A method and system for detecting appearance defects of a tungsten electrode of an argon arc welding, and a storage medium

Abstract

The application provides a kind of argon arc welding tungsten electrode appearance defect detection method, system and storage medium, it is related to tungsten electrode appearance detection technical field, the sample electrode of known defect is scanned three-dimensionally first, is discretized into several space points and is endowed with each point defect existence label. By applying composite sine wave excitation current, the eddy current signal when the electrode uniform speed spiral passes through the detection coil is collected, the full matrix data is obtained by laser ultrasonic scanning simultaneously and the three-dimensional gray scale model is reconstructed. A double-branch deep learning model is constructed, the eddy current feature vector and ultrasonic image block features are fused, and the mapping relationship between them and the defect label is learned. The same detection is performed on the electrode to be tested, and the trained model outputs the defect confidence of each position. Combined with the preset defect severity weight and spatial key area weight, the comprehensive quality index is calculated, and the electrode is determined to be qualified or not according to the qualified threshold. The application realizes high-precision, automatic comprehensive detection of surface and internal defects.

Description

Technical Field

[0001] This invention relates to the field of tungsten electrode appearance inspection technology, specifically to a method, system, and storage medium for detecting appearance defects in argon arc welded tungsten electrodes. Background Technology

[0002] In argon arc welding, the tungsten electrode serves as the carrier of the electric arc, and the microscopic integrity of its surface and the uniformity of its internal material directly determine the stability of the arc and the welding quality. During the manufacturing process, fluctuations in powder metallurgy, sintering, or subsequent grinding processes can easily lead to defects in the electrode, such as longitudinal / transverse microcracks, internal porosity, and non-metallic inclusions. These defects, especially microscopic defects at the micrometer or sub-millimeter level, are so small and inconspicuous that they completely exceed the resolution limits of human vision, rendering traditional manual inspection methods essentially ineffective. Even under strong magnification, internal defects remain undetectable, posing a primary challenge to quality control.

[0003] Traditional inspection methods primarily rely on manual visual inspection or single non-destructive testing techniques. However, manual inspection is inefficient and easily affected by subjective factors, while single techniques, due to their limitations, cannot simultaneously cover the accurate identification of both minute surface defects and internal three-dimensional defects. With the upgrading of industrial quality control, higher demands are placed on the comprehensiveness, automation, and reliability of defect detection. Although various physical inspection methods exist, no single technique can achieve stable, reliable, and high-precision identification and classification in real industrial inspection environments when faced with various microscopic defects invisible to the naked eye on tungsten electrodes. Developing a new method that can break through the detection limits of single techniques and deeply integrate multi-source information to significantly improve the confidence level of micro-defect identification has become an inevitable technological development direction for solving the comprehensive quality control of tungsten electrodes.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and storage medium for detecting appearance defects in argon arc welding tungsten electrodes, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for detecting visual defects in argon arc welding tungsten electrodes, comprising the following steps:

[0008] Step 1: Discretize the sample tungsten electrode with known defect distribution into several points with spatial relative information, and assign a corresponding defect presence label to each point based on the defect category, which includes surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions.

[0009] Step 2: Apply a composite sinusoidal excitation current to the detection coil. As the sample tungsten electrode moves through the central hole of the detection coil in a relative spiral motion along the axial direction of the detection coil, acquire multiple sets of eddy current detection signals. The eddy current detection signals include at least the sub-segment of the sample tungsten electrode in the detection coil, the rotation angle, and the real and imaginary parts of the induced voltage at both ends of the coil.

[0010] Step 3: Perform ultrasonic testing on the sample tungsten electrode to obtain its three-dimensional grayscale model. Based on the sub-segment, extract the corresponding region image block from the three-dimensional grayscale model and retrieve the defect presence marker of the corresponding point based on the sub-segment.

[0011] Step 4: Construct a deep learning classification model to learn the mapping relationship between rotation angle, real and imaginary components, and defect existence markers between region image blocks and corresponding points in each sub-segment;

[0012] Step 5: Based on the deep learning classification model, obtain the confidence level of each point of the tungsten electrode to be tested belonging to various types of defects, and judge them according to the preset defect tolerance standard to determine whether the tungsten electrode to be tested is qualified.

[0013] Furthermore, assigning the existence marker of the defect specifically includes:

[0014] The discretization into several points with spatial relative information specifically involves: obtaining the three-dimensional coordinates of each point on the sample tungsten electrode to form a set of coordinate positions; the defect existence marker is a four-digit code, and the four codes from left to right correspond one-to-one with surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions, respectively. When a point belongs to a certain type of defect, the code of that position in the defect existence marker corresponding to that point is set to 1, otherwise it is set to 0.

[0015] Furthermore, the acquisition process of the eddy current detection signal specifically includes:

[0016] The composite sinusoidal excitation current is formed by the linear superposition of high-frequency and low-frequency components, wherein the frequency range of the high-frequency component is [1.2MHz, 2MHz], the frequency range of the low-frequency component is [200kHz, 400kHz], and the current amplitude of the high-frequency component is smaller than that of the low-frequency component.

[0017] The real and imaginary components of the induced voltage at both ends of the coil are synchronously acquired by a phase-sensitive detector. The phase-sensitive detector is based on the lock-in amplification principle and uses orthogonal reference signal pairs corresponding to each frequency component in the composite sinusoidal excitation current to extract the real and imaginary components of each frequency component from the induced voltage through orthogonal adjustment.

[0018] The relative spiral motion is a uniform spiral motion, and the rotation angle and moving speed of the sample tungsten electrode satisfy the following relationship: ;in, This indicates the moving speed, specifically the speed at which the sample tungsten electrode moves along its axis. The value represents the rotational speed, and D represents the diameter of the sample tungsten electrode.

[0019] Furthermore, the ultrasonic detection specifically includes:

[0020] An ultrasonic testing device is used to scan the sample tungsten electrode to obtain full matrix data of its internal structure. The full matrix data is preprocessed and processed using synthetic aperture focusing technology to reconstruct a three-dimensional grayscale model of the sample tungsten electrode. The three-dimensional grayscale model consists of the three-dimensional coordinates and grayscale values ​​of each point, wherein the grayscale values ​​of each point are obtained by coherent superposition of the preprocessed ultrasonic signals.

[0021] Furthermore, reconstructing the three-dimensional grayscale model specifically includes:

[0022] The three-dimensional space to be reconstructed is discretized into a regular grid. For each grid point, the propagation time delay of each acquired signal at that point is calculated based on the ultrasonic propagation principle. The signal amplitude at the corresponding time delay is extracted from the preprocessed ultrasonic signal and weighted coherently superimposed to obtain the gray value of that grid point. The gray values ​​of all grid points together constitute a three-dimensional gray model.

[0023] Furthermore, during the uniform spiral motion of the sample tungsten electrode, the eddy current detection signal is sampled at equal time intervals, and the timestamp of each sampling is recorded. The real and imaginary components of the induced voltage collected at each timestamp are combined to form a complex impedance signal. On the complex plane formed by the real component as the horizontal axis and the imaginary component as the vertical axis, each complex impedance signal corresponds to a coordinate point. As the sample tungsten electrode performs eddy current scanning, the continuously collected complex impedance signals form a trajectory on the complex plane. Based on all the coordinate points on the trajectory, a polynomial fitting method is used to obtain the functional expression of the trajectory, which is used as the ideal trajectory function.

