match the preset horizontal resolution of the image, the system uses bilinear interpolation to calculate the pixel values ​​at non-sampling point locations. For any coordinate in the image... Find its four nearest known points in the physical grid. The gray value of this point is calculated by weighted average: The final generated matrix This refers to the two-dimensional B-Scan grayscale image, where the high grayscale value areas visually reflect the defect reflection surfaces inside the steel component.

[0027] It is worth noting that the preset image lateral resolution is determined based on a joint statistical analysis of the ultrasonic probe beam width and encoder accuracy. The system first measures the -6dB beam width of the probe at the target depth. (For example, 2.0mm), according to the spatial sampling theorem, in order to avoid image aliasing and ensure the lateral continuity of minute defects, select... One-tenth to one-fifth of the image resolution is used as a baseline. Simultaneously, the minimum physical step of the statistical encoder... Finally selected The preset horizontal resolution of the image (e.g., 0.2 mm) is used to achieve the best balance between imaging accuracy and data size.

[0028] It is worth noting that the preset bit depth was determined based on a comprehensive evaluation of the human eye's visual resolution threshold and computer storage efficiency. Psychophysical experiments were used to measure the minimum perceptible difference (JND) of the human eye in response to grayscale changes. The results showed that the human eye can typically distinguish 60 to 100 grayscale levels. To provide sufficient visual redundancy and adapt to common image processing standards, the closest power of 2 to this range, i.e., 8 bits (256 levels), was selected as the preset bit depth.

[0029] For example, a temporal augmentation signal sequence containing 200 frames, each frame being 1024 points in length. Global maximum envelope amplitude. Minimum amplitude The bit depth is set to 8 bits. For the 200th sample point in the 50th frame of the signal, its Hilbert envelope amplitude is... Substitute the values ​​into the formula to obtain the grayscale value. Based on the scan location information, the physical location corresponding to this frame is... The system maps the grayscale vector to the corresponding column of the B-Scan image. The final generated B-Scan image has a size of [size missing]. The pixels are sufficient to clearly display the echo brightness of a defect at a depth of 200 points and a horizontal position of 10mm, which is 128 (medium brightness).

[0030] In step S14, the gradient magnitude distribution of the two-dimensional B-Scan grayscale image is calculated, the region of interest (ROI) for the defect is determined based on the local gradient maxima in the gradient magnitude distribution, and the geometric boundary of the ROI is extracted as the initial evolution contour, including: The gradient magnitude distribution is obtained by calculating the derivatives of the two-dimensional B-Scan grayscale image in the time and depth directions using the gradient operator. Local maxima points are extracted from the gradient magnitude distribution, and a double threshold hysteresis connection algorithm is applied to connect strong and weak edge pixels to form a binarized edge map. A morphological closing operation is performed on the binarized edge map to obtain a closed connected region after filling the gap. The closed connected region is defined as the region of interest of the defect, and the outer contour of the closed connected region is extracted as the initial evolution contour.

[0031] It should be noted that the calculation of derivatives and the synthesis of gradient magnitude distributions using the gradient operator are implemented using Sobel operator convolution. This operator contains two... The convolutional kernels are used to detect edge responses in the horizontal (time axis) and vertical (depth axis) directions, respectively. Let... Given a 2D B-Scan grayscale image as input, the horizontal convolution kernel... and vertical convolution kernel They are defined as follows: Calculate the horizontal gradient components separately and vertical gradient components ( (This represents a convolution operation). Subsequently, each pixel in the image is calculated. gradient magnitude : The resulting matrix This refers to the gradient magnitude distribution, where high magnitude regions correspond to defect boundaries in the image where grayscale changes drastically.

[0032] It should be noted that the application of the hysteresis thresholding algorithm to connect pixels with strong and weak edges is based on the core idea of ​​Canny edge detection. This algorithm first defines a high threshold. and a low threshold Traverse the gradient magnitude distribution The following decision logic is executed: if Then mark the pixel as a strong edge point; if If so, then mark that pixel as a background point; if If a pixel is found to be a weak edge, it is marked as such. Then, connectivity analysis is performed. For each weak edge, it is checked whether a strong edge exists within its 8-neighborhood. If a strong edge exists, it is included as a true edge; otherwise, it is suppressed as background. The resulting binary image is the binarized edge map, where a pixel value of 1 (or 255) represents an edge, and 0 represents background.

