Efficient flaw detection method of ultrasonic flaw detector for automatic detection of welding seam

Abstract

The invention provides an efficient flaw detection method of an ultrasonic flaw detector for automatic detection of a welding seam, and belongs to the technical field of welding seam detection.The efficient flaw detection method comprises the steps that firstly, a welding seam three-dimensional digital model is established, an optimal scanning path is determined, reflection echo signals are processed in real time through an adaptive beam forming algorithm, and defect features are extracted and subjected to quantitative analysis; and establishing a neural network model based on the defect characteristic spectrum, and performing deep learning training by using the sample library. A multi-angle scanning strategy is adopted to implement full-coverage detection, and cross scanning is applied in a key area to improve the detection probability. And calculating the accurate position of the defect by using a time-difference method, and constructing a defect three-dimensional distribution diagram by combining a multi-angle data fusion technology. The optimal filtering parameter is determined through the acoustic feature optimization function, and the defect signal-to-noise ratio is improved. And finally, a quality evaluation report is automatically generated based on a detection result, so that the problems of defect identification accuracy and detection efficiency in weld flaw detection are effectively solved.

Description

Technical Field

[0001] The invention belongs to the technical field of weld detection, and in particular relates to a high-efficiency flaw detection method of an ultrasonic flaw detector for automatic weld detection. Background Art

[0002] Weld quality inspection is a key link in the field of industrial manufacturing. Ultrasonic non-destructive testing has become the mainstream inspection method due to its advantages such as no radiation hazard and large detection depth. Traditional ultrasonic weld flaw detection technology mainly uses a single probe fixed angle scanning method to determine the internal defects of the weld by manually interpreting the ultrasonic echo signal. This method has been widely used in the weld quality assessment of important engineering structures such as pressure vessels, pipelines, and bridges.

[0003] However, traditional ultrasonic flaw detection technology has obvious limitations in complex weld inspection. First, fixed-angle scanning is difficult to fully cover all areas of special-shaped welds, resulting in a high rate of missed defect detection; second, manual interpretation of ultrasonic signals depends on the experience of inspectors, which is highly subjective and has poor consistency; third, a single parameter setting cannot adapt to welds of different thicknesses and material properties, reducing signal quality; finally, traditional signal processing methods have limited noise suppression capabilities, and defect signals are easily masked in high-noise environments.

[0004] In the face of increasing industrial safety requirements, how to achieve efficient and accurate automatic defect recognition under complex weld conditions, improve detection coverage and reduce human interpretation errors has become a core issue that needs to be urgently addressed in current ultrasonic weld flaw detection technology. Existing technologies are difficult to balance detection efficiency and recognition accuracy, especially in welds with complex geometric shapes and in high-noise environments. In other words, the existing technology has technical problems such as low defect recognition accuracy and insufficient detection efficiency during ultrasonic weld flaw detection. Summary of the invention

[0005] In view of this, the present invention provides an efficient flaw detection method of an ultrasonic flaw detector for automated weld detection, which can solve the technical problems in the prior art of low defect recognition accuracy and insufficient detection efficiency during ultrasonic weld flaw detection.

[0006] The present invention is implemented as follows: The present invention provides an efficient flaw detection method of an ultrasonic flaw detector for automatic weld detection, including: establishing a three-dimensional digital model of the weld and determining the optimal scanning path of the ultrasonic probe; configuring phased array ultrasonic probe parameters, setting the focusing depth to two-thirds of the weld thickness, and adjusting the probe incident angle to a first angle value; setting a time domain synchronous acquisition threshold and starting an automatic gain compensation system; applying an adaptive beamforming algorithm to process reflected echo signals and extract defect feature signals; establishing a neural network model according to defect features and performing deep learning training; adopting a multi-angle scanning strategy to implement full coverage detection; calculating the defect position by a time difference method, and constructing a three-dimensional defect distribution map in combination with multi-angle data fusion technology; using an acoustic feature optimization function to determine the optimal filtering parameters and improve the defect signal-to-noise ratio; generating a quality assessment report based on the detection results, and applying the UltraDefectNet model for high-precision defect classification.

[0007] Among them, the optimal scanning path specifically refers to the probe movement trajectory calculated based on the weld geometry and the acoustic properties of the weld material, which enables the ultrasonic beam to penetrate the entire weld area at the optimal angle.

[0008] Among them, the adaptive beamforming algorithm specifically refers to a signal processing method that achieves dynamic focusing and deflection of the ultrasonic beam by adjusting the transmission and reception timing delays of each array element of the phased array ultrasonic probe, thereby obtaining higher resolution and stronger penetration ability.

[0009] The first angle value specifically refers to the incident angle that can produce the strongest shear wave conversion efficiency, which is calculated based on the ratio of the longitudinal wave sound velocity of the weld material to the shear wave sound velocity of the weld material.

[0010] Among them, the acoustic feature optimization function is used to determine the optimal signal processing parameters according to the characteristics of the weld material and the detection environment. The input includes the acoustic impedance value of the weld material, the weld thickness parameters, the background noise spectrum characteristics, the expected defect type indicators and the detection sensitivity requirements. The output is the optimal bandpass filter frequency range, the optimal signal gain curve, the optimal time domain window length and the optimal noise suppression threshold.

[0011] Among them, the optimal bandpass filter frequency range is used to set the upper and lower limits of the passband frequency of the signal filter; the optimal signal gain curve is used to set the gain adjustment parameters of the automatic gain compensation system; the optimal time domain window length is used to set the sampling time of the time domain synchronous acquisition threshold; the optimal noise suppression threshold is used to set the working threshold of the noise suppression system.

[0012] The specific structure of the UltraDefectNet model is a hybrid architecture combining a multi-layer convolutional neural network and a bidirectional long short-term memory network. It includes five convolutional layers for extracting spatial features from ultrasonic reflection echo signals, and three bidirectional long short-term memory network layers for capturing temporal features. It also uses a self-attention mechanism to achieve weight allocation for different frequency components.

[0013] Among them, the number of self-attention heads is related to the acoustic properties of the weld material. The calculation formula for the number of heads is four times the ratio of the longitudinal wave speed of the weld material to the transverse wave speed of the weld material. The scaling factor in the attention mechanism is proportional to the first angle value, and finally the defect type and defect size estimation results are output through the fully connected layer.

[0014] Among them, the steps of establishing the training data set in the UltraDefectNet model training process include collecting typical defect ultrasonic reflection echo signals generated under various welding processes, labeling each defect type and recording the defect size parameters and defect location information, and using data enhancement technology to simulate the signal performance in different noise environments.

[0015] The training data set includes multiple sets of labeled data including porosity defects, crack defects, slag inclusion defects, and unfusion defects. At the same time, a validation set is established for UltraDefectNet model evaluation. The validation set data comes from actual weld inspection cases and has no overlap with the training set, ensuring the generalization ability of the UltraDefectNet model.

[0016] Compared with the prior art, the present invention provides an efficient flaw detection method for an ultrasonic flaw detector for automated weld detection. The efficient flaw detection method proposed by the present invention integrates phased array ultrasonic technology, adaptive signal processing and deep learning algorithms to form a complete set of automated weld detection solutions. This method achieves full coverage detection of welds with complex geometric shapes by establishing a three-dimensional digital model of the weld and calculating the optimal scanning path, combined with the dynamic focusing capability of the phased array ultrasonic probe, and significantly improves the detection coverage rate.

[0017] In order to solve the problems of poor signal quality and inconsistent manual interpretation in traditional technologies, the present invention uses adaptive beamforming algorithms and acoustic feature optimization functions to achieve intelligent adjustment of signal processing parameters, effectively improving the defect signal-to-noise ratio by at least 5 decibels, so that tiny defect signals can be clearly identified even in a high-noise environment. At the same time, the introduced UltraDefectNet deep learning model replaces traditional manual interpretation. Through training on 200,000 sets of annotated data, it achieves a defect detection accuracy rate of more than 95%, and the defect location positioning error is less than 5% of the weld thickness, which greatly reduces the error of manual interpretation.

[0018] Through the combined application of multi-angle scanning strategy and multi-angle data fusion technology, the method of the present invention successfully solves the core technical problems of low defect recognition accuracy and insufficient detection efficiency under complex weld conditions. This method not only realizes the high automation of the detection process and reduces manual intervention, but also automatically generates a weld quality assessment report based on the detection results, providing a more reliable technical guarantee for industrial safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention.

[0020] Figure 2 This is a comparison chart of the detection rates of various weld defects at different incident angles in Example 2.

[0021] Figure 3 This is a spectrum comparison diagram of the ultrasonic signal before and after processing in Example 2. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0023] like Figure 1 FIG. 1 is a flow chart of an efficient flaw detection method of an ultrasonic flaw detector for automatic weld detection provided by the present invention. The method comprises the following steps:

[0024] S01. Establish a three-dimensional digital model of the weld and perform meshing processing, and determine the optimal scanning path of the ultrasonic probe through spatial modeling;

[0025] S02, configuring the array element parameters of the phased array ultrasonic probe, setting the focus depth to two-thirds of the weld thickness, and adjusting the probe incident angle to a first angle value;

[0026] S03, setting the time domain synchronous acquisition threshold and starting the automatic gain compensation system, and implementing dynamic gain control for weld areas of different thicknesses;

[0027] S04. Apply adaptive beamforming algorithm to process reflected echo signals in real time, extract defect characteristic signals and conduct multi-dimensional quantitative analysis;

[0028] S05. Establish a defect recognition neural network model based on the frequency spectrum characteristics of defect characteristic signals, and use a known defect sample library for deep learning training;

[0029] S06. Use multi-angle scanning strategy to implement full coverage detection, and use cross scanning in key areas to increase the probability of detection;

[0030] S07. Calculate the precise location of defects by using the time difference method and construct a three-dimensional distribution map of weld defects by combining multi-angle data fusion technology;

[0031] S08. Use the acoustic feature optimization function to determine the optimal filtering parameters, eliminate structural noise and scattering interference, and improve the defect signal-to-noise ratio by at least 5 decibels;

[0032] S09. Compare and analyze the defect type and defect size with standard specifications, and apply the pre-trained UltraDefectNet model for high-precision defect classification; generate a weld quality assessment report.