[0024] Furthermore, the deep learning classification model is constructed as follows:

[0025] For the acquired eddy current detection signal, a feature vector is constructed based on the rotation angle, real component value, and imaginary component value at each time stamp; the deviation between the coordinate point of the complex impedance signal and the ideal trajectory function at each time stamp is calculated, and the deviation is obtained by calculating the vertical distance from the actual coordinate point to the ideal trajectory; the rotation angle, real component value, imaginary component value, and deviation are combined to form a numerical eddy current feature vector.

[0026] From the 3D grayscale model, extract the regional image patch that corresponds spatially to the sub-segment to serve as the regional image patch feature;

[0027] The eddy current feature vector, the regional image patch feature, and the defect presence label of the corresponding point are associated to form a training sample; the data of all spatial location points are sorted to form the final training sample set, and divided into training set, validation set and test set in a ratio of 6:2:2; the eddy current feature vector and regional image patch feature in the training set are used as input, and the defect presence label of the corresponding point is used as label to train the deep learning classification model.

[0028] The deep learning classification model employs a dual-branch convolutional neural network, comprising: a one-dimensional convolutional branch for processing eddy feature vectors and a three-dimensional convolutional branch for processing region image patch features; a feature fusion module for concatenating the feature vectors output from the fully connected layers at the ends of the one-dimensional and three-dimensional convolutional branches; and a classifier consisting of at least one fully connected layer, whose input is the feature vector concatenated by the feature fusion module. The final output layer of the classifier uses a Softmax activation function to output a defect presence marker for each point, as well as a confidence distribution vector indicating which point belongs to each defect category.

[0029] Further, determining whether the tungsten electrode to be tested is qualified specifically includes:

[0030] Eddy current testing and ultrasonic testing are performed on the tungsten electrode to be tested to obtain its eddy current testing signal and ultrasonic three-dimensional grayscale model; eddy current feature vectors and regional image block features corresponding to sub-segments of the tungsten electrode to be tested are extracted; the eddy current feature vectors and regional image block features are input into the trained deep learning classification model to obtain the confidence distribution vector output by the model; the maximum value in the confidence distribution vector is used as the reliability index. If the reliability index is greater than the preset confidence threshold, the point is marked as a suspected defect location.

[0031] The defect tolerance standard is based on a comprehensive quality index, and the specific calculation formula is as follows:

[0032]

[0033] in, The overall quality index of the tungsten electrode to be tested. Let be the severity weight of the q-th defect category. Let be the confidence level that the a-th suspected defect location belongs to the q-th defect category. Let be the spatial weight of the a-th suspected defect location; a is the index of the suspected defect location; A is the number of suspected defect locations; q is the defect category index; Q is the number of defect categories.

[0034]

[0035] in, Let a be the coordinates of the a-th suspected defect location. The coordinates are the center of the critical area of ​​the tungsten electrode to be tested. Parameters for controlling the width of the influence range of critical areas. This is a coefficient used to adjust the weight magnitude of key areas;

[0036] The comprehensive quality index of the tungsten electrode to be tested is compared with a preset pass threshold. If the comprehensive quality index of the tungsten electrode to be tested is not greater than the pass threshold, the tungsten electrode to be tested is determined to be qualified. If the comprehensive quality index of the tungsten electrode to be tested is greater than the pass threshold, the tungsten electrode to be tested is determined to be unqualified.

[0037] The present invention also provides a system for detecting appearance defects in argon arc welding tungsten electrodes. This system is used to implement the aforementioned method for detecting appearance defects in argon arc welding tungsten electrodes, and includes:

[0038] The defect category marking module is used to discretize the sample tungsten electrode with known defect distribution into several points with spatial relative information, and assign a corresponding defect presence mark to each point based on the defect category. The defect categories include surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions.

[0039] The eddy current scanning module is used to apply a composite sinusoidal excitation current to the detection coil. As the sample tungsten electrode passes through the central hole of the detection coil in a relative helical motion along the axial direction of the detection coil, multiple sets of eddy current detection signals are acquired. The eddy current detection signals include at least the sub-segment of the sample tungsten electrode in the detection coil, the rotation angle, and the real and imaginary parts of the induced voltage at both ends of the coil.

[0040] The ultrasonic testing module is used to perform ultrasonic testing on the sample tungsten electrode to obtain its three-dimensional grayscale model. Based on the sub-segment, the corresponding area image block is extracted from the three-dimensional grayscale model, and the defect presence marker of the corresponding point is retrieved based on the sub-segment.

[0041] The classification model building module is used to build a deep learning classification model to learn the mapping relationship between rotation angle, real and imaginary components, and defect existence markers between region image blocks and corresponding points in each sub-segment.

[0042] The detection and judgment module is used to obtain the confidence level of each point of the tungsten electrode to be detected belonging to various types of defects based on a deep learning classification model, and to judge it according to the preset defect tolerance standard to determine whether the tungsten electrode to be detected is qualified.

[0043] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for detecting appearance defects in argon arc welding tungsten electrodes.

[0044] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0045] This solution achieves integrated, high-precision detection of surface and internal defects in argon arc welded tungsten electrodes by fusing full-matrix data from eddy current scanning and laser ultrasound, and incorporating deep learning for collaborative analysis. Its advantages lie in effectively overcoming the limitations of traditional single-detection techniques: composite eddy current excitation and phase-sensitive detection enhance the sensitivity and discrimination of micro-cracks on the surface; non-contact laser ultrasound full-matrix acquisition and synthetic aperture imaging technology achieve high-resolution three-dimensional localization and identification of internal defects. Finally, through multimodal feature fusion and a deep learning model, the system can automatically and accurately classify and quantify various defects across the entire electrode range, and make objective judgments based on a comprehensive quality index. This significantly improves the reliability, automation level, and applicability of the detection, meeting the stringent quality requirements of tungsten electrodes in industrial settings. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0047] Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0049] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0050] Example:

[0051] Please see Figure 1 The present invention provides a technical solution:

[0052] A method for detecting visual defects in argon arc welding tungsten electrodes, comprising the following steps:

[0053] Step 1: Discretize the sample tungsten electrode with known defect distribution into several points with spatial relative information, and assign a corresponding defect presence label to each point based on the defect category, which includes surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions.