[0033] It should be noted that performing morphological closing operations on the binarized edge map aims to bridge edge breaks caused by noise or weak gradients. The morphological closing operation is defined as dilation followed by erosion. A structuring element of a preset size is selected. (For example (rectangle), for binary images Perform the calculation: This operation fills in contour gaps and holes smaller than the structuring element size, merging discrete edge fragments into a complete closed region, i.e., the closed connected region. The system then uses a contour tracing algorithm (such as the Suzuki algorithm) to extract the outermost set of pixel coordinates of this connected region as the initial evolving contour.

[0034] It is worth noting that the high threshold and low threshold The determination is based on statistical analysis of the gradient histograms of historical defect images. The system calculates the cumulative distribution function (CDF) of gradient magnitudes in historical samples, and selects the magnitude at which the CDF reaches 90% as the mean. and set This is to ensure that weak but continuous defect boundaries can be preserved. The size of the structural element is determined based on statistical analysis of the width of edge fracture gaps in historical samples. The system statistically analyzes the distribution of the maximum fracture distance at the defect edge, selects its 95th percentile (e.g., 4 pixels) as a benchmark, and sets the size of the structural element to an odd number (e.g., 5 pixels) that is slightly larger than this benchmark.

[0035] For example, for a The local region of a pixel using B-Scan. In coordinates... At this point, obtained through Sobel operation Calculate the gradient magnitude The threshold determined based on historical statistics is... .because This point is marked as a strong edge. (In coordinates...) At this point, the gradient magnitude is It belongs to the weak edge ( However, due to its proximity to strong edge points Adjacent, therefore retained. If There is a breakpoint with a width of 2 pixels (amplitude less than 40) at that point. After the structuring element is closed, the breakpoint is filled, forming a continuous contour. The final extracted initial evolution contour is a set of closed coordinate sequences. .

[0036] In step S15, a total energy functional containing an internal smooth energy term and an external gradient energy term is constructed. A greedy iterative algorithm is used to drive the initial evolution profile to shrink towards the high gradient edge, resulting in a refined defect profile, including: The total energy functional is established, wherein the inner smooth energy term is used to constrain the continuity and curvature of the contour, and the outer gradient energy term is constructed based on the gray-level gradient field of the region of interest of the defect, and is used to attract the contour to the high gradient edge; Calculate the energy change of the total energy functional when the control point on the initial evolution profile moves to a neighborhood position; The displacement vector of the control point is determined according to the principle of energy minimization, and the position of the control point is iteratively updated until the energy change value is less than the preset energy convergence threshold, thereby obtaining the refining defect profile.

[0037] It should be noted that the total energy functional is established using a Discrete Active Contour Model (often called the Snake model). For the... Control points The closed contour formed (where Its total energy functional Defined as the weighted sum of the energies at all control points: in, This is the continuous component in the internal smoothing energy term, used to force the control points to maintain an equidistant distribution; its calculation formula is: ,in This is the average distance between all adjacent points on the contour; The curvature component in the internal smoothing energy term is used to limit sharp angles in the contour; its calculation formula is as follows: ; The external gradient energy term directly references the gradient magnitude distribution calculated in step S14. Its calculation formula is That is, the larger the gradient value, the lower the energy; These are the corresponding weighting coefficients.

[0038] It should be noted that the contour shrinkage driven by the Greedy Iteration Algorithm follows a local optimum search logic. In each iteration, the system traverses every control point on the contour. For the current control point Systematically examine its performance (or Each candidate location within the spatial neighborhood (A total of 9 or 25 candidate points). For each candidate position, substitute into the above formula to calculate... Move to Local energy value at time Selecting the option that enables The neighborhood where the minimum value is reached As the target location. If Not equal to the current position Then a displacement vector is generated. and update the control points to After completing one round of updates for all control points, calculate the change in total energy across the entire map. .like If the energy convergence threshold is less than the preset energy threshold, or if the preset maximum number of iterations is reached, the evolution stops, and the set of control points at the final position is output as the refining defect profile.