[0033] Among them, the optimal scanning path specifically refers to the probe movement trajectory calculated based on the weld geometry and the acoustic properties of the weld material, which enables the ultrasonic beam to penetrate the entire weld area at the optimal angle.

[0034] Among them, the adaptive beamforming algorithm specifically refers to a signal processing method that achieves dynamic focusing and deflection of the ultrasonic beam by adjusting the transmission and reception timing delays of each array element of the phased array ultrasonic probe, thereby obtaining higher resolution and stronger penetration ability.

[0035] Among them, multi-angle data fusion technology specifically refers to the computational processing technology that integrates ultrasonic reflection data obtained at different incident angles through spatial mapping and probability statistics methods to generate more complete and accurate weld defect morphology information.

[0036] Among them, the first angle value specifically refers to the incident angle that can produce the strongest shear wave conversion efficiency calculated based on the ratio of the longitudinal wave sound velocity of the weld material to the shear wave sound velocity of the weld material, usually in the range of 45 degrees to 70 degrees.

[0037] The time domain synchronous acquisition threshold specifically refers to the amplitude setting value that triggers the data acquisition system to start recording when the amplitude of the ultrasonic signal exceeds a certain multiple of the background noise, which is used to filter out invalid signals and improve system efficiency.

[0038] Among them, the acoustic feature optimization function is used to determine the optimal signal processing parameters according to the characteristics of the weld material and the detection environment. The input includes the acoustic impedance value of the weld material, the weld thickness parameter, the background noise spectrum characteristics, the expected defect type index and the detection sensitivity requirements. The output is the optimal bandpass filter frequency range, the optimal signal gain curve, the optimal time domain window length and the optimal noise suppression threshold; the optimal bandpass filter frequency range is used to set the upper and lower limits of the passband frequency of the signal filter; the optimal signal gain curve is used to set the gain adjustment parameters of the automatic gain compensation system; the optimal time domain window length is used to set the sampling time of the time domain synchronous acquisition threshold; the optimal noise suppression threshold is used to set the working threshold of the noise suppression system.

[0039] Among them, the specific structure of the UltraDefectNet model is a hybrid architecture combining a multi-layer convolutional neural network and a bidirectional long short-term memory network, which includes five convolutional layers for extracting spatial features in ultrasonic reflection echo signals, and three bidirectional long short-term memory network layers for capturing temporal features. The self-attention mechanism is used to realize the weight allocation of different frequency components. The number of self-attention heads is related to the acoustic properties of the weld material. The formula for calculating the number of heads is four times the ratio of the longitudinal wave speed of the weld material to the transverse wave speed of the weld material. The scaling factor in the attention mechanism is proportional to the first angle value. Finally, the defect type and defect size estimation results are output through the fully connected layer. The UltraDefectNet model adopts a residual connection structure to improve the stability of deep network training, and applies batch normalization after each layer of convolution to improve the training speed.

[0040] Among them, the steps of establishing the training data set in the UltraDefectNet model training process specifically include collecting ultrasonic reflection echo signals of typical defects generated under various welding processes, annotating each defect type and recording the defect size parameters and defect location information, using data enhancement technology to simulate the signal performance under different noise environments, and constructing a total of 200,000 sets of annotated data including various types of porosity defects, crack defects, slag inclusion defects and unfusion defects. At the same time, a validation set is established for UltraDefectNet model evaluation. The validation set data comes from actual weld inspection cases and has no overlap with the training set to ensure the generalization ability of the UltraDefectNet model.

[0041] Among them, the steps of UltraDefectNet model training specifically include first performing initial training in a small batch gradient descent manner, setting the learning rate to the inverse of the weld thickness parameter and gradually decaying it, performing a verification test every fifty training cycles and recording the UltraDefectNet model performance indicators, and adopting a learning rate adjustment strategy when there is no significant improvement in the performance of three consecutive verification tests. During the training process, weld defect detection accuracy and defect location positioning accuracy are used as joint loss function indicators, and the UltraDefectNet model parameter update is implemented using an adaptive moment estimation algorithm. The training process is completed when the UltraDefectNet model reaches a detection accuracy of more than 95% and the defect location positioning error is less than 5% of the weld thickness parameter.

[0042] The specific implementation of the above steps is described in detail below.

[0043] The specific implementation method of step S01 is to first collect the weld shape parameters, including the geometric dimension information such as weld length, width, thickness, groove angle, etc., and then use three-dimensional modeling software to build an accurate digital model of the weld. Then, the finite element meshing technology is used to refine the weld model, and the mesh size of the weld area is usually set to 1 / 8 to 1 / 10 of the ultrasonic wavelength to ensure the calculation accuracy. Subsequently, the propagation path of the ultrasonic wave in the weld material is simulated based on the ray tracing algorithm, while considering the acoustic impedance distribution of the weld material, and calculating the coverage area of ​​the ultrasonic beam at each incident point and incident angle. Finally, the A* path planning algorithm is used to determine the optimal scanning path of the ultrasonic probe, which can ensure that the ultrasonic beam penetrates the entire weld area at the optimal angle and the scanning time is the shortest. The purpose of this step is to provide accurate probe motion control instructions for subsequent flaw detection processes through digital modeling technology, thereby improving detection efficiency and defect detection rate.

[0044] The specific implementation method of step S02 is to first set the number of array elements of the phased array ultrasonic probe according to the weld thickness parameters. Usually, a 32-element probe is used for welds with a thickness of less than 10 mm, a 64-element probe is used for welds with a thickness of 10 mm to 30 mm, and a 128-element probe is used for welds with a thickness of more than 30 mm. Then, the position of two-thirds of the weld thickness is calculated as the ultrasonic focusing depth. This depth setting can achieve better near-surface defect detection capabilities while ensuring the sensitivity of weld root detection. Then, the first angle value is calculated based on the ratio of the longitudinal wave velocity to the shear wave velocity of the weld material. The calculation formula is: ,in is the shear wave speed, is the longitudinal wave speed, and the angle value is usually in the range of 45° to 70°. Finally, the transmission delay time of each array element is adjusted through the phased array control system so that the ultrasonic beam is incident on the weld at the first angle value. The purpose of this step is to ensure that the ultrasonic wave can enter the weld at the best angle and focusing state, thereby improving the sensitivity of defect detection.

[0045] The specific implementation method of step S03 is to first measure the amplitude of the weld background noise, and then set the time domain synchronous acquisition threshold to 2.5 to 3.5 times the average amplitude of the background noise. This setting can effectively filter invalid signals while not missing small-amplitude defect echoes. Then start the automatic gain compensation system, which automatically adjusts the gain parameters according to the attenuation law of sound waves in the material. For carbon steel welds, the typical attenuation compensation value is 0.5 to 2dB / mm; for stainless steel welds, the typical attenuation compensation value is 0.3 to 1.5dB / mm; for aluminum alloy welds, the typical attenuation compensation value is 0.2 to 1dB / mm. Then, for different thickness areas of the weld, a regional dynamic gain control strategy is adopted, that is, a low gain setting is used in the thin weld area, and a high gain setting is used in the thick weld area. The gain difference is proportional to the thickness change, and the proportional coefficient is usually 0.8 to 1.2dB / mm. Finally, a mapping relationship table between weld thickness and gain parameters is established for the system to query and call in real time during the scanning process. The purpose of this step is to improve the adaptability of the flaw detection system to weld areas of different thicknesses by optimizing signal acquisition parameters and ensure the consistency of detection sensitivity.

[0046] The specific implementation method of step S04 is to first apply the delay-sum beamforming basic algorithm to obtain the initial echo signal, and then use the minimum variance distortion-free response beamforming algorithm to adaptively optimize the signal. The algorithm can automatically suppress interference signals from non-target directions and improve the signal-to-noise ratio of the target defect signal. Then, the optimized echo signal is envelope detected to extract characteristic parameters such as signal amplitude, phase, and frequency. Subsequently, the wavelet transform is applied to perform time-frequency analysis on the signal, and the dB4 wavelet basis function is used for 5-layer decomposition to extract the energy distribution characteristics of each frequency band. Then, the extracted features are reduced in dimension by combining the principal component analysis method, and the principal component with a contribution rate of 95% is retained as the defect feature vector. Finally, the defect feature vector is quantized and encoded to form a standardized defect feature descriptor for subsequent defect identification and classification. The purpose of this step is to extract valuable defect feature information from complex echo signals through advanced signal processing algorithms to provide a reliable data basis for defect identification.