[0054] In this embodiment, assigning the defect existence marker specifically includes:

[0055] First, the sample tungsten electrodes with known defect distribution are discretized, i.e., high-precision three-dimensional coordinate data of the sample tungsten electrodes are obtained through three-dimensional scanning technology to form a set of coordinate positions. Specifically, a non-contact three-dimensional scanner (such as a laser scanner or structured light scanner) is used to scan the sample tungsten electrodes from all directions to obtain dense point cloud data of their surface. Each point cloud data point contains three-dimensional spatial coordinates (X, Y, Z), and these coordinates constitute a set of points with spatial relative information. Further, in order to optimize data processing efficiency and retain key spatial information, the original point cloud data is preprocessed, including the following sub-steps: point cloud denoising, point cloud simplification, and spatial index construction. Point cloud denoising specifically includes: using a statistical filtering algorithm to remove outlier noise points, specifically by calculating the average distance between each point and its b nearest neighbors, and removing points whose mean distance exceeds a preset standard deviation threshold based on a Gaussian distribution model. Here, the value of b is determined according to the point cloud density, and is usually set to 10 to 30; the standard deviation threshold is generally set to 1.0 to 2.0 to ensure that the true surface features are preserved while removing noise. Point cloud simplification specifically includes: uniformly downsampling the point cloud using a voxel grid filtering method while ensuring no loss of defect features. The voxel grid size needs to be set in a balance between processing speed and accuracy, typically ranging from 1 / 50 to 1 / 100 of the electrode diameter. For example, for a tungsten electrode with a diameter of 4 mm, the voxel size can be set to 0.04 mm to 0.08 mm. To accelerate subsequent spatial queries, a KD-tree (K-Dimensional Tree) spatial index structure is established on the processed point cloud to achieve efficient nearest neighbor search and region lookup.

[0056] Based on known defect categories, each discretized point is assigned a corresponding defect presence marker. This defect presence marker is a four-digit code, with each of the four digits from left to right corresponding to a longitudinal surface crack, a transverse surface crack, internal porosity, and internal inclusions, respectively. Specifically, the first digit indicates the presence of a longitudinal surface crack, the second indicates the presence of a transverse surface crack, the third indicates the presence of internal porosity, and the fourth indicates the presence of internal inclusions. When a point belongs to a defect of a certain category, the code at that position in the defect presence marker is 1; otherwise, it is 0. For example, marker 1001 indicates that the area where this point is located contains a longitudinal surface crack and internal inclusions, but not a transverse surface crack or internal porosity.

[0057] The method for determining the spatial neighborhood is as follows: A spherical neighborhood is constructed centered on the current point. The value of the neighborhood radius R is based on the typical size characteristics of the defect. By statistically analyzing a large number of known defect samples, the size distribution of various defects is obtained: Longitudinal surface cracks typically have a length of 0.5 mm to 5 mm and a width of 0.01 mm to 0.1 mm. Transverse surface cracks have a length comparable to longitudinal cracks, but their direction is perpendicular to the axial direction of the sample tungsten electrode. Internal pore diameters typically range from 0.05 mm to 0.5 mm, and internal inclusion sizes typically range from 0.02 mm to 0.3 mm. Based on these defect size characteristics, the neighborhood radius R should at least cover the maximum size of typical defects, while avoiding confusion between different defect regions due to an excessively large neighborhood. Specifically, the value of R ranges from 0.5 mm to 1.0 mm, with a preferred value of 0.7 mm. The R-value was optimized through cross-validation: different R-values ​​(0.3mm, 0.5mm, 0.7mm, 1.0mm, 1.2mm) were tested on the training set to assess their impact on defect classification accuracy, and the R-value that resulted in the highest accuracy was selected. For the detection of internal defects (porosity and inclusions), industrial CT (computed tomography) technology was used to obtain the internal three-dimensional structure data of the sample tungsten electrode. The CT data was registered and fused with the surface three-dimensional scan data, and the location and extent of internal defects were identified using a three-dimensional image segmentation algorithm, thereby determining the presence of internal defects at each discrete point. The assignment of defect presence markers was achieved using specialized annotation software that simultaneously displays three-dimensional point cloud data and corresponding CT slice data. Annotations were performed manually by experienced quality inspectors or assisted by semi-automatic algorithms. The annotation results were reviewed by multiple people to ensure accuracy. Through the above discretization and marking methods, a training sample set with spatial relative information and accurate defect markings was constructed, laying the data foundation for subsequent training of a deep learning-based defect detection model.

[0058] Step 2: Apply a composite sinusoidal excitation current to the detection coil. As the sample tungsten electrode moves through the central hole of the detection coil in a relative spiral motion along the axial direction of the detection coil, acquire multiple sets of eddy current detection signals. The eddy current detection signals include at least the sub-segment of the sample tungsten electrode in the detection coil, the rotation angle, and the real and imaginary parts of the induced voltage at both ends of the coil.

[0059] In this embodiment, the acquisition process of the eddy current detection signal specifically includes:

[0060] The composite sinusoidal excitation current is formed by the linear superposition of high-frequency and low-frequency components, and its mathematical expression is:

[0061]

[0062] in, This represents the composite sinusoidal excitation current at time t. For low-frequency components, For high-frequency components, The amplitude of the low-frequency component current. The amplitude of the high-frequency component current. The phase difference between the two components is, in this embodiment... t represents the time variable, indicating the time progression calculated from the moment the excitation current is applied, measured in seconds (s). This expression describes the instantaneous amplitude variation of the excitation current in the time domain. In practical systems, this signal is generated by a function generator or a direct digital frequency synthesis (DDS) module and applied to the detection coil as a continuous time-varying voltage or current signal.

[0063] The high-frequency component has a frequency range of [1.2MHz, 2MHz]. The eddy currents generated by the high-frequency excitation are concentrated on the surface layer of the material and are used to detect surface defects. According to the skin depth formula, the skin depth is calculated to be 0.12mm to 0.16mm in the frequency range of 1.2MHz to 2MHz. This depth range is just enough to effectively monitor microcracks on the surface of the tungsten electrode (the depth is usually 0.05mm to 0.15mm) without being excessively disturbed by the internal structure. The low-frequency component has a frequency range of [200kHz, 400kHz]. The eddy currents generated by the low-frequency excitation have a larger penetration depth and are used to detect internal defects. Under the same conditions, the skin depth corresponding to 200kHz to 400kHz is 0.31mm to 0.44mm. This depth can effectively cover the near-surface area of ​​the tungsten electrode and is used to detect internal defects such as pores and inclusions (usually located in the depth range of 0.2mm to 0.4mm below the surface). The setting of the current amplitude ratio is based on the following considerations: the high-frequency component current amplitude Less than the amplitude of the low-frequency component current The specific proportional relationship is as follows: =0.3 to 0.6, with an optimal value of 0.45. This ratio was determined through signal-to-noise ratio optimization experiments: while keeping the total excitation current constant, adjusting the high- and low-frequency amplitude ratio revealed that when... At a value of 0.45, the ratio of surface crack signal to noise is highest, while the detection sensitivity of internal defects remains at an acceptable level. The peak value of the total excitation current is controlled within the range of 50mA to 100mA to avoid overheating effects and ensure linear response.

[0064] The real and imaginary components of the induced voltage across the coil are synchronously acquired using a phase-sensitive detector. This phase-sensitive detector, based on the phase-locked amplification principle, extracts the real and imaginary components of each frequency component from the induced voltage through orthogonal adjustment using orthogonal reference signal pairs corresponding to each frequency component of the composite sinusoidal excitation current. Parallel digital phase-locked detection is implemented using a field-programmable gate array (FPGA), and both high- and low-frequency components are demodulated in real time. The sampling rate is set to at least 10 times that of the highest frequency component (i.e., ≥20 MS / s) to ensure signal integrity.

