Defect positioning method based on fusion of weld defect features and trajectory tracking data

Abstract

The invention relates to a defect positioning method based on fusion of weld defect features and trajectory tracking data, and belongs to the technical field of weld defect detection and positioning. The method comprises the following steps: capturing welding seam track dynamic data and defect feature data, constructing a dynamic coordinate system based on a welding seam initial feature point, and establishing double-data-set reference mapping; performing multi-physics field interference decoupling correction on the trajectory data, and performing cross-modal feature purification and core feature consistency verification on the defect data; converting the preprocessed data into a feature form adaptive to fusion, and constructing a welding process-defect formation mechanism association network to regulate and control fusion weight; and finally, reconstructing a three-dimensional dynamic contour of the welding seam, calling dynamic positioning logic to position the defect, and outputting a result carrying the process-defect causal confidence coefficient. The positioning precision is improved through multi-dimensional data fusion and mechanism association, and technical support is provided for welding quality management and control.

Description

Technical Field

[0001] This invention belongs to the field of weld defect detection and location technology, and is a defect location method based on the fusion of weld defect features and trajectory tracking data. Background Technology

[0002] In the industrial manufacturing sector, welds, as the core structure connecting components, directly determine the overall strength, safety, and service life of equipment or structures. Accurate location of weld defects is a crucial step in ensuring welding quality. As the manufacturing industry continuously increases its demands for product precision and reliability, traditional weld defect location technologies are gradually becoming insufficient to meet the inspection needs under complex working conditions, highlighting a growing technological bottleneck.

[0003] Currently, most mainstream weld defect detection methods in the industry rely on single-modal sensor data, such as ultrasonic testing and radiographic testing. While these methods can acquire partial physical properties or morphological features of defects, they lack a unified spatiotemporal reference and logical correlation between different types of data. For example, trajectory tracking systems can record the welding torch's movement trajectory and changes in process parameters, but this data and defect feature data are often in independent coordinate systems. The spatiotemporal resolution is not adapted to the dynamic changes during the welding process, making it difficult to achieve accurate mapping between the two and preventing the tracing of the root cause of defects from the process-defect correlation dimension.

[0004] In the data preprocessing stage, existing technologies are insufficient in handling multi-physics interference. Interference factors such as mechanical vibration and arc thermal disturbance during welding can easily cause coordinate drift in trajectory data. Traditional correction models are mostly designed for a single interference source and do not consider the coupling effect of mechanical vibration and thermal deformation, resulting in limited correction accuracy. Simultaneously, defect feature data often contains pseudo-defect signals such as welding spatter and oxide scale. Existing feature purification methods mostly rely on single-mode noise suppression strategies and lack a consistency verification mechanism for cross-modal features, making it difficult to effectively remove pseudo-signals and thus affecting the authenticity of defect features and subsequent positioning accuracy.

[0005] Furthermore, existing defect location methods generally lack a deep integration of welding processes and defect formation mechanisms. Most techniques locate defects solely based on surface morphological features, failing to establish a quantitative correlation between process parameter fluctuations and defect formation. This results in location results that only reflect the spatial position of the defect, unable to simultaneously provide the causal relationship between process anomalies and defect formation. Moreover, weld 3D contour reconstruction is mostly based on static data, neglecting the dynamic impact of thermal deformation during welding on the contour datum, further exacerbating location errors and failing to meet the practical needs of high-precision, intelligent welding quality control. Therefore, developing a weld defect location method capable of deep fusion of multi-source data and synergistic correlation between process and mechanism has become an urgent technical problem to be solved in the industry. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, this invention provides a defect localization method based on the fusion of weld defect features and trajectory tracking data; The objective of this invention can be achieved through the following technical solutions: A defect localization method based on the fusion of weld defect features and trajectory tracking data includes: S1: Capture dynamic trajectory data and defect feature data of the weld formation process; adjust the spatiotemporal resolution of trajectory data based on arc energy fluctuation characteristics, and enhance the multi-dimensional capture density of defect features through molten pool morphology changes; construct a dynamic coordinate system based on weld initiation feature points, and establish a benchmark mapping relationship between the two datasets through real-time coordinate system calibration. S2: Perform multi-physics field interference decoupling correction on the trajectory dynamic data, construct a coordinate drift correction model, and use the process parameter fluctuation law to back-verify and correct coordinate deviation; perform cross-modal feature purification on the defect feature data, and eliminate false defect signals through core defect feature consistency verification. S3: The preprocessed trajectory data and defect feature data are transformed into feature forms that meet the fusion requirements; a welding process-defect formation mechanism association network is constructed, and the scenario-based fusion weights of the features are adjusted to generate a fusion feature body containing the causal correlation between process and defect; S4: Reconstruct the three-dimensional dynamic contour of the weld based on the trajectory dynamic data, calculate the contour thermal deformation amount as the positioning reference correction parameter; call the adapted dynamic positioning logic based on the fused feature body, perform the corresponding positioning operation on the defect, and output the defect positioning result carrying the process-defect causal confidence.