[0047] The specific implementation method of step S05 is to first analyze the spectrum distribution of the defect characteristic signal and extract the characteristic parameters such as peak frequency, valley frequency, bandwidth, energy distribution, etc. in the spectrum. Then, a multi-layer perceptron neural network model is constructed, the number of nodes in the input layer is consistent with the characteristic dimension, the hidden layer adopts a 3-5 layer structure, the number of nodes in each layer is 1.5-2 times that of the input layer, and the number of nodes in the output layer is consistent with the number of defect types. Then, the known defect sample library data is imported, which contains characteristic signal data of typical welding defects such as pores, cracks, slag inclusions, and unfusion, and each type of defect contains at least 1000 groups of samples. Subsequently, the back propagation algorithm is used to train the neural network, the initial value of the learning rate is set to 0.01, and the cosine annealing strategy is used to gradually reduce the learning rate. Finally, the model performance is evaluated by the cross-validation method, and the model training is completed when the recognition accuracy reaches more than 90%. The purpose of this step is to establish a defect recognition model based on spectral features and use machine learning methods to improve the accuracy and stability of defect classification.

[0048] The specific implementation method of step S06 is to first design a multi-angle scanning strategy based on the first angle value, and usually use the first angle value ±5°, ±10°, ±15°, a total of 7 angles for scanning. Then the weld is segmented longitudinally, and the length of each segment is 70% to 80% of the total length of the probe array element, ensuring that there is a 20% to 30% overlap area between adjacent scanning segments. Then, cross-scanning is performed on key areas of the weld, such as the weld root and heat-affected zone, that is, a path perpendicular to the conventional scanning direction is used for supplementary inspection, and the spacing of the cross-scanning is usually 50% of the length of the probe array element. Subsequently, the data of each scanning point is collected and preprocessed in real time to ensure that the data quality meets the inspection requirements. Finally, a weld full coverage inspection database is established to record the inspection results of each position at different angles. The purpose of this step is to improve the defect detection rate through a multi-angle, full coverage scanning strategy, especially for directional defects such as cracks, multi-angle scanning can significantly improve the detection probability.

[0049] The specific implementation method of step S07 is to first perform time delay correction on the collected multi-angle reflected echo signals to eliminate the time deviation caused by the difference in sound path at different angles. Then the time difference method is used to calculate the defect position, and the calculation formula is: ,in is the defect depth, is the speed of sound, is the round trip time of the sound wave. Then, the lateral position of the defect is determined by the principle of triangulation, and the optimal position estimate of the defect is calculated by combining the measurement results at different angles. The multi-angle data is then subjected to Bayesian fusion processing, that is, weights are assigned according to the reliability of the data at each angle, and a probability distribution model of the defect position is constructed. Finally, a three-dimensional distribution map of weld defects is constructed based on the fused data, and three-dimensional voxel rendering technology is used to intuitively display the spatial distribution of defects in the weld. The purpose of this step is to improve the defect location accuracy through multi-angle data fusion, and generate intuitive defect distribution visualization results, providing a reliable basis for weld quality assessment.

[0050] The specific implementation method of step S08 is to first determine the input parameters of the acoustic feature optimization function, including the acoustic impedance value of the weld material, the weld thickness parameter, the background noise spectrum characteristics, the expected defect type index and the detection sensitivity requirements. Then the genetic algorithm is used to solve the optimal filter parameters, the population size is set to 50, the number of iterations is set to 100, the crossover probability is set to 0.8, and the mutation probability is set to 0.1. Then the bandpass filter parameters are set according to the optimization results. Usually, for a 2.5MHz center frequency probe, the optimal bandpass range is 1.5MHz to 3.5MHz; for a 5MHz center frequency probe, the optimal bandpass range is 3MHz to 7MHz. Then the adaptive Wiener filtering algorithm is applied to further suppress the structural noise. The algorithm can automatically adjust the filter coefficient according to the power spectral density ratio of the signal to the noise. Finally, the optimization effect is verified by comparing the signal-to-noise ratio before and after signal processing. The signal-to-noise ratio after optimization should be no less than 5dB. The purpose of this step is to effectively eliminate structural noise and scattering interference by optimizing the signal processing parameters, and improve the clarity and recognizability of the defect signal.

[0051] The specific implementation method of step S09 is to first collect defect information of all detection positions, including defect type, location, size and other parameters. Then load the pre-trained UltraDefectNet model, which can more accurately identify and classify weld defects based on deep learning algorithms. Then compare the defect parameters with standard specifications such as GB / T5777 to evaluate the severity and allowability of each defect. Subsequently, the overall quality grade of the weld is calculated based on the evaluation results, which is usually divided into four levels: level 1 (excellent), level 2 (good), level 3 (qualified) and unqualified. Finally, a weld quality assessment report is automatically generated, including a defect distribution map, defect type statistics, severe defect warnings and overall quality evaluation conclusions. The purpose of this step is to objectively evaluate the quality of the weld based on the test results and standard specifications, and generate a standardized quality assessment report to provide a basis for welding process improvement and quality control.

[0052] The detailed structure of the UltraDefectNet model is a hybrid architecture that combines a multi-layer convolutional neural network with a bidirectional long short-term memory network. The input layer of the model receives the preprocessed ultrasonic reflection echo signal, and the signal length is usually 1024 to 4096 sampling points. The first convolutional layer contains 32 Convolution kernel, step size is 1, using ReLU activation function, followed by The second convolutional layer contains 64 Convolution kernel, step size is 1, using ReLU activation function, followed by The third convolutional layer contains 128 Convolution kernel, step size is 1, using ReLU activation function, followed by The fourth convolutional layer contains 256 Convolution kernel, step size is 1, using ReLU activation function, followed by The fifth convolutional layer contains 512 Convolution kernel, step size is 1, using ReLU activation function, followed by Max pooling layer. These five convolutional layers are used to extract spatial features from ultrasonic reflection echo signals, such as waveform morphology, peak distribution, etc. The output of the convolutional layer is connected to three bidirectional long short-term memory network layers, each of which contains 256 memory units to capture temporal features, such as the time correlation and frequency change trend of the signal. The self-attention mechanism is applied after the bidirectional long short-term memory network layer. The number of heads is calculated as four times the ratio of the longitudinal wave sound velocity of the weld material to the shear wave sound velocity of the weld material. For a typical carbon steel weld, the number of heads is usually 8 to 12. The scaling factor in the self-attention mechanism is proportional to the first angle value, and the calculation formula is the first angle value divided by 10. The output of the self-attention layer is connected to three fully connected layers, with the number of nodes being 512, 256, and 128 respectively. The last fully connected layer is connected to the softmax classifier to output the probability distribution of the defect type, while the other branch outputs the defect size estimation result through the regressor. Batch normalization technology is applied after each convolution layer in the model to accelerate the training process, and a residual connection structure is added between deep networks to avoid the gradient vanishing problem and improve training stability.

[0053] The specific implementation method of establishing the training data set of the UltraDefectNet model is to first collect the ultrasonic reflection echo signals of typical defects generated under various welding processes, including typical defect signals such as porosity defects, crack defects, slag inclusion defects, and unfusion defects generated under processes such as manual arc welding, submerged arc welding, gas shielded welding, and laser welding. For each defect, ultrasonic tomography technology is used to accurately measure the defect size parameters, and the defect type is confirmed by combining metallographic analysis methods to establish accurate annotation information. Then, data enhancement technology is used to expand the training samples, including adding Gaussian noise of different intensities, random time offset, amplitude scaling, frequency modulation, etc., to simulate the signal performance under different noise environments and detection conditions. Through these data enhancement methods, the original defect signal is expanded to 200,000 sets of annotated data, of which porosity defects account for 30%, crack defects account for 25%, slag inclusion defects account for 20%, unfusion defects account for 15%, and other types of defects account for 10%. Then the training set and the validation set are divided. The validation set data comes from actual weld inspection cases and has no overlap with the training set. Usually, the ratio of the training set to the validation set is 8:2. Finally, the data set is standardized to make the signals from different sources comparable. The standardization method uses z-score standardization, that is, ,in is the signal mean, is the signal standard deviation.

[0054] The specific implementation method of UltraDefectNet model training is to first initialize the network parameters. The convolution layer parameters use the He initialization method, and the recurrent neural network layer parameters use the orthogonal initialization method. Then set the learning rate to the inverse of the weld thickness parameter. For 20mm thick welds, the initial learning rate is set to 0.05. Then use the small batch gradient descent method for training. The batch size is set to 64, and each batch of data is randomly extracted from the training set. During the training process, the cross entropy loss function is used to evaluate the defect classification performance, and the root mean square error loss function is used to evaluate the defect size estimation performance. The two form a joint loss function in a ratio of 7:3. A verification test is performed every fifty training cycles, and the classification accuracy and positioning accuracy on the verification set are calculated. When the performance improvement of three consecutive verification tests is less than 0.5%, the learning rate is reduced to half of the current value and training continues. Parameter update is implemented using the adaptive moment estimation algorithm. Set to 0.9, Set to 0.999, Set as The training process continues until the model reaches a detection accuracy of more than 95% on the validation set and the defect location error is less than 5% of the weld thickness parameter, or the maximum number of training cycles is 500. The resulting UltraDefectNet model can perform high-precision classification and size estimation of weld defects, providing a reliable basis for weld quality assessment.

[0055] The mathematical model or calculation process involved in the present invention is described in detail below.

[0056] The ray tracing algorithm in step S01 involves the calculation of the reflection and refraction of ultrasound at the material interface, which is specifically expressed as follows:

[0057] ;

[0058] In the formula, is the angle of incidence; is the refraction angle; is the speed of sound in the incident medium; is the speed of sound in the refracting medium.