[0065] The relative helical motion is a uniform helical motion, achieved through a precision motion control system. Furthermore, the rotation angle and moving speed of the sample tungsten electrode satisfy the following relationship: ;in, The moving speed (mm / s) represents the speed at which the sample tungsten electrode moves along its axis. The value represents the rotational speed (rad / s), and D represents the diameter of the sample tungsten electrode (mm). For a standard tungsten electrode with a diameter of 2.4 mm, the moving speed ranges from 2 mm / s to 5 mm / s, and the rotational speed is adjusted accordingly to 15 rad / s to 30 rad / s. The motion control accuracy requires an axial positioning error ≤0.01 mm and an angular positioning error ≤1°. During the process of the sample tungsten electrode spiraling uniformly through the detection coil, multiple sets of eddy current detection signals are acquired at fixed time intervals. The acquisition interval is determined based on the Nyquist sampling theorem and motion parameters, i.e., at least two data points are acquired within each rotation cycle to ensure sufficient spatial sampling. Each set of eddy current detection signals also includes a timestamp and a sequence number.

[0066] Step 3: Perform ultrasonic testing on the sample tungsten electrode to obtain its three-dimensional grayscale model. Based on the sub-segment, extract the corresponding region image block from the three-dimensional grayscale model and retrieve the defect presence marker of the corresponding point based on the sub-segment.

[0067] In this embodiment, the ultrasonic detection specifically includes:

[0068] Define a linear scanning path parallel to the axial direction on the surface of the tungsten electrode sample, located on its side surface. The path length should cover the effective section of the tungsten electrode sample to be detected, typically the entire working length of the electrode, for example, 50 mm to 150 mm. Distribute this path uniformly into N nodes. The set of nodes that constitute the excitation and reception points is defined by the spatial sampling theorem and the detection resolution requirements. The specific value of N is determined by both the spatial sampling theorem and the detection resolution requirements: to effectively distinguish minute defects (such as cracks or pores ≥ 0.1 mm), the node spacing should not exceed half the smallest defect size. For a length of 150 mm, when N = 64, the node spacing is approximately 2.34 mm. To ensure sensitivity to even smaller defects, N can preferably be set to 128 or 256, thereby reducing the node spacing to approximately 1.17 mm or 0.59 mm. Node positioning is achieved using a high-precision translation stage (positioning accuracy better than 0.01 mm).

[0069] A non-contact laser ultrasonic testing system is employed, the core of which comprises a pulsed laser and a laser interferometer. The pulsed laser is used to excite ultrasonic waves; the short-pulse laser (pulse width on the order of nanoseconds, e.g., 10 ns) is focused onto the surface of a tungsten electrode sample through an optical lens group, forming a micrometer-sized spot. After the laser energy is absorbed by the electrode surface, broadband ultrasonic waves (mainly including longitudinal waves, transverse waves, and surface waves) are generated at the excitation point due to thermoelastic or ablation effects. The single-pulse energy of the pulsed laser is precisely controlled, typically on the order of millijoules (e.g., 1 mJ to 10 mJ), to ensure a sufficient signal-to-noise ratio for the ultrasonic signal while avoiding damage to the tungsten electrode surface. A laser interferometer is selected as the receiver to accurately measure the minute vibrations on the sample surface caused by the ultrasonic waves propagating to the receiving point.

[0070] The minimum center-to-center distance between the pulsed laser spot and the probe spot of the laser interferometer projected onto the tungsten electrode surface of the sample is ≥1 mm. This distance is set for two main reasons: First, to ensure that the probe spot of the laser interferometer is not affected by the strong plasma flash and surface thermal interference generated during pulsed laser excitation, thus avoiding signal saturation or severe deterioration of the signal-to-noise ratio; second, to provide sufficient path for the propagation of ultrasound in the material, allowing different wave modes (such as longitudinal and transverse waves) to be separated, facilitating subsequent signal analysis. In practical systems, this distance d is typically set to a fixed value between 1 mm and 3 mm.

[0071] The data acquisition process is as follows: At each node, a pulsed laser emits a laser beam for excitation, while simultaneously, a laser interferometer receives ultrasonic signals at all N nodes. Thus, for a single excitation, a 1×N row vector signal can be acquired. This process is repeated, traversing all N excitation points, and finally, all data are combined into an N×N two-dimensional signal matrix, i.e., the full matrix data. Each element in the full matrix data is a time-series signal that fully characterizes the physical process of ultrasonic waves propagating from the i-th excitation point to the j-th receiving point, including all waveform components from reflection from the electrode surface, scattering from internal defects, and reflection from the bottom surface.

[0072] Each element in the full matrix data represents the physical process of ultrasonic waves propagating from the excitation point to the receiving point, including all waveform components from reflection from the electrode surface, scattering from internal defects, and reflection from the bottom surface;

[0073] Each element in the full matrix data undergoes preprocessing. The preprocessing steps include at least time-base zero alignment, applying bandpass filtering with a passband frequency range of [1MHz, 30MHz], and energy attenuation compensation for each signal in the two-dimensional signal matrix using an exponential gain function. Time-base zero alignment: Due to potential slight time delays in the laser excitation and electron acquisition paths, time-base zero alignment is necessary for each scan signal. Specifically, the starting point of the surface wave (or direct longitudinal wave) propagating directly from the excitation point to the receiving point in each signal is selected as the time zero, and this starting point of all signals is aligned to the same time position using a digital delay line or interpolation algorithm. This ensures the timing accuracy of subsequent imaging and analysis. Bandpass filtering: Each signal is filtered using a bandpass filter (such as a Butterworth filter) with a passband frequency range of [1MHz, 30MHz]. This frequency range is determined based on the propagation characteristics of ultrasonic waves in tungsten materials and the size of the target defect. The lower frequency limit of 1MHz is primarily used to filter out low-frequency noise (such as environmental vibration and electronic equipment noise); the upper frequency limit of 30MHz is used to retain high-frequency components that can effectively distinguish minute defects (such as those at the 0.1mm level) while suppressing unwanted noise at higher frequencies. The filter order is typically chosen from 4th to 8th order to achieve a balance between passband flatness and transition band steepness.

[0074] The energy attenuation compensation refers to applying a gain function that grows exponentially with time to the signal for compensation, and its mathematical expression is:

[0075]

[0076] in, To compensate for the signal at the i-th excitation point and the j-th receiving point at time t, Let be the signal at time t at the i-th excitation point and the j-th receiving point. The compensation coefficient is t, which is a time variable; i represents the index of the excitation point, and j represents the index of the receiving point.

[0077] The compensation coefficient The value of is crucial, and its specific determination method is as follows: The amplitude attenuation curve of ultrasound in a defect-free tungsten sample is measured experimentally, and the attenuation coefficient (unit: dB / MHz / cm) is fitted. Compensation coefficient. Related to the attenuation coefficient and center frequency, it approximately satisfies For tungsten materials, typical attenuation coefficients range from 1 to 5 dB / MHz / cm. For an ultrasonic signal with a center frequency of 15 MHz, if the attenuation coefficient is taken as 3 dB / MHz / cm, the calculated compensation coefficient is approximately... In practical applications, the compensation coefficient can be... to Fine-tuning is performed within a certain range to ensure that the echo amplitude of deep defects is on a similar order of magnitude to that of near-surface defects, facilitating subsequent imaging and identification.