[0039] It is worth noting that the weighting coefficients (Continuity weights) and The determination of the curvature weight is based on statistical analysis of historical defect morphologies. The system extracts manually annotated real defect contours from the historical database and calculates the average curvature variance and point spacing variance of their edges. A grid search method is used to find a set of parameters that maximizes the IoU (Intersection over Union) between the contour evolving under the drive of these parameters and the ground truth. For example, for crack-type defects, a larger value is typically selected. A value (e.g., 1.0) is used to maintain the smoothness of the contour and prevent it from getting caught in image noise; while for pore-like defects, a smaller value is selected. The value is set to allow the contour to adapt to the circular boundary. The preset energy convergence threshold is determined based on the asymptotic analysis of the energy decay curve. Analyzing the historical convergence process, when the rate of change of total energy is less than 0.1% of the initial energy, the contour position has basically stabilized. Therefore, the value corresponding to this ratio is used as the preset energy convergence threshold.

[0040] For example, consider a control point on the contour. Current coordinates The point before it. lie in The last point lie in Average spacing Calculate the current energy. . Assume the gradient magnitude at this point is... ,but .set up Then the total energy Examine neighboring points Assuming the gradient magnitude at this point (Closer to the edge), New (distance )for ,new for ,new New total energy .because Energy is significantly reduced, and the system determines the displacement vector. Update control points to The iteration ends when the total energy change of all points in the entire graph is less than a preset energy convergence threshold (e.g., 0.5).

[0041] In step S16, a morphological quantification index is calculated based on the refined defect contour. The morphological quantification index is then input into a pre-trained support vector machine classifier, which outputs a defect type label, including: Traverse the set of boundary points of the refined defect contour, calculate the pixel area and pixel perimeter, and convert them into physical area and physical perimeter by combining the system calibration coefficients; Calculate the shape compactness factor based on the physical area and the physical perimeter; Construct a multidimensional feature vector that includes the physical area, the physical perimeter, and the shape compactness factor; The pre-trained support vector machine classifier with radial basis function kernel is invoked to calculate the decision function value of the multidimensional feature vector. Based on the decision function value, the defect is determined to be of the crack type or the porosity type, and the defect type label is generated.

[0042] It should be noted that the calculation of geometric parameters by traversing the boundary point set of the refined defect profile is implemented using a polygon geometry analytical method. Assume the refined defect profile consists of a set of clockwise ordered coordinate points. First, the pixel area is calculated using the discrete form of Green's Formula. : Secondly, the pixel perimeter is calculated using the Euclidean distance accumulation method. : Subsequently, the preset system calibration coefficients are read. (Unit: mm / pixel), the physical area is obtained by performing a dimensional conversion using the following formula. With physical perimeter : It should be noted that calculating the shape compactness factor and constructing the multidimensional feature vector aims to extract size-independent morphological features to distinguish between strip-shaped and clumpy defects. Shape compactness factor The calculation formula is as follows: The closer the factor value is to 1, the closer the defect morphology is to a circle (such as pores); the closer the value is to 0, the more elongated the defect morphology is (such as cracks). After normalizing the above three physical quantities, the system constructs a multidimensional feature vector. .

[0043] It should be noted that the decision-making process using a pre-trained Support Vector Machine (SVM) classifier is implemented based on a high-dimensional mapping using the Radial Basis Function (RBF). The decision function of this classifier... Defined as: in, For support vectors, The category labels corresponding to the support vectors are set (+1 represents cracks, -1 represents pores). For Lagrange multipliers, For kernel function parameters, This is the bias term. The system calculates the input vector. The decision function value, if If so, a crack label is generated; if Then, a pore label is generated.

[0044] It is worth noting that the system calibration coefficients The determination was based on calibration experiments using a standard CSK-IA test block. By scanning a standard reflective aperture of known diameter (e.g., 1.0 mm) on the test block, the pixel diameter of the aperture in the image was measured, and the average of the ratios was taken as the mean. The model parameters of the SVM classifier ( The parameters were determined based on offline training on a historical defect database. The system collects historical defect samples verified by metallographic experiments and uses a grid search combined with 5-fold cross-validation to find the parameter combination that maximizes the classification F1 score in the parameter space as the preset model parameters.