[0059] This equation is based on Snell's law and describes the refraction behavior of sound waves at the interface of materials with different acoustic impedances. Determined by the probe position and the normal direction of the weld surface, the incident medium sound velocity Usually it is the longitudinal wave speed, which can be measured by the pulse echo method. Select the corresponding material sound velocity value according to the required wave type (longitudinal wave or shear wave). This equation is used to accurately calculate the propagation path of ultrasonic waves in the weld, providing a theoretical basis for determining the optimal scanning path. This equation takes into account the coupling effect of the acoustic properties and geometric relationship of the material, and can accurately predict the propagation behavior of ultrasonic waves in complex weld structures.

[0060] The evaluation function of the A* path planning algorithm is expressed as follows:

[0061] ;

[0062] In the formula, For Node The total evaluation function value of ; From the starting point to the node the actual cost of For slave nodes The estimated cost to reach the destination.

[0063] in, The calculation method is:

[0064] ;

[0065] In the formula, For Node The parent node of From parent node to node The cost.

[0066] and The calculation method is:

[0067] ;

[0068] In the formula, For Node Euclidean distance to the target point; For Node The weld coverage at the position ranges from 0 to 1; and is the weight coefficient, usually The value range is 0.3~0.7. The value range is 0.3 to 0.7, and .

[0069] This evaluation function combines the two factors of path length and weld coverage. By adjusting the weight coefficient, the relationship between detection efficiency and coverage can be balanced. The use of this equation can generate the optimal scanning path to ensure that the ultrasonic beam fully covers the weld, while minimizing the probe movement distance and improving detection efficiency.

[0070] The calculation formula of the first angle value in step S02 is:

[0071] ;

[0072] In the formula, is the first angle value, i.e., the incident angle that produces the strongest shear wave conversion efficiency; is the shear wave speed; is the longitudinal wave speed.

[0073] For typical carbon steel materials, About 5900~6100m / s, About 3200~3300m / s, calculated Usually it is in the range of 32° to 34°. This angle needs to take into account the sound velocity measurement error and introduce a correction term:

[0074] ;

[0075] In the formula, It is a correction item with a value range of -2° to 2°, which is determined through actual tests.

[0076] This equation is based on the shear wave critical angle theory and is used to determine the incident angle that produces the strongest shear wave. At this angle, the energy efficiency of converting incident longitudinal waves into shear waves is the highest, which is beneficial to improving the sensitivity of defect detection. The accurate calculation of this angle is crucial for the parameter setting of phased array ultrasonic probes and directly affects the performance of the detection system.

[0077] The calculation formula of the phased array ultrasonic probe element delay time is:

[0078] ;

[0079] In the formula, For the The delay time of each array element; For the The distance from the array element to the focal point; is the distance from the reference array element to the focal point; is the speed of sound in the wedge.

[0080] in, The calculation method is:

[0081] ;

[0082] In the formula, For the The array element Coordinates on the axis; and The focus points are and Coordinates on the axis.

[0083] When considering the refraction at the weld interface, the above formula needs to be modified to:

[0084] ;

[0085] In the formula, For the The distance from the array element to the incident point on the weld surface; is the distance from the incident point to the focal point on the weld surface; is the distance from the reference array element to the incident point on the weld surface; is the distance from the incident point to the focal point on the weld surface; is the speed of sound in the weld material.

[0086] This equation is based on the Fermat minimum time principle and is used to calculate the delay time of each array element of the phased array ultrasonic probe to achieve the focusing and deflection of the ultrasonic beam. By accurately controlling the transmission and reception timing of each array element, dynamic focusing can be achieved to improve the detection resolution and sensitivity. This equation takes into account the difference in the propagation speed of sound waves in different media and can accurately control the formation process of the ultrasonic beam.

[0087] The gain value calculation formula of the automatic gain compensation system in step S03 is:

[0088] ;

[0089] In the formula, The depth is The gain value at , in dB; is the initial gain value, usually set to 20-30dB; It is the first-order coefficient, which indicates the attenuation coefficient of sound waves in the material, and its unit is dB / mm; It is the quadratic coefficient, which is used to compensate for the scattering loss caused by the increase in depth. The unit is dB / mm², and the value is usually 0.001~0.01dB / mm²; is the sound wave propagation depth, in mm.

[0090] For different materials, Typical values ​​are: carbon steel 0.5 ~ 2dB / mm, stainless steel 0.3 ~ 1.5dB / mm, aluminum alloy 0.2 ~ 1dB / mm.

[0091] This equation is based on the attenuation law of sound waves in materials, taking into account two mechanisms: absorption attenuation and scattering attenuation. The first-order term represents the absorption of sound energy by the material, which is related to the material properties and frequency; the second-order term represents the scattering loss caused by the increase in the sound wave propagation path. This equation can achieve dynamic compensation for the echo signals of defects at different depths to ensure the consistency of detection sensitivity.

[0092] The relationship between the gain difference and thickness change of the regional dynamic gain control is expressed as:

[0093] ;

[0094] In the formula, is the gain difference, in dB; is the proportionality coefficient, the unit is dB / mm, and the value range is 0.8~1.2dB / mm; is the thickness change value, in mm.

[0095] This equation describes the linear relationship of gain adjustment in different thickness areas. Through this equation, the gain parameters can be automatically adjusted according to the change of weld thickness to ensure the consistency of detection sensitivity. It needs to be determined through experiments based on the actual material and equipment characteristics, and its value affects the system's adaptability to thickness changes. The application of this equation can solve the problem of inconsistent detection sensitivity caused by uneven weld thickness.

[0096] The weight vector calculation formula of the minimum variance distortion-free response beamforming algorithm in step S04 is:

[0097] ;

[0098] In the formula, is the weight vector with dimension , is the number of array elements; is the covariance matrix of the received signal, with dimension ; is the guidance vector in the target direction, with dimension ; for The conjugate transpose of .

[0099] Covariance matrix The calculation method is:

[0100] ;

[0101] In the formula, is the signal vector received by each array element, with dimension ; Represents the expected operation.

[0102] In practical applications, the sample covariance matrix is ​​usually used to estimate :

[0103] ;

[0104] In the formula, is the number of sampling points, usually ranging from 100 to 1000.

[0105] Steering vector The calculation method is:

[0106] ;

[0107] In the formula, is the angular frequency; For the The time delay of an array element relative to a reference array element.

[0108] The algorithm can effectively suppress noise and interference signals and improve the signal-to-noise ratio of target defect signals by minimizing the output power while keeping the gain in the target direction constant. The equation makes full use of the principle of spatial filtering and achieves optimal suppression of interference in non-target directions by adaptively adjusting the weights of each array element signal. The algorithm has excellent performance in complex noise environments and is a key technology for achieving high-resolution and high-signal-to-noise ratio defect detection.

[0109] The calculation formula of wavelet transform is:

[0110] ;

[0111] In the formula, is the wavelet transform coefficient; is the scale parameter, which controls the expansion and contraction of the wavelet; is the translation parameter, which controls the position of the wavelet; is the signal to be analyzed; is the wavelet basis function; express The complex conjugate of .

[0112] The dB4 wavelet basis function is characterized by having four vanishing moments, which is suitable for analyzing the mutation characteristics in non-stationary signals and can effectively extract the characteristic information in the defect echo signal. By analyzing the signal at different scales, the wavelet transform can obtain time domain and frequency domain information at the same time, which is particularly suitable for analyzing non-stationary signals such as ultrasonic defect echoes. The transform can separate different frequency components in the signal, making it easier to extract defect characteristics.

[0113] The calculation formula of eigenvalue and eigenvector of principal component analysis method is:

[0114] ;

[0115] In the formula, is the covariance matrix of the feature data; For the eigenvalues; is the corresponding feature vector.

[0116] Covariance matrix The calculation method is:

[0117] ;

[0118] In the formula, For the Feature samples; is the mean vector of all samples; is the sample size.

[0119] The contribution rate of the principal component is calculated as follows:

[0120] ;

[0121] In the formula, For the The contribution rate of the principal components; is the total number of features.

[0122] Select the one that meets the conditions Principal components:

[0123] ;

[0124] In the formula, is the number of principal components that need to be retained.

[0125] The calculation formula of the eigenvector after dimensionality reduction is:

[0126] ;

[0127] In the formula, is the feature vector after dimensionality reduction; For the former The projection matrix composed of eigenvectors; is the original feature vector.

[0128] The principal component analysis method converts possibly correlated variables into linearly uncorrelated variables through orthogonal transformation, finds the main features in the data, and achieves feature dimensionality reduction. This method can effectively reduce the dimension of the feature vector and improve the efficiency and accuracy of subsequent defect identification.

[0129] The loss function of the neural network model in step S05 is:

[0130] ;

[0131] In the formula, is the cross entropy loss function value; is the sample size; is the number of defect categories; For sample Belongs to category The true label (0 or 1); Predict samples for the model Belongs to category probability.

[0132] This loss function is used to evaluate the difference between the model's prediction and the true label. The smaller the value, the better the model performance. The cross entropy loss function has a large penalty for misclassification and can effectively promote the model's learning process.

[0133] The calculation formula for weight update in the back-propagation algorithm is:

[0134] ;

[0135] In the formula, For the Layer Neuron to Layer The connection weights of neurons; is the learning rate; is the loss function Weight The partial derivative of .

[0136] The calculation of partial derivatives involves the chain rule, and the weight parameters are updated by back-propagating the error. This equation is the core of neural network training, and the model output gradually approaches the true label by iteratively adjusting the weight parameters.