[0078] Three-dimensional imaging was performed on the preprocessed full matrix data using a time-delay overlay algorithm. The space occupied by the sample tungsten electrode was discretized into a three-dimensional pixel (voxel) mesh. For each voxel point P(x,y,z) in the mesh, the ultrasonic wave propagation time to all excitation and receiver point pairs (i,j) was calculated. Then, all data in the entire matrix at the corresponding time... Amplitude value at point Perform coherent superposition, and the superposition result is used as the gray value of the voxel point P. By traversing all voxels across the entire imaging area, the three-dimensional grayscale model of the sample tungsten electrode can be obtained. This model intuitively reflects the acoustic impedance changes of the electrode's internal structure, with defects (such as pores and inclusions) appearing as areas with different grayscale values ​​than the surrounding material.

[0079] Finally, based on the sub-segment (which spatially corresponds strictly to the sub-segment defined in step 2 for eddy current detection), the corresponding 3D region image block is extracted from the 3D grayscale model. Simultaneously, based on the axial position and rotation angle information of this sub-segment, defect presence markers for all spatial points within this sub-segment are retrieved from the coordinate position set and defect presence markers established in step 1. This step achieves precise spatial alignment between the ultrasonic detection data and the pre-known ground truth defect markers, providing crucial training samples for subsequent training of a model capable of associating multimodal detection signals with defect types.

[0080] Step 4: Construct a deep learning classification model to learn the mapping relationship between rotation angle, real and imaginary components, and defect existence markers between region image blocks and corresponding points in each sub-segment;

[0081] In this embodiment, reconstructing the three-dimensional grayscale model specifically includes:

[0082] The preprocessed full matrix data is processed using synthetic aperture focusing technology to reconstruct a three-dimensional grayscale model of the internal structure of the sample tungsten electrode. The three-dimensional space to be reconstructed is discretized into a regular grid composed of M pixels. The specific implementation method is as follows: a three-dimensional bounding box is determined according to the geometric dimensions of the sample tungsten electrode, and the bounding box is defined by X, Y, and Z dimensions respectively. , , The spacing is uniformly divided. The value of the spacing is based on the diffraction limit of ultrasound, and theoretically should be less than or equal to half the wavelength of the ultrasound to ensure resolution. For an ultrasound with a center frequency of 15MHz in tungsten (longitudinal wave velocity c is approximately 5170m / s), the wavelength is approximately 0.34mm, therefore the voxel size... , , The preferred setting is 0.1 mm to 0.15 mm. The value of M is determined by both the imaging area volume and the voxel size. For example, for an electrode area of ​​50 mm × 4 mm × 4 mm, when using a 0.1 mm voxel, M = 50 × 40 × 40 = 80,000.

[0083] For each signal in the full matrix data, calculate the total sound wave propagation time delay from the excitation point to any pixel in the regular grid, and then to the receiving point. The specific calculation formula is as follows:

[0084]

[0085] in, This represents the total sound wave propagation time delay from the signal at the i-th excitation point and the j-th receiving point to the k-th pixel. This represents the k-th pixel, where k is the pixel index. This represents the straight-line distance from the k-th pixel to the i-th excitation point. represents the straight-line distance from the k-th pixel to the j-th receiving point; c represents the longitudinal wave velocity of the ultrasonic wave in the tungsten electrode material of the sample, which is a preset known constant. Its value is determined by using a standard sample calibration method, employing a tungsten standard block of the same material, known thickness, and without defects, measuring the flight time of the ultrasonic wave across the known thickness, and calculating the precise sound velocity value. For pure tungsten material, this value is typically in the range of 5170±20 m / s. Accurately setting the c value is crucial for the focusing quality of the reconstructed image; an error exceeding ±1% will result in significant image blurring.

[0086] For each pixel in the regular grid, it is treated as a potential scattering point. All node pairs in the full matrix data are traversed, and the total acoustic propagation time delay for each node pair is calculated. The scattering time delay is then extracted from the corresponding compensated signal. The signal amplitude at that moment;

[0087] An amplitude weighting factor related to the propagation distance is applied to the signal amplitude extracted from the compensated signal to compensate for the diffusion attenuation of the sound wave propagating in the sample tungsten electrode material. The formula for calculating the amplitude weighting factor is as follows:

[0088]

[0089] in, This represents the magnitude weighting factor of the k-th pixel;

[0090] The amplitude weighting factor is introduced based on the spherical wave diffusion attenuation model of ultrasound propagation in a homogeneous isotropic medium. This factor compensates for the geometric diffusion effect of the sound wave propagating from the excitation point to the scattering point and then to the receiving point.

[0091] The weighted signal amplitudes contributed by all nodes at a given pixel are coherently superimposed. This coherent superposition process enhances signals from real defect scatterers by in-phase superposition, while canceling out-of-phase signals from noise or spurious scatterers by out-of-phase superposition, thus significantly improving the image's signal-to-noise ratio and contrast. The final grayscale values ​​of all pixels together constitute the three-dimensional grayscale model, which is stored as a three-dimensional array, with each element representing the ultrasonic reflection intensity at the corresponding spatial location.

[0092] In this embodiment, the eddy current detection signal is sampled at equal time intervals during the uniform spiral motion of the sample tungsten electrode, and the timestamp of each sampling is recorded. The real and imaginary components of the induced voltage collected at each timestamp are combined to form a complex impedance signal. On a complex plane with the real component as the horizontal axis and the imaginary component as the vertical axis, each complex impedance signal corresponds to a coordinate point. As the sample tungsten electrode performs a spiral scan, the continuously collected complex impedance signals form a trajectory on the complex plane. Based on all the coordinate points on the trajectory, a polynomial fitting method is used to obtain the functional expression of the trajectory, which serves as the ideal trajectory function. Specifically, a cubic polynomial is used for fitting, and the coefficients in the polynomial are determined by the least squares method. The goodness of fit is required to be no less than 0.95 to ensure the accuracy of the trajectory description.

[0093] In this embodiment, the deep learning classification model is constructed as follows:

[0094] For the acquired eddy current detection signals, a feature vector is constructed based on the rotation angle, real component value, and imaginary component value at each time stamp. The deviation between the coordinate point of the complex impedance signal at each time stamp and the ideal trajectory function is calculated. This deviation is obtained by calculating the vertical distance from the actual coordinate point to the ideal trajectory, reflecting the degree of deviation of the current monitoring position from the ideal defect-free state. The rotation angle, real component value, imaginary component value, and deviation are combined to form a numerical eddy current feature vector. In practical applications, to enhance the temporal context information, feature vectors from K consecutive time steps are typically used to form a feature sequence, where K ranges from 5 to 10, corresponding to a rotation angle range of approximately 15-30°.

[0095] From the 3D grayscale model, extract the regional image patch that corresponds spatially to the sub-segment to serve as the regional image patch feature;

[0096] The eddy current feature vector, the regional image patch feature, and the defect presence label of the corresponding point are associated to form a training sample; the data of all spatial location points are sorted to form the final training sample set, and divided into training set, validation set and test set in a ratio of 6:2:2; the eddy current feature vector and regional image patch feature in the training set are used as input, and the defect presence label of the corresponding point is used as label to train the deep learning classification model.