[0045] For example, for a refined defect profile, the pixel area is calculated. Pixel, pixel perimeter Pixels. Known system calibration coefficients. In physical transformation, Perform compactness calculations. Discriminant construction of feature vectors Inputting the SVM classifier, the decision function value (hypothesis) is calculated to obtain... Since the compactness is significantly lower than the circularity threshold (typically >0.8) and the decision value is positive, the system outputs the defect type label as crack.

[0046] In step S17, based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index, a preset decision matrix is ​​queried to determine the final defect assessment level, including: The basic risk coefficient is determined based on the defect type label. If the defect type label is a crack, the basic risk coefficient of the first value is set. Calculate the spatial distribution density of the refining defect profile within the scanning area, and use the spatial distribution density to weight and correct the basic risk coefficient to obtain the corrected risk coefficient. By combining the corrected risk coefficient with the physical size data, a severity value is obtained; The threshold range to which the severity value belongs is queried in the preset decision matrix, and the final defect assessment level, which includes rework, monitoring, or pass / fail conclusions, is obtained.

[0047] It should be noted that determining the basic risk coefficient This is achieved by consulting a pre-set defect risk mapping table. This table defines the correspondence between defect types and dimensionless risk values. If the defect type label generated in step S16 is "crack," then... (i.e., the first value); if the label is an air vent, then set This setting reflects the physical fact that planar defects (cracks) are significantly more sensitive to stress concentration than volumetric defects (porosity).

[0048] It should be noted that calculating the spatial distribution density and the corrected risk coefficient aims to quantify the nonlinear weakening effect of defect clusters on structural strength. The system first uses the centroid of the current defect as the center of a circle with a preset influence radius. Construct a local search neighborhood. Traverse the refined defect contours of all identified defects (including the current defect itself) within this neighborhood and calculate the sum of their physical areas. Then, the spatial distribution density is calculated using the following formula. : Furthermore, the adjusted risk coefficient is calculated using an exponential growth model. : in, This represents the clustering impact factor. The formula indicates that the denser the defect distribution, the greater the potential disruptive risk exhibits a superlinear growth.

[0049] It is worth noting that the aforementioned clustering influence factor The determination was based on multi-defect stress coupling analysis under finite element analysis (FEA). The system first constructs a parametric mechanical model of the welding area, setting different spatial distribution densities. A random defect cluster sample (values ​​ranging from 0.0 to 0.5). After applying a standard yield load to the model, the equivalent stress concentration factor for each sample is calculated. For a large number of simulation-generated The data were subjected to exponential regression analysis, and the growth rate parameter of the best-fit curve was extracted as the preset parameter. The value (e.g., 10.0). This parameter objectively reflects the physical characteristic that when the defect spacing is smaller than the size of the plastic zone, the interference of stress fields leads to a sharp increase in the probability of local material failure.

[0050] Furthermore, the exponential regression analysis employs the least squares method for fitting to determine the growth rate parameter; specifically, the least squares method is used to solve for the parameter that minimizes the sum of squared errors between data pairs, and the resulting value is the clustering influence factor, which is a dimensionless clustering coupling coefficient with a typical value range of 5-15, the specific value depending on the material properties and defect type.

[0051] It should be noted that the severity value is obtained by mapping the adjusted risk coefficient to physical size data. This is achieved using a linear weighted product method. The physical area calculated in step S16 is directly referenced. (unit: The calculation formula is: Get the value Then, the system queries a preset decision matrix. This matrix contains three mutually exclusive threshold intervals and their corresponding evaluation conclusions. The conclusion is immediate rejection; if The conclusion is to prioritize monitoring; if The conclusion is "Accept".

[0052] It is worth noting that the preset radius of influence The determination of the plastic zone radius is based on the estimation of the plastic zone size in fracture mechanics. For common steel structural materials (such as Q345), the plastic zone radius at a typical crack tip is calculated based on its yield strength and fracture toughness. Select 3 times as (For example, 5.0 mm) to cover areas where stresses interfere with each other. The threshold value... and The determination was based on a joint analysis of historical failure cases and industry standards (such as AWS D1.1). The system statistically analyzed historical data to determine the minimum severity value that led to structural failure, and selected the lower limit of its safety margin (e.g., 5.0 mm). 2 As Select 40% as (e.g., 2.0mm) 2 ).