[0137] The learning rate is adjusted using the cosine annealing strategy, and the calculation formula is:

[0138] ;

[0139] In the formula, For the The learning rate of the iteration; is the minimum learning rate, usually set to 0.01 times the initial learning rate; is the maximum learning rate, that is, the initial learning rate; is the current iteration number; is the total number of iterations.

[0140] The cosine annealing strategy can use a larger learning rate in the early stage of training to quickly approach the optimal solution, and use a smaller learning rate in the later stage of training to fine-tune the parameters, effectively avoiding the oscillation problem caused by too large a learning rate and the slow convergence problem caused by too small a learning rate. This strategy can balance the training speed and model performance, and improve the efficiency and effect of neural network training.

[0141] The formula for calculating the defect depth using the time difference method in step S07 is:

[0142] ;

[0143] In the formula, is the defect depth, in mm; is the propagation speed of sound waves in the material, in mm / μs; is the round trip time of the sound wave, in μs.

[0144] This formula is based on the principle of ultrasonic propagation. By measuring the time from the emission to the reception of the ultrasonic wave, the depth position of the reflection point is calculated. The coefficient 2 indicates that the sound wave needs to travel back and forth. This equation is the basic method for locating defects in ultrasonic testing. It is simple and effective.

[0145] The calculation formula for determining the lateral position of a defect based on the triangulation principle is:

[0146] ;

[0147] In the formula, is the lateral distance of the defect relative to the incident point, in mm; is the defect depth, in mm; is the incident angle of the ultrasonic beam.

[0148] When the refraction effect is taken into account, the formula is modified to:

[0149] ;

[0150] In the formula, is the refraction angle, calculated according to Snell's law: , and are the speed of sound in the incident medium and the refracting medium, respectively, is the angle of incidence.

[0151] The triangulation principle uses geometric relationships to determine the lateral position of the defect, and combined with depth information, the two-dimensional coordinates of the defect can be determined. This equation takes into account the geometric characteristics of the sound wave propagation path and can accurately locate the defect position.

[0152] The Bayesian weight calculation formula for multi-angle data fusion is:

[0153] ;

[0154] In the formula, For the The weight of the angle data; For the The measurement uncertainty of the angle data; is the number of angles.

[0155] Measurement uncertainty Related to the signal-to-noise ratio, the calculation formula is:

[0156] ;

[0157] In the formula, is the proportionality coefficient, which is usually 1; For the The signal-to-noise ratio of the angle data.

[0158] The calculation formula of the defect position coordinates after fusion is:

[0159] ;

[0160] In the formula, is the defect position coordinate after fusion; For the The defect position coordinates are measured at 3 angles.

[0161] Multi-angle data fusion technology improves the accuracy of defect location by weight distribution and weighted averaging according to the reliability of data from each angle. This method can effectively utilize complementary information obtained from multiple angles, reduce the uncertainty of single angle measurement, and improve the reliability of defect location.

[0162] The expression of the acoustic feature optimization function in step S08 is:

[0163] ;

[0164] In the formula, To optimize the objective function value; is a signal processing parameter vector, including bandpass filter frequency range, signal gain curve parameters, time domain window length and noise suppression threshold; For parameters The corresponding signal-to-noise ratio; For parameters The corresponding resolution; For parameters The corresponding computational complexity; , and are weight coefficients, which respectively represent the importance of signal-to-noise ratio, resolution and computational complexity. The value ranges from 0.5 to 0.7. The value ranges from 0.2 to 0.4. The value is 0.1~0.2, and .

[0165] This function takes into account the signal processing effect and computational efficiency, aiming to find the best signal processing parameters, achieve high signal-to-noise ratio and high resolution while maintaining reasonable computational complexity. By optimizing this function with a genetic algorithm, the signal processing parameters that best suit the current detection conditions can be obtained. The design of this function reflects the balance between effect and efficiency in engineering applications.

[0166] The frequency domain expression of the adaptive Wiener filtering algorithm is:

[0167] ;

[0168] In the formula, is the filter frequency response function; is the power spectrum density of the defect signal; is the power spectral density of the noise.

[0169] The calculation formula for the signal-to-noise ratio improvement is:

[0170] ;

[0171] In the formula, is the signal-to-noise ratio improvement value, in dB; is the signal-to-noise ratio after filtering; is the signal-to-noise ratio before filtering.

[0172] The adaptive Wiener filter algorithm automatically adjusts the filter coefficient according to the power spectral density ratio of the signal and the noise, and can suppress the noise to the greatest extent while retaining the target signal. This algorithm is the optimal linear filter under the minimum mean square error criterion, and is suitable for the case where the noise and signal overlap in the spectrum. This equation is widely used in signal processing and can effectively improve the signal-to-noise ratio of weld defect signals.

[0173] The calculation formula for z-score standardization is:

[0174] ;

[0175] In the formula, is the standardized signal; is the original signal; is the signal mean; is the signal standard deviation.

[0176] in, and The calculation method is:

[0177] ;

[0178] ;

[0179] In the formula, is the number of signal sample points; For the The value of the sample points.

[0180] This standardization method converts the signal into a form with a mean of 0 and a standard deviation of 1, eliminating the scale differences of signals from different sources and making the data comparable. Standardization is crucial for machine learning algorithms such as neural networks, which can accelerate model convergence and improve model generalization capabilities.

[0181] The parameter update formula of the adaptive moment estimation algorithm is:

[0182] ;

[0183] ;

[0184] ;

[0185] ;

[0186] ;

[0187] In the formula, and are the first-order moment and second-order moment estimates, respectively; is the current gradient; and is the momentum coefficient, The value is usually 0.9. The value is usually 0.999; and is the bias-corrected moment estimate; is the updated parameter; is the learning rate; is a small constant to prevent division by zero errors, usually set to .

[0188] The adaptive moment estimation algorithm combines the advantages of the momentum method and the adaptive learning rate method, and can automatically adjust the learning rate of each parameter to accelerate the training process of the neural network. The algorithm realizes adaptive adjustment of the direction and step size of parameter update by accumulating the first-order and second-order moments of the gradient, effectively solving common problems in deep neural network training, such as difficulty in learning rate selection and slow convergence.

[0189] Among them, the calculation formula for the number of heads of the self-attention mechanism in the UltraDefectNet model is:

[0190] ;

[0191] In the formula, is the number of heads of the self-attention mechanism; is the longitudinal wave velocity of the weld material; is the shear wave velocity of the weld material.

[0192] This formula designs the correlation between the number of self-attention heads and the acoustic properties of the material, so that the model structure can adapt to the characteristics of different materials. For a typical carbon steel weld, the number of heads is usually 8 to 12. This design enables the model to better capture the characteristics of ultrasonic reflection signals in different materials and improve the adaptability and generalization ability of the model.

[0193] Among them, the calculation formula of the scaling factor in the self-attention mechanism is:

[0194] ;

[0195] In the formula, is the scaling factor in the self-attention mechanism; is the first angle value.

[0196] This formula establishes a proportional relationship between the scaling factor and the first angle value, so that the self-attention mechanism can automatically adjust the intensity of attention distribution according to the detection angle. This design takes into account the impact of the ultrasonic incident angle on the characteristics of the reflected signal and can improve the model's ability to process data at different angles.

[0197] Among them, the initial value setting formula of the learning rate in UltraDefectNet model training is:

[0198] ;

[0199] In the formula, is the initial learning rate; It is the weld thickness parameter, the unit is mm.

[0200] This formula associates the learning rate with the weld thickness, taking into account the impact of the physical properties of the inspection object on model training. For a 20 mm thick weld, the initial learning rate is set to 0.05. This design can automatically adjust the training parameters according to the characteristics of the inspection object and improve the applicability of the model.

[0201] The method for calculating the output parameters of the acoustic feature optimization function includes:

[0202] The formula for calculating the optimal bandpass filter frequency range is:

[0203] ;

[0204] ;

[0205] In the formula, and are the lower and upper frequency limits of the bandpass filter respectively; is the probe center frequency; is the nominal bandwidth of the probe; is the acoustic impedance value of the weld material; The reference acoustic impedance value is usually taken as the acoustic impedance of steel 45.7×10 6 kg / (m²·s); and is the adjustment factor, usually The value ranges from 0.1 to 0.3. The value ranges from 0.2 to 0.4.

[0206] This formula takes into account the influence of the material acoustic impedance on the optimal filtering frequency band, and can automatically adjust the filtering parameters according to the material characteristics, thereby improving the adaptability of signal processing.

[0207] Specifically, the principle of the present invention is: the technical principle of the present invention is based on the deep integration of the physical mechanism of ultrasonic detection and artificial intelligence technology, and realizes efficient identification of weld defects through multi-dimensional optimization. First, a three-dimensional digital model of the weld is established and meshed, which provides an accurate geometric basis for ultrasonic propagation path planning. Based on the theory of sound wave propagation, the present invention calculates the optimal scanning path of the ultrasonic probe to ensure that the ultrasonic beam can penetrate the entire area of ​​the weld at the optimal angle, fundamentally improving the detection coverage.

[0208] Phased array ultrasonic technology is the core physical basis of the present invention. By controlling the transmission and reception timing delays of each array element, dynamic focusing and deflection of the ultrasonic beam are achieved. The present invention sets the focusing depth at two-thirds of the weld thickness. This setting is based on the attenuation law of ultrasonic waves in the weld, and can achieve the best weld coverage effect while ensuring penetration. The incident angle is set to the first angle value (within the range of 45 to 70 degrees), which is the angle that can produce the strongest shear wave conversion efficiency calculated based on the acoustic wave mode conversion theory, effectively improving the detection capability of vertical defects.