[0097] The deep learning classification model employs a dual-branch convolutional neural network, comprising: an input layer that receives a sequence of eddy current features of dimension [H, 4], where H is the time step and 4 is the feature dimension. The model includes a one-dimensional convolutional branch for processing the eddy current feature vector and a three-dimensional convolutional branch for processing region image patch features; a feature fusion module that concatenates the feature vectors output from the fully connected layers at the ends of the one-dimensional and three-dimensional convolutional branches; and a classifier consisting of at least one fully connected layer, whose input is the feature vector concatenated by the feature fusion module. The final output layer of the classifier uses a Softmax activation function to output a defect presence marker for each point, and a confidence distribution vector indicating the point's classification into each defect category. Each element in the vector contains its coordinates and the confidence level of belonging to a particular defect category.

[0098] Supervised training is performed on this two-branch deep learning classification model, using an optimizer to iteratively update model parameters to minimize the weighted cross-entropy loss function. Model training is considered complete when both of the following conditions are met: First, the training loss function value decreases by less than a preset threshold over 20 consecutive epochs, indicating convergence. This threshold is set based on the magnitude of the loss function and training stability, ensuring timely stopping when the loss changes only slightly to avoid unnecessary computation. Second, the multi-class macro-F1 score calculated on the validation set does not increase over 15 consecutive training epochs to prevent overfitting. Training terminates immediately when either of the above conditions is triggered, and the optimal model parameters from the validation set are saved. The macro-F1 score comprehensively considers precision and recall across all classes and is a robust metric for evaluating multi-class classification performance. On the test set, a model is considered excellent if its overall classification accuracy for all defect categories is not lower than threshold 1, and its recall for each defect category is not lower than threshold 2. These thresholds were determined based on the actual needs of industrial inspection: an overall accuracy of 95% ensures the reliability of the inspection system, and a recall rate of 90% per class guarantees that the false negative rate for various defects, especially dangerous defects (such as cracks), is kept within an acceptable range. The final model saved is the parameter version that achieves the highest macro F1 score on the validation set.

[0099] Step 5: Based on the deep learning classification model, obtain the confidence level of each point of the tungsten electrode to be tested belonging to various types of defects, and judge them according to the preset defect tolerance standard to determine whether the tungsten electrode to be tested is qualified.

[0100] In this embodiment, determining whether the tungsten electrode to be tested is qualified specifically includes:

[0101] For the tungsten electrode under test, repeat the detection process described in steps 2 and 3 to obtain its complete eddy current detection signal and ultrasonic 3D grayscale model. Specifically, the electrode under test is passed through the eddy current detection coil with the exact same parameters as in the training phase (including the same composite sinusoidal excitation current frequency and amplitude, and the same uniform helical motion parameters v and w), and the same laser ultrasonic scanning configuration (including the same number of nodes N, spot spacing, etc.) is used to acquire the full matrix data. The ultrasonic 3D grayscale model is then generated through a completely consistent preprocessing and 3D reconstruction process. This consistency ensures the consistency of the feature space from the training model to the actual application. Analysis positions are selected along the axial direction of the tungsten electrode under test at fixed spatial intervals. The value of the spatial interval is based on two factors: ensuring the comprehensiveness of the detection and considering computational efficiency. Typically, the spatial interval is set to 1 / 4 to 1 / 2 of the tungsten electrode diameter. For a standard 2.4mm diameter electrode, the spatial interval can be 0.6mm to 1.2mm, with 1.0mm being the preferred value. At each analysis location, the eddy current feature vector of the corresponding sub-segment is extracted and the corresponding regional image patch features are extracted from the three-dimensional grayscale model. The extraction method is exactly the same as that used when constructing the training set in step 4.

[0102] The eddy current feature vector and region image patch features are input into the trained deep learning classification model to obtain the confidence distribution vector output by the model. The maximum value in the confidence distribution vector is used as the reliability index. If the reliability index is greater than a preset confidence threshold, the point is marked as a suspected defect location. The determination of the confidence threshold is based on the model's performance statistics on the test set. Specifically, the reliability index distribution curve of all samples on the test set is plotted, with the primary goal of ensuring a high-confidence true positive rate. Typically, the confidence threshold is set between 0.7 and 0.9. A preferred method is to observe the impact of the confidence threshold on precision and recall on the test set, using it as a variable. The threshold that achieves a high F1 score and a precision exceeding 90% is selected as the confidence threshold, for example, 0.85. This threshold setting aims to effectively filter out most locations where the model's prediction is uncertain, focusing on high-confidence anomalous signals for subsequent comprehensive judgment, balancing the risks of missed detections and false positives.

[0103] The defect tolerance standard is based on a comprehensive quality index, and the specific calculation formula is as follows:

[0104]

[0105] in, The overall quality index of the tungsten electrode to be tested. Let be the severity weight of the q-th defect category. Let be the confidence level that the a-th suspected defect location belongs to the q-th defect category. Let q be the spatial weight of the a-th suspected defect location; a is the index of the suspected defect location; A is the number of suspected defect locations; q is the defect category index; and Q is the number of defect categories.

[0106] The comprehensive quality index constructs a multi-dimensional quantitative evaluation system. The formula distinguishes the essential differences in the impact of different types of defects on product function by introducing defect severity weights; for example, crack defects are given higher weights because they are prone to structural failure. A spatial weighting function identifies key functional areas of the product, allowing the same defect to have different impacts on overall quality depending on its location. Finally, the defect confidence score output by the model objectively reflects the probability of the defect's existence. This systematic integration of defect severity, locational criticality, and probability of existence allows the final comprehensive quality index to more accurately represent the total potential quality risk of a workpiece, rather than irrationally equating all abnormal signals. Its rationality lies in simulating the complex factors considered by human experts when making comprehensive scrap judgments and transforming them into a repeatable and quantifiable calculation model. In this comprehensive quality index formula, the dependent variable is the final calculated comprehensive quality index (…). ), which means the overall quality risk quantification value of the tungsten electrode to be tested; while the independent variables are the various contributing factors that constitute this risk value, mainly including the confidence level of the defect category predicted by the model ( ), preset defect severity weights ( ) and the calculated spatial location weights ( These independent variables are able to construct a comprehensive quality index because they capture three indispensable dimensions of quality assessment: It provides probabilistic evidence of whether a certain defect exists, which is the factual basis for risk calculation; It answers the question "If the defect exists, how serious are the consequences?" based on prior knowledge, and acts as a regulator of risk magnitude. This assesses whether the location of the defect would worsen the consequences, thus amplifying the risk impact. The combined effect of these three factors makes... This model can serve as a comprehensive, weighted expected risk value to objectively characterize the quality status of electrodes. At any suspected defect location, as the model's confidence in its classification of a particular defect increases, especially for defect types with high severity weights, its contribution to the overall quality index significantly increases. Similarly, when a defect appears near the center of a pre-defined critical area, its spatial weight increases due to the exponential function, thus increasing the overall quality index. Conversely, a low-confidence defect, a defect type with a low severity weight, or a defect far from the critical area contributes very little to the overall quality index. Therefore, the value of the overall quality index monotonically increases with the occurrence and increase of high-confidence defects, high-risk defects, and defects located in critical areas, accurately depicting the worsening trend of workpiece quality risk.