[0053] For example, a crack defect is detected, and its physical area is... The basic risk coefficient is obtained by looking up the table. Set the radius of influence. The area of ​​this region is approximately Three defects, including itself, were detected within this radius, with a total area of... Calculate density ,set up ,but Calculate severity Preset threshold .because The severity value falls within the middle range. The system's final defect assessment level is designated as "key monitoring".

[0054] In summary, this invention achieves deep denoising and feature enhancement of weak defect signals against a complex scattering noise background by acquiring ultrasonic signal sequences and using a convolutional neural network to generate time-frequency feature masks for weighted reconstruction. By constructing B-Scan grayscale images based on scanning positions and combining total energy functional theory with a greedy iterative algorithm to refine the initial contour, it achieves adaptive extraction of defect geometric boundaries and accurate morphological quantification. Furthermore, by fusing morphological features, support vector machine classification results, and defect spatial distribution density, a multi-dimensional risk assessment model is established, enabling intelligent identification of welding defect types and scientific determination of hazard levels. This method effectively solves the problems of low signal-to-noise ratio, blurred defect contours, and single assessment dimensions in existing technologies for detecting weld straightening areas, significantly improving the accuracy and engineering practicality of non-destructive testing of steel structures.

[0055] Reference Figure 2 The second embodiment of the present invention provides a non-destructive testing system for welded straightening areas of steel components based on ultrasound, comprising: The signal acquisition and transformation module is used to control the ultrasonic acquisition device to continuously scan along the welding correction area of ​​the steel component, acquire the original ultrasonic signal sequence, and perform a short-time Fourier transform on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence. The deep enhancement module is used to input the two-dimensional time-frequency matrix sequence into the pre-trained convolutional neural network model one by one to generate the corresponding time-frequency feature mask sequence, and perform frame-by-frame dot multiplication operation between the time-frequency feature mask sequence and the two-dimensional time-frequency matrix sequence, and perform inverse transformation reconstruction to obtain the time-domain enhanced signal sequence. The image construction module is used to map the time-domain enhanced signal sequence into a two-dimensional B-Scan grayscale image based on the scanning position information of the ultrasound acquisition device. The initial contour extraction module is used to calculate the gradient magnitude distribution of the two-dimensional B-Scan grayscale image, determine the region of interest of the defect based on the local gradient maxima in the gradient magnitude distribution, and extract the geometric boundary of the region of interest of the defect as the initial evolution contour. The contour refinement module is used to construct a total energy functional that includes an internal smooth energy term and an external gradient energy term. It uses a greedy iterative algorithm to drive the initial evolved contour to shrink towards the high gradient edge to obtain a refined defect contour. The intelligent classification module is used to calculate the morphological quantification index based on the refined defect contour, input the morphological quantification index into a pre-trained support vector machine classifier, and output the defect type label. The assessment and decision module is used to query a preset decision matrix based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index to determine the final defect assessment level.

[0056] It should be noted that the ultrasonic-based non-destructive testing system for welded straightening areas of steel components provided in this embodiment of the invention is used to execute all the process steps of the ultrasonic-based non-destructive testing method for welded straightening areas of steel components described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, and therefore will not be repeated.

[0057] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as an ultrasonic-based non-destructive testing program for weld straightening areas of steel components. When the processor executes the computer program, it implements the steps in the various embodiments of the ultrasonic-based non-destructive testing method for weld straightening areas of steel components described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above system embodiments, such as the signal acquisition and conversion module.

[0058] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0059] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0060] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0061] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0062] If the modules / units integrated into the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or system capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0063] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0064] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A non-destructive testing method for welded straightening areas of steel components based on ultrasonic waves, characterized in that, include: The ultrasonic acquisition device is controlled to continuously scan along the welding correction area of ​​the steel component to obtain the original ultrasonic signal sequence, and a short-time Fourier transform is performed on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence. The two-dimensional time-frequency matrix sequence is input into a pre-trained convolutional neural network model to generate a corresponding time-frequency feature mask sequence. The time-frequency feature mask sequence is then multiplied frame by frame with the two-dimensional time-frequency matrix sequence, and an inverse transform is performed to reconstruct the time-domain enhanced signal sequence. Based on the scanning position information of the ultrasound acquisition device, the time-domain enhanced signal sequence is mapped into a two-dimensional B-Scan grayscale image; The gradient magnitude distribution of the two-dimensional B-Scan grayscale image is calculated, the region of interest of the defect is determined based on the local gradient maxima of the gradient magnitude distribution, and the geometric boundary of the region of interest of the defect is extracted as the initial evolution contour. A total energy functional containing an internal smooth energy term and an external gradient energy term is constructed. A greedy iterative algorithm is used to drive the initial evolution profile to shrink towards the high gradient edge to obtain a refined defect profile. Based on the refined defect contour, a morphological quantification index is calculated, and the morphological quantification index is input into a pre-trained support vector machine classifier to output a defect type label. Based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index, a preset decision matrix is ​​queried to determine the final defect assessment level.