[0007] As a preferred embodiment of the present invention, the method for adjusting the spatiotemporal resolution of the arc energy fluctuation characteristics trajectory data is as follows: The arc energy fluctuation coefficient is calculated by the real-time change rate of arc voltage and the effective value of welding current to quantify the degree of dynamic change of arc energy. Based on the fluctuation coefficient, a step-by-step adjustment rule is set. For coefficients below a preset threshold, a baseline spatiotemporal resolution is adopted, and for coefficients exceeding the threshold, the resolution is increased according to a gradient. The position of the weld fusion line is synchronously associated, and the resolution sensitivity is further increased in the area near the fusion line to perform dynamic resolution adjustment adapted to the arc energy fluctuation and the key areas of the weld.

[0008] Specifically, the method for adjusting the spatiotemporal resolution of the arc energy fluctuation characteristics trajectory data is as follows: By simultaneously acquiring the molten pool contour, temperature field distribution, and liquid metal flow trajectory through infrared imaging and visual sensing, quantitative parameters of the molten pool dynamic morphology are extracted, and a quantitative characterization model of molten pool morphology changes is constructed. Morphology-defect correlation rules are established to analyze the mapping relationship between different morphological change characteristics and potential defect types, and to determine the key dimensions and triggering conditions for defect feature capture. Dynamic capture density control is implemented, and based on the detected high-risk morphological change characteristics, the sampling frequency of ultrasonic echo signals and the layered scanning density of X-ray images are simultaneously increased. The capture dimension combination is adapted in combination with the characteristics of the weld formation stage to achieve precise matching between the multi-dimensional capture density of defect features and the molten pool morphology risk.

[0009] Specifically, the dynamic coordinate system is constructed from the weld initiation feature points, and the specific method is as follows: By combining visual recognition with laser contour scanning, the starting point of the weld fusion line or the starting position of the bevel root is located. Combined with the starting coordinate reference in the welding process document, the physical anchor point of the coordinate system origin is determined. Define the dynamic coordinate axis direction, with the welding travel direction as the X-axis, and update the axial direction in real time by tracking the movement trajectory of the welding torch; use the weld section normal direction as the Y-axis, and dynamically calibrate the workpiece surface geometric features; use the molten pool depth direction as the Z-axis, and correct the axial perpendicularity by combining the molten depth data from ultrasonic testing. Real-time coordinate system calibration is performed. By using high-frequency acquired trajectory dynamic data and weld formation feedback, the coordinate offset caused by thermal deformation is calculated, and the origin and axial parameters of the coordinate axes are periodically corrected.