[0107] The severity weight The severity is determined using the Delphi method (expert survey method) combined with historical failure analysis data. The specific values ​​are as follows: longitudinal surface cracks (q=1) and transverse surface cracks (q=2) are given the highest severity weight, typically set to 1.0, because they may become stress concentration sources and propagate under the influence of the electric arc, leading to electrode breakage or arc instability; internal pores (q=3) interfere with current distribution and heat dissipation, and are given a medium weight, typically set to 0.6 to 0.8; internal inclusions (q=4) have a relatively small impact and are given a lower weight, typically set to 0.3 to 0.5. An example set of values ​​is: Severity(1)=1.0, Severity(2)=1.0, Severity(3)=0.7, Severity(4)=0.4.

[0108]

[0109] in, Let a be the coordinates of the a-th suspected defect location. This refers to the coordinates of the center of the critical area of ​​the tungsten electrode to be tested. This critical area is typically the section of the tungsten electrode that is most frequently clamped in the welding fixture or has the highest current density. The parameter used to control the width of the influence range of the critical region is related to the axial length of the critical region, and is usually set to one-quarter of the axial length. For example, if the length of the critical region is 10mm, then... 2.5mm is acceptable. This setting allows for... It exhibits a significant spatial weight amplification effect within a range of approximately ±5 mm. This is a coefficient used to adjust the weight magnitude of critical areas. Its value determines the penalty for defects within the critical area, and it is typically set between 0.5 and 2.0 based on engineering experience. If the reliability requirements of the critical area are extremely high, A higher value, such as 1.5, can be chosen, meaning that defects in the critical area have a significant impact on... The contribution will be significantly amplified (up to 2.5 times that of non-critical areas).

[0110] The comprehensive quality index of the tungsten electrode to be tested is compared with a preset pass threshold. If the comprehensive quality index of the tungsten electrode to be tested is not greater than the pass threshold, the tungsten electrode to be tested is determined to be qualified. If the comprehensive quality index of the tungsten electrode to be tested is greater than the pass threshold, the tungsten electrode to be tested is determined to be unqualified.

[0111] The determination of the acceptance threshold is based on statistical process control principles and product quality standards. The specific method involves collecting a large number of electrode samples that have been confirmed as qualified through traditional manual testing or long-term use verification, calculating the distribution of their comprehensive quality index, and taking the percentile of this distribution (e.g., the 95th or 97.5th percentile) as the reference benchmark for the acceptance threshold. Simultaneously, adjustments can be made based on the customer's acceptable quality level. For example, if the 95th percentile of the calculated comprehensive quality index for qualified electrodes is 5.0, and the customer's acceptable quality level requirement is 2.5%, then the acceptance threshold can be initially set at 5.0, and fine-tuned after verifying its effectiveness in small-batch trials. This threshold setting method ensures that the judgment standard is consistent with the quality level of historically qualified products.

[0112] Please see Figure 2 The present invention also provides a system for detecting appearance defects in argon arc welding tungsten electrodes. This system is used to implement the aforementioned method for detecting appearance defects in argon arc welding tungsten electrodes, and includes:

[0113] The defect category marking module is used to discretize the sample tungsten electrode with known defect distribution into several points with spatial relative information, and assign a corresponding defect presence mark to each point based on the defect category. The defect categories include surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions.

[0114] The eddy current scanning module is used to apply a composite sinusoidal excitation current to the detection coil. As the sample tungsten electrode passes through the central hole of the detection coil in a relative helical motion along the axial direction of the detection coil, multiple sets of eddy current detection signals are acquired. The eddy current detection signals include at least the sub-segment of the sample tungsten electrode in the detection coil, the rotation angle, and the real and imaginary parts of the induced voltage at both ends of the coil.

[0115] The ultrasonic testing module is used to perform ultrasonic testing on the sample tungsten electrode to obtain its three-dimensional grayscale model. Based on the sub-segment, the corresponding area image block is extracted from the three-dimensional grayscale model, and the defect presence marker of the corresponding point is retrieved based on the sub-segment.

[0116] The classification model building module is used to build a deep learning classification model to learn the mapping relationship between rotation angle, real and imaginary components, and defect existence markers between region image blocks and corresponding points in each sub-segment.

[0117] The detection and judgment module is used to obtain the confidence level of each point of the tungsten electrode to be detected belonging to various types of defects based on a deep learning classification model, and to judge it according to the preset defect tolerance standard to determine whether the tungsten electrode to be detected is qualified.

[0118] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for detecting appearance defects in argon arc welding tungsten electrodes.

[0119] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0120] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0121] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0122] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for detecting visual defects in argon arc welding tungsten electrodes, characterized in that, The specific steps include: Step 1: Discretize the sample tungsten electrode with known defect distribution into several points with spatial relative information, and assign a corresponding defect presence label to each point based on the defect category, which includes surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions. Step 2: Apply a composite sinusoidal excitation current to the detection coil. As the sample tungsten electrode moves through the central hole of the detection coil in a relative spiral motion along the axial direction of the detection coil, acquire multiple sets of eddy current detection signals. The eddy current detection signals include at least the sub-segment of the sample tungsten electrode in the detection coil, the rotation angle, and the real and imaginary parts of the induced voltage at both ends of the coil. Step 3: Perform ultrasonic testing on the sample tungsten electrode to obtain its three-dimensional grayscale model. Based on the sub-segment, extract the corresponding region image block from the three-dimensional grayscale model and retrieve the defect presence marker of the corresponding point based on the sub-segment. Step 4: Construct a deep learning classification model to learn the mapping relationship between rotation angle, real and imaginary components, and defect existence markers between region image blocks and corresponding points in each sub-segment; Step 5: Based on the deep learning classification model, obtain the confidence level of each point of the tungsten electrode to be tested belonging to various types of defects, and judge them according to the preset defect tolerance standard to determine whether the tungsten electrode to be tested is qualified. The defect tolerance standard is based on a comprehensive quality index, and the specific calculation formula is as follows: in, The overall quality index of the tungsten electrode to be tested. Let be the severity weight of the q-th defect category. Let be the confidence level that the a-th suspected defect location belongs to the q-th defect category. Let be the spatial weight of the a-th suspected defect location; a is the index of the suspected defect location; A is the number of suspected defect locations; q is the defect category index; Q is the number of defect categories. in, Let a be the coordinates of the a-th suspected defect location. The coordinates are the center of the critical area of ​​the tungsten electrode to be tested. Parameters for controlling the width of the influence range of critical areas. This is a coefficient used to adjust the weight magnitude of key areas; The comprehensive quality index of the tungsten electrode to be tested is compared with a preset pass threshold. If the comprehensive quality index of the tungsten electrode to be tested is not greater than the pass threshold, the tungsten electrode to be tested is determined to be qualified. If the comprehensive quality index of the tungsten electrode to be tested is greater than the pass threshold, the tungsten electrode to be tested is determined to be unqualified.

2. The method for detecting appearance defects in argon arc welding tungsten electrodes according to claim 1, characterized in that, Assigning the existence marker of the defect specifically includes: The discretization into several points with spatial relative information specifically involves: obtaining the three-dimensional coordinates of each point on the sample tungsten electrode to form a set of coordinate positions; the defect existence marker is a four-digit code, and the four codes from left to right correspond one-to-one with surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions, respectively. When a point belongs to a certain type of defect, the code of that position in the defect existence marker corresponding to that point is set to 1, otherwise it is set to 0.