2. The method for non-destructive testing of welded straightening areas of steel components based on ultrasound according to claim 1, characterized in that, The controlled ultrasonic acquisition device continuously scans along the welding correction area of ​​the steel component to acquire the original ultrasonic signal sequence, and performs a short-time Fourier transform on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence, including: The ultrasound acquisition device is controlled to acquire the original A-scan signal sequence at a preset sampling rate; The Hanning window is selected as the window function, and a preset window length and overlap rate are set. A sliding window short-time Fourier transform is performed on each frame of the original A-scan signal sequence. The amplitude spectrum of the transformation result is calculated to generate the two-dimensional time-frequency matrix sequence, wherein the horizontal axis of the two-dimensional time-frequency matrix represents time, the vertical axis represents frequency, and the matrix element values ​​represent frequency energy distribution.

3. The method for non-destructive testing of welded straightening areas of steel components based on ultrasound according to claim 1, characterized in that, The process of inputting the two-dimensional time-frequency matrix sequence into a pre-trained convolutional neural network model to generate a corresponding time-frequency feature mask sequence, performing frame-by-frame dot-multiplication between the time-frequency feature mask sequence and the two-dimensional time-frequency matrix sequence, and then performing an inverse transform reconstruction to obtain a time-domain enhanced signal sequence includes: The deep time-frequency clustering pattern in the two-dimensional time-frequency matrix sequence is extracted using the convolutional neural network model, and the time-frequency feature mask sequence with values ​​between 0 and 1 is output, wherein the mask value of the region corresponding to the defect echo is close to 1, and the mask value of the noise region is close to 0. The time-frequency feature mask sequence is multiplied element-wise with the two-dimensional time-frequency matrix sequence to suppress broadband noise energy with low mask values, resulting in a weighted time-frequency matrix sequence. Perform a short-time inverse Fourier transform on the weighted time-frequency matrix sequence to restore the phase information and obtain the time-domain enhanced signal sequence with improved signal-to-noise ratio.

4. The non-destructive testing method for welded straightening areas of steel components based on ultrasound according to claim 1, characterized in that, The step of mapping the time-domain enhanced signal sequence into a two-dimensional B-Scan grayscale image based on the scanning position information of the ultrasound acquisition device includes: Perform a Hilbert transform on each frame of the time-domain enhanced signal sequence and calculate the signal envelope amplitude; The signal envelope amplitude is mapped to a gray value of a preset bit depth to obtain a one-dimensional gray vector sequence; Based on the scanning position information, the one-dimensional grayscale vector sequence is spatially arranged and interpolated in the scanning direction to generate the two-dimensional B-Scan grayscale image.

5. The method for non-destructive testing of welded straightening areas of steel components based on ultrasound according to claim 1, characterized in that, The calculation of the gradient magnitude distribution of the two-dimensional B-Scan grayscale image, the determination of the region of interest (ROI) based on the local gradient maxima in the gradient magnitude distribution, and the extraction of the geometric boundary of the ROI as the initial evolution contour include: The gradient magnitude distribution is obtained by calculating the derivatives of the two-dimensional B-Scan grayscale image in the time and depth directions using the gradient operator. Local maxima points are extracted from the gradient magnitude distribution, and a double threshold hysteresis connection algorithm is applied to connect strong and weak edge pixels to form a binarized edge map. A morphological closing operation is performed on the binarized edge map to obtain a closed connected region after filling the gap. The closed connected region is defined as the region of interest of the defect, and the outer contour of the closed connected region is extracted as the initial evolution contour.