[0010] Specifically, the baseline mapping relationship of the two datasets includes: Spatial reference mapping: The spatial coordinates of the defect feature dataset are mapped to the spatial framework of the trajectory dynamic dataset through the coordinate transformation rules of the dynamic coordinate system. This establishes a one-to-one correspondence between the spatial location of the defect features and the spatiotemporal coordinates of the trajectory data, and associates the spatial markers of the key areas of the weld. Time reference mapping: A time alignment reference is established based on welding process nodes, and the acquisition time of defect feature data is associated with the sampling time of trajectory dynamic data. The correspondence between the two in the welding time sequence is strengthened by process stage labels. Feature association mapping: Establish association rules between the process-spatiotemporal features of trajectory dynamic data and the morphological-attribute features of defect feature data, and map the process anomaly features of trajectory data with the risk features of defect data.

[0011] Specifically, the method for constructing the coordinate drift correction model is as follows: The vibration signal is spectrally decomposed using a modal separation algorithm to distinguish between the rigid displacement component and the flexible deformation component of the mechanical vibration; combined with the three-dimensional temperature field distribution acquired by an infrared thermal imager, the material temperature-deformation coefficient matrix is ​​introduced to calculate the coordinate offset caused by the arc thermal disturbance. A dynamic coupling correction model is constructed. Based on the time-varying coupling law of mechanical vibration and thermal disturbance during welding, a three-dimensional coordinate drift function containing a time decay factor is established to transform the decoupled disturbance components into correction parameters for the trajectory coordinates. Perform reverse verification of process parameters, extract transient fluctuation characteristics of welding current and voltage to construct interference-process correlation model, compare the model-predicted drift with the actual measurement residual, iteratively optimize the coefficient matrix of the coupled correction model, and dynamically adapt the coordinate correction accuracy to the process fluctuation state.

[0012] Specifically, the method for performing cross-modal feature purification is as follows: Variational mode decomposition (VMD) algorithm is used to separate defect vibration signals from structural noise in ultrasonic echo signals. Intrinsic vibration modes determined solely by the physical properties of the defect are selected based on modal energy ratio. Multi-scale edge detection and morphological thinning are combined to extract the topological skeleton and edge connectivity features of the defect region from the X-ray images, while simultaneously suppressing pseudo-contours formed by welding spatter. The grayscale stretching range of the X-ray images is optimized based on the frequency characteristics of the ultrasonic vibration modes. At the same time, the spatial boundary of the X-ray topology is used to constrain the effective analysis range of the ultrasonic modes, forming a mutual reinforcement mechanism of the two modal features and weakening the detection blind zone of a single mode.

[0013] Specifically, the core defect feature consistency verification includes: A multi-dimensional matching framework is constructed, establishing the correlation verification dimensions of core defect features from three levels: spatial location, morphological features, and physical properties. The spatial verification features are based on the positional coincidence of the spatial verification features in the weld coordinate system; the morphological features are based on the consistency of the contour geometric features; and the physical properties are based on the intrinsic matching relationship between ultrasonic vibration characteristics and X-ray attenuation characteristics. Cross-modal cross-validation is performed, constraining the effective analysis boundary of the X-ray topology by the spatial distribution range of the ultrasonic vibration modes, and at the same time verifying the authenticity of the ultrasonic vibration signal source by the defect contour morphology of the X-ray image, forming a mutually corroborating verification logic. A pseudo-signal identification mechanism is established. For features that appear only in a single mode or have contradictory multi-dimensional matching, the mechanism is compared with a feature library of common interference sources in welding processes to identify them as pseudo-defect signals and remove them.

[0014] Specifically, the method for transforming the preprocessed trajectory data and defect feature data into a feature format suitable for fusion requirements is as follows: Based on the welding time sequence, continuous spatiotemporal segments are divided, the changing trend of trajectory coordinates of each segment is extracted, the welding process parameter fluctuation index of the corresponding segment is associated, and spatiotemporal information and process information are coupled through matrix operation to form a process-spatiotemporal coupling feature that characterizes the process state at a specific spatiotemporal location. Extract the core morphological parameters of the defects, associate them with the physical attribute data of the defects, and combine them with the feature weight rules of the defect type to construct morphological-attribute association features that are deeply bound to the morphology and attributes. The transformed features are uniformly mapped to a preset feature space. Through time alignment, the features are bound to welding process stage labels. Dimension normalization is performed synchronously. Through feature scaling, parameters of different physical dimensions are unified to the same numerical range.