3. The method for detecting appearance defects in argon arc welding tungsten electrodes according to claim 1, characterized in that, The acquisition process of the eddy current detection signal specifically includes: The composite sinusoidal excitation current is formed by the linear superposition of high-frequency and low-frequency components, wherein the frequency range of the high-frequency component is [1.2MHz, 2MHz], the frequency range of the low-frequency component is [200kHz, 400kHz], and the current amplitude of the high-frequency component is smaller than that of the low-frequency component. The real and imaginary components of the induced voltage at both ends of the coil are synchronously acquired by a phase-sensitive detector. The phase-sensitive detector is based on the lock-in amplification principle and uses orthogonal reference signal pairs corresponding to each frequency component in the composite sinusoidal excitation current to extract the real and imaginary components of each frequency component from the induced voltage through orthogonal adjustment. The relative spiral motion is a uniform spiral motion, and the rotation angle and moving speed of the sample tungsten electrode satisfy the following relationship: ;in, This indicates the moving speed, specifically the speed at which the sample tungsten electrode moves along its axis. The value represents the rotational speed, and D represents the diameter of the sample tungsten electrode.

4. The method for detecting appearance defects in argon arc welding tungsten electrodes according to claim 1, characterized in that, The ultrasonic detection specifically includes: An ultrasonic testing device is used to scan the sample tungsten electrode to obtain full matrix data of its internal structure. The full matrix data is preprocessed and processed using synthetic aperture focusing technology to reconstruct a three-dimensional grayscale model of the sample tungsten electrode. The three-dimensional grayscale model consists of the three-dimensional coordinates and grayscale values ​​of each point, wherein the grayscale values ​​of each point are obtained by coherent superposition of the preprocessed ultrasonic signals.

5. The method for detecting appearance defects in argon arc welding tungsten electrodes according to claim 4, characterized in that, Reconstructing the three-dimensional grayscale model specifically includes: The three-dimensional space to be reconstructed is discretized into a regular grid. For each grid point, the propagation time delay of each acquired signal at that point is calculated based on the ultrasonic propagation principle. The signal amplitude at the corresponding time delay is extracted from the preprocessed ultrasonic signal and weighted coherently superimposed to obtain the gray value of that grid point. The gray values ​​of all grid points together constitute a three-dimensional gray model.

6. The method for detecting visual defects in argon arc welding tungsten electrodes according to claim 2, characterized in that: During the uniform spiral motion of the sample tungsten electrode, the eddy current detection signal is sampled at equal time intervals, and the timestamp of each sampling is recorded. The real and imaginary components of the induced voltage collected at each timestamp are combined to form a complex impedance signal. On the complex plane formed by the real component as the horizontal axis and the imaginary component as the vertical axis, each complex impedance signal corresponds to a coordinate point. As the sample tungsten electrode performs eddy current scanning, the continuously collected complex impedance signals form a trajectory on the complex plane. Based on all the coordinate points on the trajectory, a polynomial fitting method is used to obtain the functional expression of the trajectory, which is used as the ideal trajectory function.

7. The method for detecting visual defects in argon arc welding tungsten electrodes according to claim 6, characterized in that... Construct the deep learning classification model: For the acquired eddy current detection signal, a feature vector is constructed based on the rotation angle, real component value, and imaginary component value at each time stamp; the deviation between the coordinate point of the complex impedance signal and the ideal trajectory function at each time stamp is calculated, and the deviation is obtained by calculating the vertical distance from the actual coordinate point to the ideal trajectory; the rotation angle, real component value, imaginary component value, and deviation are combined to form a numerical eddy current feature vector. From the 3D grayscale model, extract the regional image patch that corresponds spatially to the sub-segment to serve as the regional image patch feature; The eddy current feature vector, the regional image patch feature, and the defect presence label of the corresponding point are associated to form a training sample; the data of all spatial location points are sorted to form the final training sample set, and divided into training set, validation set and test set in a ratio of 6:2:2; the eddy current feature vector and regional image patch feature in the training set are used as input, and the defect presence label of the corresponding point is used as label to train the deep learning classification model. The deep learning classification model employs a dual-branch convolutional neural network, comprising: a one-dimensional convolutional branch for processing eddy current feature vectors, and a three-dimensional convolutional branch for processing region image patch features. A feature fusion module is used to concatenate the feature vectors output by the fully connected layers at the end of the one-dimensional convolutional branch and the three-dimensional convolutional branch; a classifier is composed of at least one fully connected layer, whose input is the feature vector concatenated by the feature fusion module; the final output layer of the classifier adopts the Softmax activation function and outputs the defect existence label of each point and the confidence distribution vector of each point belonging to each defect category.

8. The method for detecting appearance defects in argon arc welding tungsten electrodes according to claim 7, characterized in that, Determining whether the tungsten electrode to be tested is qualified specifically includes: Eddy current testing and ultrasonic testing are performed on the tungsten electrode to be tested to obtain its eddy current testing signal and ultrasonic three-dimensional grayscale model; eddy current feature vectors and regional image block features corresponding to sub-segments of the tungsten electrode to be tested are extracted; the eddy current feature vectors and regional image block features are input into the trained deep learning classification model to obtain the confidence distribution vector output by the model; the maximum value in the confidence distribution vector is used as the reliability index. If the reliability index is greater than the preset confidence threshold, the point is marked as a suspected defect location.

9. A system for detecting visual defects in argon arc welding tungsten electrodes, characterized in that, The argon arc welding tungsten electrode appearance defect detection system is used to implement the argon arc welding tungsten electrode appearance defect detection method according to any one of claims 1-8, including: The defect category marking module is used to discretize the sample tungsten electrode with known defect distribution into several points with spatial relative information, and assign a corresponding defect presence mark to each point based on the defect category. The defect categories include surface longitudinal cracks, surface transverse cracks, internal pores, and internal inclusions. The eddy current scanning module is used to apply a composite sinusoidal excitation current to the detection coil. As the sample tungsten electrode passes through the central hole of the detection coil in a relative helical motion along the axial direction of the detection coil, multiple sets of eddy current detection signals are acquired. The eddy current detection signals include at least the sub-segment of the sample tungsten electrode in the detection coil, the rotation angle, and the real and imaginary parts of the induced voltage at both ends of the coil. The ultrasonic testing module is used to perform ultrasonic testing on the sample tungsten electrode to obtain its three-dimensional grayscale model. Based on the sub-segment, the corresponding area image block is extracted from the three-dimensional grayscale model, and the defect presence marker of the corresponding point is retrieved based on the sub-segment. The classification model building module is used to build a deep learning classification model to learn the mapping relationship between rotation angle, real and imaginary components, and defect existence markers between region image patches and corresponding points in each sub-segment. The detection and judgment module is used to obtain the confidence level of each point of the tungsten electrode to be detected belonging to various types of defects based on a deep learning classification model, and to judge it according to the preset defect tolerance standard to determine whether the tungsten electrode to be detected is qualified.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements a method for detecting appearance defects in argon arc welding tungsten electrodes as described in any one of claims 1-8.

Images