[0209] In the field of signal processing, the present invention uses a combination of adaptive beamforming algorithm and acoustic feature optimization function to achieve a significant improvement in signal quality. Adaptive beamforming enhances the spatial resolution of the echo signal inside the weld by adjusting the array element parameters in real time; the acoustic feature optimization function intelligently calculates key parameters such as the optimal bandpass filter frequency range and signal gain curve according to the weld material characteristics and the detection environment, effectively suppresses structural noise and scattering interference, and improves the defect signal-to-noise ratio by at least 5 decibels, so that tiny defects can be reliably detected even in a high-noise environment.

[0210] The UltraDefectNet model is the innovation of the present invention in the field of artificial intelligence. Its hybrid architecture of multi-layer convolutional neural network and bidirectional long short-term memory network is designed according to the characteristics of ultrasonic signals. Five convolutional layers extract spatial features, three-layer bidirectional long short-term memory network captures temporal features, and the self-attention mechanism realizes the weight distribution of different frequency components, which perfectly matches the time-frequency characteristics of ultrasonic defect signals. In particular, the design that the number of self-attention heads is related to the acoustic properties of weld materials enables the model to adapt to the acoustic properties of different materials and enhances the generalization ability of the algorithm.

[0211] Multi-angle data fusion technology is the key to solving complex weld detection problems in this invention. By integrating ultrasonic data acquired from different angles through spatial mapping and probability statistics methods, not only the detection coverage is improved, but also the defect characteristics can be described from multiple dimensions, significantly enhancing the reliability of defect identification. Combined with the precise location information of defects calculated by the time difference method, this invention successfully constructs a high-precision three-dimensional distribution map of weld defects, providing comprehensive and objective data support for weld quality assessment.

[0212] A specific embodiment 1 of the present invention is provided below, and the specific implementation method of each step in this embodiment 1 is described in detail as follows.

[0213] The specific implementation method of step S01 is to first collect the weld shape parameters, including the weld length, width, thickness, groove angle and other geometric dimension information, and then use 3D modeling software to build an accurate digital model of the weld. Then, the finite element meshing technology is used to refine the weld model. The mesh size of the weld area is usually set to 1 / 8 to 1 / 10 of the ultrasonic wavelength to ensure the calculation accuracy. Then, the propagation path of the ultrasonic wave in the weld material is simulated based on the ray tracing algorithm. The algorithm uses the reflection and refraction laws of sound waves at the material interface and calculates the propagation direction of the sound wave according to Snell's law, which is specifically expressed as:

[0214] ;

[0215] In the formula, is the angle of incidence; is the refraction angle; is the speed of sound in the incident medium; is the speed of sound in the refracting medium.

[0216] At the same time, the acoustic impedance distribution of the weld material is considered, and the coverage area of ​​the ultrasonic beam at each incident point and incident angle is calculated. Finally, the A* path planning algorithm is used to determine the optimal scanning path of the ultrasonic probe. The evaluation function of the algorithm is expressed as:

[0217] ;

[0218] In the formula, For Node The total evaluation function value of ; From the starting point to the node the actual cost of For slave nodes The estimated cost to reach the end point is calculated as:

[0219] ;

[0220] In the formula, For Node Euclidean distance to the target point; For Node The weld coverage at the position ranges from 0 to 1; and is the weight coefficient, usually The value range is 0.3~0.7. The value range is 0.3 to 0.7, and The purpose of this step is to provide accurate probe motion control instructions for the subsequent flaw detection process through digital modeling technology, thereby improving detection efficiency and defect detection rate.

[0221] The specific implementation method of step S02 is to first set the number of array elements of the phased array ultrasonic probe according to the weld thickness parameters. Usually, a 32-element probe is used for welds with a thickness of less than 10 mm, a 64-element probe is used for welds with a thickness of 10 mm to 30 mm, and a 128-element probe is used for welds with a thickness of more than 30 mm. Then, the position of two-thirds of the weld thickness is calculated as the ultrasonic focusing depth. This depth setting can obtain better near-surface defect detection capabilities while ensuring the sensitivity of weld root detection. Then, the first angle value is calculated based on the ratio of the longitudinal wave velocity to the shear wave velocity of the weld material. The calculation formula is:

[0222] ;

[0223] In the formula, is the first angle value, i.e., the incident angle that produces the strongest shear wave conversion efficiency; is the shear wave speed; is the longitudinal wave speed.

[0224] For typical carbon steel materials, About 5900~6100m / s, About 3200~3300m / s, calculated Usually it is in the range of 32° to 34°. This angle needs to take into account the sound velocity measurement error and introduce a correction term:

[0225] ;

[0226] In the formula, is a correction term with a value range of -2° to 2°, which is determined through actual testing. Finally, the transmission delay time of each array element is adjusted through the phased array control system so that the ultrasonic beam is incident on the weld at the first angle value. The calculation formula for the delay time of the phased array ultrasonic probe element is:

[0227] ;

[0228] In the formula, For the The delay time of each array element; For the The distance from the array element to the focal point; is the distance from the reference array element to the focal point; is the speed of sound in the wedge. When considering the refraction at the weld interface, the above formula needs to be modified to:

[0229] ;

[0230] In the formula, For the The distance from the array element to the incident point on the weld surface; is the distance from the incident point to the focal point on the weld surface; is the distance from the reference array element to the incident point on the weld surface; is the distance from the incident point to the focal point on the weld surface; The purpose of this step is to ensure that the ultrasonic wave can enter the weld at the best angle and focus state to improve the sensitivity of defect detection.

[0231] The specific implementation method of step S03 is to first measure the amplitude of the weld background noise, and then set the time domain synchronous acquisition threshold to 2.5 to 3.5 times the average amplitude of the background noise. This setting can effectively filter invalid signals while not missing small-amplitude defect echoes. Then start the automatic gain compensation system, which automatically adjusts the gain parameters according to the attenuation law of sound waves in the material. The gain value calculation formula of the automatic gain compensation system is:

[0232] ;

[0233] In the formula, The depth is The gain value at , in dB; is the initial gain value, usually set to 20-30dB; It is the first-order coefficient, which indicates the attenuation coefficient of sound waves in the material, and its unit is dB / mm; It is the quadratic coefficient, which is used to compensate for the scattering loss caused by the increase in depth. The unit is dB / mm², and the value is usually 0.001~0.01dB / mm²; is the sound wave propagation depth, in mm.

[0234] For different materials, The typical values ​​are: carbon steel 0.5 ~ 2dB / mm, stainless steel 0.3 ~ 1.5dB / mm, aluminum alloy 0.2 ~ 1dB / mm. Then, for different thickness areas of the weld, a regional dynamic gain control strategy is adopted, that is, a low gain setting is used in the thin weld area, and a high gain setting is used in the thick weld area. The gain difference is proportional to the thickness change. The relationship between the gain difference and thickness change of the regional dynamic gain control is expressed as:

[0235] ;

[0236] In the formula, is the gain difference, in dB; is the proportionality coefficient, the unit is dB / mm, and the value range is 0.8~1.2dB / mm; is the thickness change value, in mm.

[0237] Finally, a mapping table between weld thickness and gain parameters is established for the system to query and call in real time during the scanning process. The purpose of this step is to improve the adaptability of the flaw detection system to weld areas of different thicknesses by optimizing signal acquisition parameters and ensure the consistency of detection sensitivity.

[0238] The specific implementation method of step S04 is to first apply the delay-sum beamforming basic algorithm to obtain the initial echo signal, and then use the minimum variance distortion-free response beamforming algorithm to adaptively optimize the signal. The algorithm can automatically suppress interference signals from non-target directions and improve the signal-to-noise ratio of the target defect signal. The weight vector calculation formula of the minimum variance distortion-free response beamforming algorithm is:

[0239] ;

[0240] In the formula, is the weight vector with dimension , is the number of array elements; is the covariance matrix of the received signal, with dimension ; is the guidance vector in the target direction, with dimension ; for The conjugate transpose of .

[0241] Covariance matrix The calculation method is:

[0242] ;

[0243] In the formula, is the signal vector received by each array element, with dimension ; Represents the expected operation.

[0244] Then, the optimized echo signal is subjected to envelope detection to extract characteristic parameters such as signal amplitude, phase, and frequency. Then, wavelet transform is applied to perform time-frequency analysis on the signal, and the dB4 wavelet basis function is used to perform 5-layer decomposition to extract the energy distribution characteristics of each frequency band. The calculation formula of wavelet transform is:

[0245] ;

[0246] In the formula, is the wavelet transform coefficient; is the scale parameter, which controls the expansion and contraction of the wavelet; is the translation parameter, which controls the position of the wavelet; is the signal to be analyzed; is the wavelet basis function; express The complex conjugate of .

[0247] Then, the extracted features are reduced in dimension by combining the principal component analysis method, and the principal component with a contribution rate of 95% is retained as the defect feature vector. The formula for calculating the eigenvalue and eigenvector of the principal component analysis method is:

[0248] ;

[0249] In the formula, is the covariance matrix of the feature data; For the eigenvalues; is the corresponding feature vector.

[0250] Finally, the defect feature vector is quantized and encoded to form a standardized defect feature descriptor for subsequent defect identification and classification. The purpose of this step is to extract valuable defect feature information from complex echo signals through advanced signal processing algorithms to provide a reliable data basis for defect identification.