[0015] Specifically, the welding process-defect formation mechanism correlation network includes: Using process-spatiotemporal coupling characteristics as the core input, it incorporates welding material parameters, groove structure parameters, and spatial coordinate information in a dynamic coordinate system to construct an input system covering process dynamics and spatial positioning. It has built-in welding thermodynamics and fluid dynamics models to establish a quantitative correlation between process characteristics and the spatial location and formation mechanism of defects, and to perform a deep mapping between process fluctuations and defect generation areas and their underlying mechanisms. Based on the dynamic adjustment of feature weights in the welding stage, it outputs the location results of defects in the dynamic coordinate system, risk level, and appropriate process adjustment suggestions.

[0016] Specifically, the three-dimensional dynamic profile of the weld includes: The spatiotemporal coordinate sequence and process parameter fluctuation characteristics of dynamic trajectory data are analyzed to construct a spatial position-process state mapping relationship, transforming discrete trajectory points into three-dimensional feature points containing forming influencing factors. Based on the spatial topological relationship of feature points and process association rules, three-dimensional contours of each stage are generated and formed into a continuous contour framework through temporal splicing. Combined with contour deviation feedback from real-time sensing, the detailed parameters of the framework are dynamically corrected to keep the reconstructed contour consistent with the actual forming state of the weld in spatiotemporal terms.

[0017] Specifically, the method for performing the corresponding location operation on the defect is as follows: A feature-coordinate dynamic binding mechanism is established. Using the dynamic coordinate system as the spatial reference, the cross-modal purified defect features are associated with the coordinate system spatial parameters, the spatial attributes of the defect features are extracted, and the feature data are mapped to the initial three-dimensional coordinates under the dynamic coordinate system through matrix transformation. The process traceability-assisted positioning correction is performed by calling the process-spatiotemporal coupling features in the welding process-defect formation mechanism association network, comparing the process fluctuation area corresponding to the initial coordinates of the defect, and inferring the possible extension boundary of the defect through the spatiotemporal range of the process anomaly, thus correcting the coordinate deviation caused by the sensor blind zone in the initial positioning, and forming a two-way verification of feature positioning and process traceability. Multi-source feedback closed-loop calibration is performed, and the weld seam three-dimensional dynamic contour is reconstructed in real time. The defect location coordinates are compared with the geometric features of the contour, and the accurate three-dimensional coordinates of the defect in the dynamic coordinate system and the location information of the corresponding welding process stage are output.

[0018] The beneficial effects of this invention are as follows: (1) By setting up a dynamic coordinate system based on the starting feature point of the weld and calibrating it in real time, and combining the dual dataset benchmark mapping mechanism of space, time and feature association, and at the same time constructing a multi-physics field interference decoupling correction model, the problem of spatiotemporal misalignment and coordinate drift of trajectory dynamic data and defect feature data in traditional technology is solved, and the dual datasets are accurately associated and the trajectory coordinates are calibrated, providing a unified and reliable benchmark framework for subsequent data fusion.

[0019] (2) By setting up a collaborative processing logic with cross-modal feature purification and core defect feature consistency verification, the pain points of large single-modal noise interference and difficulty in removing false defect signals in traditional defect feature extraction are specifically addressed, thereby improving the purity and authenticity of defect features and ensuring that the subsequent fusion feature body can accurately reflect the essential attributes of the defect.