6. The method for non-destructive testing of welded straightening areas of steel components based on ultrasound according to claim 1, characterized in that, The construction of the total energy functional, which includes an internal smooth energy term and an external gradient energy term, is used to drive the initial evolution profile to shrink towards the high gradient edge using a greedy iterative algorithm, resulting in a refined defect profile, including: The total energy functional is established, wherein the inner smooth energy term is used to constrain the continuity and curvature of the contour, and the outer gradient energy term is constructed based on the gray-level gradient field of the region of interest of the defect, and is used to attract the contour to the high gradient edge. Calculate the energy change of the total energy functional when the control point on the initial evolution profile moves to a neighborhood position; The displacement vector of the control point is determined according to the principle of energy minimization, and the position of the control point is iteratively updated until the energy change value is less than the preset energy convergence threshold, thereby obtaining the refining defect profile.

7. The method for non-destructive testing of welded straightening areas of steel components based on ultrasound according to claim 1, characterized in that, The step of calculating a morphological quantification index based on the refined defect contour, inputting the morphological quantification index into a pre-trained support vector machine classifier, and outputting a defect type label includes: Traverse the set of boundary points of the refined defect contour, calculate the pixel area and pixel perimeter, and convert them into physical area and physical perimeter by combining the system calibration coefficients; Calculate the shape compactness factor based on the physical area and the physical perimeter; Construct a multidimensional feature vector that includes the physical area, the physical perimeter, and the shape compactness factor; The pre-trained support vector machine classifier with radial basis function kernel is invoked to calculate the decision function value of the multidimensional feature vector. Based on the decision function value, it is determined whether the defect belongs to the crack type or the porosity type, and the defect type label is generated.

8. The non-destructive testing method for welded straightening areas of steel components based on ultrasound according to claim 1, characterized in that, The step of determining the final defect assessment level by querying a preset decision matrix based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index includes: The basic risk coefficient is determined based on the defect type label. If the defect type label is a crack, the basic risk coefficient of the first value is set. Calculate the spatial distribution density of the refining defect profile within the scanning area, and use the spatial distribution density to weight and correct the basic risk coefficient to obtain the corrected risk coefficient. By combining the corrected risk coefficient with the physical size data, a severity value is obtained; The threshold range to which the severity value belongs is queried in the preset decision matrix, and the final defect assessment level, which includes rework, monitoring, or pass / fail conclusions, is obtained.

9. A non-destructive testing system for welded straightening areas of steel components based on ultrasonic waves, characterized in that, include: The signal acquisition and transformation module is used to control the ultrasonic acquisition device to continuously scan along the welding correction area of ​​the steel component, acquire the original ultrasonic signal sequence, and perform a short-time Fourier transform on each frame of the original ultrasonic signal sequence to obtain a two-dimensional time-frequency matrix sequence. The deep enhancement module is used to input the two-dimensional time-frequency matrix sequence into the pre-trained convolutional neural network model one by one to generate the corresponding time-frequency feature mask sequence, and perform frame-by-frame dot multiplication operation between the time-frequency feature mask sequence and the two-dimensional time-frequency matrix sequence, and perform inverse transformation reconstruction to obtain the time-domain enhanced signal sequence. The image construction module is used to map the time-domain enhanced signal sequence into a two-dimensional B-Scan grayscale image based on the scanning position information of the ultrasound acquisition device. The initial contour extraction module is used to calculate the gradient magnitude distribution of the two-dimensional B-Scan grayscale image, determine the region of interest of the defect based on the local gradient maxima in the gradient magnitude distribution, and extract the geometric boundary of the region of interest of the defect as the initial evolution contour. The contour refinement module is used to construct a total energy functional that includes an internal smooth energy term and an external gradient energy term. It uses a greedy iterative algorithm to drive the initial evolved contour to shrink towards the high gradient edge to obtain a refined defect contour. The intelligent classification module is used to calculate the morphological quantification index based on the refined defect contour, input the morphological quantification index into a pre-trained support vector machine classifier, and output the defect type label. The assessment and decision module is used to query a preset decision matrix based on the risk coefficient corresponding to the defect type label and the physical size data in the morphological quantification index to determine the final defect assessment level.

Images