[0251] The specific implementation method of step S05 is to first analyze the spectrum distribution of the defect characteristic signal, and extract the peak frequency, valley frequency, bandwidth, energy distribution and other characteristic parameters in the spectrum. Then construct a multi-layer perceptron neural network model, the number of nodes in the input layer is consistent with the characteristic dimension, the hidden layer adopts a 3-5 layer structure, the number of nodes in each layer is 1.5-2 times that of the input layer, and the number of nodes in the output layer is consistent with the number of defect types. Then import the known defect sample library data, which contains characteristic signal data of typical welding defects such as pores, cracks, slag inclusions, and unfusion. Each type of defect contains at least 1000 groups of samples. The back propagation algorithm is then used to train the neural network, and the calculation formula for weight update is:

[0252] ;

[0253] In the formula, For the Layer Neuron to Layer The connection weights of neurons; is the learning rate; is the loss function Weight The partial derivative of .

[0254] The loss function of the neural network model is:

[0255] ;

[0256] In the formula, is the cross entropy loss function value; is the sample size; is the number of defect categories; For sample Belongs to category The true label (0 or 1); Predict samples for the model Belongs to category probability.

[0257] The initial value of the learning rate is set to 0.01, and the cosine annealing strategy is used to gradually reduce the learning rate:

[0258] ;

[0259] In the formula, For the The learning rate of the iteration; is the minimum learning rate, usually set to 0.01 times the initial learning rate; is the maximum learning rate, that is, the initial learning rate; is the current iteration number; is the total number of iterations.

[0260] Finally, the model performance is evaluated by cross-validation method, and the model training is completed when the recognition accuracy reaches more than 90%. The purpose of this step is to establish a defect recognition model based on spectral features and use machine learning methods to improve the accuracy and stability of defect classification.

[0261] The specific implementation of step S06 is the same as above and will not be described in detail here.

[0262] The specific implementation method of step S07 is to first perform time delay correction on the collected multi-angle reflected echo signals to eliminate the time deviation caused by the difference in sound path at different angles. Then the time difference method is used to calculate the defect position, and the calculation formula is:

[0263] ;

[0264] In the formula, is the defect depth, in mm; is the propagation speed of sound waves in the material, in mm / μs; is the round trip time of the sound wave, in μs.

[0265] Then the lateral position of the defect is determined by the triangulation principle. The lateral position calculation formula is:

[0266] ;

[0267] In the formula, is the lateral distance of the defect relative to the incident point, in mm; is the defect depth, in mm; is the incident angle of the ultrasonic beam.

[0268] When the refraction effect is taken into account, the formula is modified to:

[0269] ;

[0270] In the formula, is the refraction angle, calculated according to Snell's law: , and are the speed of sound in the incident medium and the refracting medium, respectively, is the angle of incidence.

[0271] Then, the multi-angle data is fused by Bayesian, that is, weights are assigned according to the reliability of the data at each angle to build a probability distribution model of the defect location. The Bayesian weight calculation formula for multi-angle data fusion is:

[0272] ;

[0273] In the formula, For the The weight of the angle data; For the The measurement uncertainty of the angle data; is the number of angles.

[0274] The calculation formula of the defect position coordinates after fusion is:

[0275] ;

[0276] In the formula, is the defect position coordinate after fusion; For the The defect position coordinates are measured at 3 angles.

[0277] Finally, a three-dimensional distribution map of weld defects is constructed based on the fused data, and three-dimensional voxel rendering technology is used to intuitively display the spatial distribution of defects in the weld. The purpose of this step is to improve the defect location accuracy through multi-angle data fusion and generate intuitive defect distribution visualization results, providing a reliable basis for weld quality assessment.

[0278] The specific implementation of step S08 is to first determine the input parameters of the acoustic feature optimization function, including the acoustic impedance value of the weld material, the weld thickness parameter, the background noise spectrum characteristics, the expected defect type index and the detection sensitivity requirements. Then, a genetic algorithm is used to solve the acoustic feature optimization function, and the expression of the function is:

[0279] ;

[0280] In the formula, To optimize the objective function value; is a signal processing parameter vector, including bandpass filter frequency range, signal gain curve parameters, time domain window length and noise suppression threshold; For parameters The corresponding signal-to-noise ratio; For parameters The corresponding resolution; For parameters The corresponding computational complexity; , and are weight coefficients, which respectively represent the importance of signal-to-noise ratio, resolution and computational complexity. The value ranges from 0.5 to 0.7. The value ranges from 0.2 to 0.4. The value is 0.1~0.2, and .

[0281] The population size is set to 50, the number of iterations is set to 100, the crossover probability is set to 0.8, and the mutation probability is set to 0.1. Then, the bandpass filter parameters are set according to the optimization results. The optimal bandpass filter frequency range is calculated as follows:

[0282] ;

[0283] ;

[0284] In the formula, and are the lower and upper frequency limits of the bandpass filter respectively; is the probe center frequency; is the nominal bandwidth of the probe; is the acoustic impedance value of the weld material; The reference acoustic impedance value is usually taken as the acoustic impedance of steel 45.7×10 6 kg / (m²·s); and is the adjustment factor, usually The value ranges from 0.1 to 0.3. The value ranges from 0.2 to 0.4.

[0285] Usually, for a 2.5MHz center frequency probe, the optimal passband range is 1.5MHz to 3.5MHz; for a 5MHz center frequency probe, the optimal passband range is 3MHz to 7MHz. Then, an adaptive Wiener filter algorithm is applied to further suppress structure noise. The algorithm can automatically adjust the filter coefficient according to the power spectral density ratio of the signal to the noise. The frequency domain expression is:

[0286] ;

[0287] In the formula, is the filter frequency response function; is the power spectrum density of the defect signal; is the power spectral density of the noise.

[0288] Finally, the optimization effect is verified by comparing the signal-to-noise ratio before and after signal processing. The calculation formula for the signal-to-noise ratio improvement is:

[0289] ;

[0290] In the formula, is the signal-to-noise ratio improvement value, in dB; is the signal-to-noise ratio after filtering; is the signal-to-noise ratio before filtering.

[0291] The signal-to-noise ratio after optimization should be improved by no less than 5dB. The purpose of this step is to effectively eliminate structural noise and scattering interference by optimizing signal processing parameters, and to improve the clarity and recognizability of defect signals.

[0292] The specific implementation method of step S09 is to first collect defect information of all detection positions, including defect type, location, size and other parameters. Then load the pre-trained UltraDefectNet model, which can more accurately identify and classify weld defects based on deep learning algorithms. The calculation formula for the number of heads of the self-attention mechanism of the UltraDefectNet model is:

[0293] ;

[0294] In the formula, is the number of heads of the self-attention mechanism; is the longitudinal wave velocity of the weld material; is the shear wave velocity of the weld material.

[0295] For a typical carbon steel weld, the number of heads is usually 8 to 12. The scaling factor calculation formula in the self-attention mechanism is:

[0296] ;

[0297] In the formula, is the scaling factor in the self-attention mechanism; is the first angle value.

[0298] Then, the defect parameters are compared with standard specifications such as GB / T5777 to evaluate the severity and allowability of each defect. The overall quality grade of the weld is then calculated based on the evaluation results, which is usually divided into four levels: level 1 (excellent), level 2 (good), level 3 (qualified) and unqualified. Finally, a weld quality assessment report is automatically generated, including defect distribution diagrams, defect type statistics, serious defect warnings, and overall quality evaluation conclusions. The purpose of this step is to objectively evaluate the weld quality based on the test results and standard specifications, and generate a standardized quality assessment report to provide a basis for welding process improvement and quality control.

[0299] In order to better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A research team used the high-efficiency flaw detection method of the present invention to improve the technology in view of the problems existing in the automatic detection process of the heat exchange tube weld. The steam generator heat exchange tube is made of austenitic stainless steel, with a wall thickness of 3.5mm and an outer diameter of 19mm. The weld is annular and has a width of about 4.2mm.

[0300] First, the research team established a three-dimensional digital model of the weld and performed meshing. The actual size parameters of the weld were collected, and an accurate model was established. The mesh size of the weld area was set to 1 / 9 of the ultrasonic wavelength, which is about 0.12 mm. After acoustic simulation, it was determined that the longitudinal wave speed in the weld material is 5950 m / s and the shear wave speed is 3270 m / s. According to Snell's law and the A* path planning algorithm, the optimal scanning path and weight coefficient were calculated. Set to 0.6, Set to 0.4.

[0301] Secondly, the team selected a 64-element phased array ultrasonic probe and set the focus depth at two-thirds of the weld thickness, or 2.33 mm. Calculate the first angle value and substitute the parameters to get Taking into account the error correction in the actual test, the final The delay time of each element of the probe is calculated according to the formula Calculated, where the probe wedge sound velocity Table 1 lists the delay time settings of some array elements.

[0302] Table 1 Phased array probe element delay time setting table

[0303]

[0304] Next, the team set the time domain synchronous acquisition threshold to 3.2 times the average amplitude of the background noise, which is about 18.4 dB. The automatic gain compensation system was started, and the parameters were set according to the material attenuation characteristics. The initial gain value The attenuation coefficient is 24dB. 1.2dB / mm, scattering loss coefficient For the area where the weld thickness changes, a regional dynamic gain control strategy is applied, and the proportional coefficient Set to 1.05dB / mm.