[0020] (3) By setting up a welding process-defect formation mechanism association network, and combining it with the positioning logic of weld three-dimensional dynamic contour and feature-coordinate binding-process tracing-multi-source closed-loop calibration based on trajectory data reconstruction, the limitation of traditional positioning that can only output defect location and cannot associate process causes is solved. The positioning accuracy is improved through dynamic benchmark correction and bidirectional verification, and finally the result carrying the process-defect causal confidence is output, providing support for welding quality control and process optimization with both accuracy and decision value. Attached Figure Description

[0021] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0022] Figure 1 This is a flowchart illustrating a defect localization method based on the fusion of weld defect features and trajectory tracking data according to the present invention. Figure 2 This is an architecture diagram of a defect localization method based on the fusion of weld defect features and trajectory tracking data according to the present invention. Detailed Implementation

[0023] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0024] Please see Figure 1-2A defect localization method based on the fusion of weld defect features and trajectory tracking data includes: S1: Capture dynamic trajectory data and defect feature data of the weld formation process; adjust the spatiotemporal resolution of trajectory data based on arc energy fluctuation characteristics, and enhance the multi-dimensional capture density of defect features through molten pool morphology changes; construct a dynamic coordinate system based on weld initiation feature points, and establish a benchmark mapping relationship between the two datasets through real-time coordinate system calibration. S2: Perform multi-physics field interference decoupling correction on the trajectory dynamic data, construct a coordinate drift correction model, and use the process parameter fluctuation law to back-verify and correct coordinate deviation; perform cross-modal feature purification on the defect feature data, and eliminate false defect signals through core defect feature consistency verification. S3: The preprocessed trajectory data and defect feature data are transformed into feature forms that meet the fusion requirements; a welding process-defect formation mechanism association network is constructed, and the scenario-based fusion weights of the features are adjusted to generate a fusion feature body containing the causal correlation between process and defect; S4: Reconstruct the three-dimensional dynamic contour of the weld based on the trajectory dynamic data, calculate the contour thermal deformation amount as the positioning reference correction parameter; call the adapted dynamic positioning logic based on the fused feature body, perform the corresponding positioning operation on the defect, and output the defect positioning result carrying the process-defect causal confidence.

[0025] As a preferred embodiment of the present invention, the method for adjusting the spatiotemporal resolution of the arc energy fluctuation characteristics trajectory data is as follows: The arc energy fluctuation coefficient is calculated by the real-time change rate of arc voltage and the effective value of welding current to quantify the degree of dynamic change of arc energy. Based on the fluctuation coefficient, a step-by-step adjustment rule is set. For coefficients below a preset threshold, a baseline spatiotemporal resolution is adopted, and for coefficients exceeding the threshold, the resolution is increased according to a gradient. The position of the weld fusion line is synchronously associated, and the resolution sensitivity is further increased in the area near the fusion line to perform dynamic resolution adjustment adapted to the arc energy fluctuation and the key areas of the weld.

[0026] Specifically, the method for adjusting the spatiotemporal resolution of the arc energy fluctuation characteristics trajectory data is as follows: By simultaneously acquiring the molten pool contour, temperature field distribution, and liquid metal flow trajectory through infrared imaging and visual sensing, quantitative parameters of the molten pool dynamic morphology are extracted, and a quantitative characterization model of molten pool morphology changes is constructed. Morphology-defect correlation rules are established to analyze the mapping relationship between different morphological change characteristics and potential defect types, and to determine the key dimensions and triggering conditions for defect feature capture. Dynamic capture density control is implemented, and based on the detected high-risk morphological change characteristics, the sampling frequency of ultrasonic echo signals and the layered scanning density of X-ray images are simultaneously increased. The capture dimension combination is adapted in combination with the characteristics of the weld formation stage to achieve precise matching between the multi-dimensional capture density of defect features and the molten pool morphology risk.

[0027] Specifically, the dynamic coordinate system is constructed from the weld initiation feature points, and the specific method is as follows: By combining visual recognition with laser contour scanning, the starting point of the weld fusion line or the starting position of the bevel root is located. Combined with the starting coordinate reference in the welding process document, the physical anchor point of the coordinate system origin is determined. Define the dynamic coordinate axis direction, with the welding travel direction as the X-axis, and update the axial direction in real time by tracking the movement trajectory of the welding torch; use the weld section normal direction as the Y-axis, and dynamically calibrate the workpiece surface geometric features; use the molten pool depth direction as the Z-axis, and correct the axial perpendicularity by combining the molten depth data from ultrasonic testing. Real-time coordinate system calibration is performed. By using high-frequency acquired trajectory dynamic data and weld formation feedback, the coordinate offset caused by thermal deformation is calculated, and the origin and axial parameters of the coordinate axes are periodically corrected.