[0305] In the signal processing phase, the team first used the minimum variance distortion-free response beamforming algorithm to process the echo signal. The weight vector of the algorithm is expressed by the formula The signal was decomposed into 5 layers using the dB4 wavelet basis function, and the time-frequency features were extracted by wavelet transform. The principal component analysis was used for dimensionality reduction, and the features with a contribution rate of 95.8% were retained. The extracted defect feature parameters are shown in Table 2.

[0306] Table 2 Statistics of defect characteristic parameters

[0307]

[0308] Subsequently, the team built a multi-layer perceptron neural network model with 25 nodes in the input layer, 4 hidden layers with 45, 40, 35 and 30 nodes in each layer, and 4 output layers. The network was trained using the back propagation algorithm, with the initial learning rate set to 0.01 and the cosine annealing strategy used to adjust the learning rate. After 50 cycles of training, the model's recognition accuracy on the validation set reached 93.2%.

[0309] During the detection implementation phase, the team adopted a multi-angle scanning strategy. In addition to the first angle value of 34.7°, six auxiliary angles of 29.7°, 39.7°, 24.7°, 44.7°, 19.7° and 49.7° were used for scanning. The longitudinal segment length of the weld is 75% of the total length of the probe array element, about 36mm, and the overlapping area of ​​adjacent segments is 12mm. The root of the weld is supplemented by cross scanning, and the spacing is 50% of the length of the probe array element, about 24mm. Figure 2The comparison of the detection rates of various weld defects at different incident angles is shown. The horizontal axis is the incident angle (°), ranging from 19.7° to 49.7°; the vertical axis is the defect detection rate (%). The detection rate curves of four defect types (pores, cracks, slag inclusions and lack of fusion) are drawn using different line types and markers in the figure. The red dotted line marks the position of the first angle value θ1=34.7°, and the highest detection point is marked in red on each curve. It can be observed from the figure that the detection rate of pore defects is as high as 93% at 34.7°; the detection rate of crack defects is 94% at 44.7°; the detection rate of slag inclusion defects is 91% at 29.7°; and the detection rate of lack of fusion defects is as high as 95% at 39.7°. This shows that there are obvious differences in the detection effects of different types of defects at different incident angles, which verifies the necessity of multi-angle scanning strategy.

[0310] For defect location, the team used the time difference method to calculate the defect depth: , and combine the principle of triangulation to determine the lateral position: . Through multi-angle data fusion technology, Bayesian weight calculation formula is applied The measurement results at different angles are weighted to finally determine the exact location of the defect. Table 3 lists some of the detected defect information.

[0311] Table 3 Weld defect detection results

[0312]

[0313] To optimize the signal processing parameters, the team applied the acoustic feature optimization function , where the weight coefficient , and They are set to 0.65, 0.25 and 0.1 respectively. The optimal bandpass filter frequency range is obtained through iterative calculation of genetic algorithm: the lower limit frequency The upper limit frequency is 2.9MHz The frequency is 6.7MHz. Adaptive Wiener filtering is used to further suppress noise, and the signal-to-noise ratio is improved after filtering. Reach 7.3dB. Figure 3 The spectrum comparison of the ultrasonic signal before and after processing is shown. The horizontal axis is frequency (MHz) and the vertical axis is normalized amplitude. The blue curve represents the original signal spectrum, the red curve represents the processed signal spectrum, and the green dotted line represents the bandpass filter response curve. The lower cutoff frequency of the bandpass filter is marked in the figure. and upper cutoff frequency , and the peak frequency of the processed signal is marked on the red curve as 5.0MHz. The processed signal spectrum is smoother, the main energy is concentrated in the bandpass range, and the background noise is significantly reduced, which effectively verifies the effectiveness of the acoustic feature optimization function in determining the optimal filtering parameters.

[0314] Finally, the research team applied the UltraDefectNet model to analyze the detection results. The number of self-attention heads in this model is calculated according to the formula The calculation results are rounded to 8 heads. The z-score method is used for standardization: , the processed data have good comparability.

[0315] Traditional weld ultrasonic flaw detection methods mainly rely on manual experience for data interpretation. The detection sensitivity is greatly affected by the skill level of the operator, and it is difficult to achieve accurate quantitative analysis of complex weld defects. The detection efficiency is low. The detection time of a single circular weld usually takes 20 to 30 minutes. It is difficult to meet production needs for batch detection of multiple welds. The analysis process of the test results is highly subjective and lacks quantitative evaluation standards, resulting in poor consistency of the evaluation results.

[0316] The efficient flaw detection method of the present invention shortens the detection time of a single annular weld to 5 to 8 minutes and improves efficiency by more than 300% by establishing a three-dimensional digital model of the weld and planning the optimal scanning path. The adaptive beamforming algorithm and multi-angle data fusion technology are used to improve the defect positioning accuracy to ±0.2mm, which is significantly improved compared with the ±0.8mm accuracy of the traditional method. The UltraDefectNet deep learning model is used for defect identification, and the classification accuracy rate reaches 93.2%, which is much higher than the 75-80% accuracy rate of the traditional method. The optimal filtering parameters are automatically determined by the acoustic feature optimization function, which improves the signal-to-noise ratio by 7.3dB and significantly enhances the detection ability of tiny defects. This method realizes the automated analysis and objective evaluation of weld inspection results, eliminates the interference of human factors, improves the consistency of evaluation results by 85%, and provides a more reliable technical guarantee for the safe operation of key welding components.

[0317] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 4 and 5 below.

[0318] Table 4 Variable explanation table (Part I)

[0319]

[0320] Table 5 Variable explanation table (Part II)

[0321]

[0322] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An efficient flaw detection method for an ultrasonic flaw detector for automated weld detection, characterized in that: include: Establish a three-dimensional digital model of the weld and determine the optimal scanning path of the ultrasonic probe; configure the phased array ultrasonic probe parameters, set the focus depth to two-thirds of the weld thickness, and adjust the probe incident angle to the first angle value; set the time domain synchronous acquisition threshold and start the automatic gain compensation system; apply the adaptive beamforming algorithm to process the reflected echo signal and extract the defect characteristic signal; according to A neural network model is established based on defect characteristics and deep learning training is performed. A multi-angle scanning strategy is used to implement full coverage detection. The defect position is calculated using the time difference method, and a three-dimensional defect distribution map is constructed using multi-angle data fusion technology. The acoustic feature optimization function is used to determine the optimal filtering parameters and improve the defect signal-to-noise ratio. A quality assessment report is generated based on the detection results, and the UltraDefectNet model is used for high-precision defect classification.

2. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 1 is characterized in that: The optimal scanning path specifically refers to the probe movement trajectory calculated based on the weld geometry and the acoustic properties of the weld material, which is used to make the ultrasonic beam penetrate the entire weld area at the optimal angle.

3. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 2, characterized in that: The adaptive beamforming algorithm specifically refers to a signal processing method that achieves dynamic focusing and deflection of the ultrasonic beam by adjusting the transmission and reception timing delays of each array element of the phased array ultrasonic probe, thereby obtaining higher resolution and stronger penetration ability.

4. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 3 is characterized in that: The first angle value specifically refers to the incident angle calculated based on the ratio of the longitudinal wave sound velocity of the weld material to the shear wave sound velocity of the weld material for generating the strongest shear wave conversion efficiency.

5. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 4, characterized in that: The acoustic feature optimization function is used to determine the optimal signal processing parameters according to the weld material characteristics and the detection environment. The input includes the weld material acoustic impedance value, weld thickness parameters, background noise spectrum characteristics, expected defect type indicators and detection sensitivity requirements. The output is the optimal bandpass filter frequency range, the optimal signal gain curve, the optimal time domain window length and the optimal noise suppression threshold.

6. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 5, characterized in that: The optimal bandpass filter frequency range is used to set the upper and lower limits of the passband frequency of the signal filter; the optimal signal gain curve is used to set the gain adjustment parameters of the automatic gain compensation system; the optimal time domain window length is used to set the sampling duration of the time domain synchronous acquisition threshold; the optimal noise suppression threshold is used to set the working threshold of the noise suppression system.

7. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 6, characterized in that: The specific structure of the UltraDefectNet model is a hybrid architecture that combines a multi-layer convolutional neural network with a bidirectional long short-term memory network. It includes five convolutional layers for extracting spatial features from ultrasonic reflection echo signals, and three bidirectional long short-term memory network layers for capturing temporal features. It also uses a self-attention mechanism to achieve weight allocation for different frequency components.

8. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 7, characterized in that: The number of self-attention heads is related to the acoustic properties of the weld material. The calculation formula for the number of heads is four times the ratio of the longitudinal wave velocity of the weld material to the transverse wave velocity of the weld material. The scaling factor in the attention mechanism is proportional to the first angle value. Finally, the defect type and defect size estimation results are output through the fully connected layer.

9. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 8, characterized in that: The steps for establishing the training data set during the UltraDefectNet model training process include collecting typical defect ultrasonic reflection echo signals generated under various welding processes, labeling each defect type and recording the defect size parameters and defect location information, and using data enhancement technology to simulate signal performance in different noise environments.

10. The high-efficiency flaw detection method of ultrasonic flaw detector for automatic weld detection according to claim 9, characterized in that: The training data set contains multiple sets of labeled data including porosity defects, crack defects, slag inclusion defects, and lack of fusion defects. At the same time, a validation set is established for UltraDefectNet model evaluation. The validation set data comes from actual weld inspection cases and has no overlap with the training set, ensuring the generalization ability of the UltraDefectNet model.

Images