[0028] Specifically, the baseline mapping relationship of the two datasets includes: Spatial reference mapping: The spatial coordinates of the defect feature dataset are mapped to the spatial framework of the trajectory dynamic dataset through the coordinate transformation rules of the dynamic coordinate system. This establishes a one-to-one correspondence between the spatial location of the defect features and the spatiotemporal coordinates of the trajectory data, and associates the spatial markers of the key areas of the weld. Time reference mapping: A time alignment reference is established based on welding process nodes, and the acquisition time of defect feature data is associated with the sampling time of trajectory dynamic data. The correspondence between the two in the welding time sequence is strengthened by process stage labels. Feature association mapping: Establish association rules between the process-spatiotemporal features of trajectory dynamic data and the morphological-attribute features of defect feature data, and map the process anomaly features of trajectory data with the risk features of defect data.

[0029] In this embodiment, the defect location scenario of butt welds in thick-walled steel structures is taken as the application object. Relying on the data acquisition system constructed by visual sensors, laser contour scanners, ultrasonic testing equipment, and welding control systems, the baseline mapping operation of the two datasets (dynamic dataset of weld trajectory and dataset of defect features) is specifically implemented. The steps are as follows: I. Implementation of Spatial Reference Mapping Defect feature spatial coordinate acquisition: The weld seam is scanned in real time during the welding process using ultrasonic testing equipment to collect the initial spatial coordinates (x0, y0, z0) of defects (such as internal lack of fusion, micro cracks), where x0 represents the position of the defect along the length of the workpiece, y0 represents the position of the defect along the width of the weld seam, and z0 represents the position of the defect along the depth of the weld seam. The initial coordinates are stored in the defect feature dataset.

[0030] Dynamic coordinate system transformation execution: The parameters of the dynamic coordinate system previously constructed based on the weld initiation feature point are called (the origin of the coordinate system is the weld fusion line initiation point, the X-axis is the welding direction, the Y-axis is the weld cross-section normal direction, and the Z-axis is the molten pool depth direction, and thermal deformation calibration has been completed), and the three-dimensional transformation matrix M of this coordinate system is obtained. The initial defect coordinates (x0, y0, z0) are substituted into the transformation formula (X, Y, Z) = M・(x0, y0, z0) to calculate the defect target coordinates (X1, Y1, Z1) within the trajectory dynamic dataset spatial framework.

[0031] One-to-one coordinate correspondence and key area marking: In the data visualization platform, load the spatiotemporal coordinate sequence of the trajectory dynamic dataset (such as the spatial position of the welding torch (X) recorded at fixed time intervals). t ,Y t Z t The system uses a coordinate matching algorithm to bind the target coordinates (X1, Y1, Z1) of the defect with the similar temporal coordinates (Xt1, Yt1, Zt1) in the trajectory coordinate sequence, achieving a one-to-one correspondence between the spatial location of the defect and the spatiotemporal coordinates of the trajectory. At the same time, the system calls the weld section data collected by the laser contour scanner in the platform to automatically identify the spatial range of the weld fusion line and the bevel transition area. The fusion line area is marked with a red dashed line and the bevel transition area is marked with a blue solid line. The defect coordinates (X1, Y1, Z1) are spatially associated with the marked areas.

[0032] II. Time Base Mapping Implementation Operation Welding process node setting: The process nodes of the butt weld of thick-walled steel structure are preset in the welding control system, which are divided into the root pass stage (from the welding start time T0 to the time T1 when the penetration depth reaches 1 / 3 of the workpiece thickness), the fill pass stage (from T1 to the time T2 when the penetration depth reaches 2 / 3 of the workpiece thickness), and the cover pass stage (from T2 to the welding end time T3). The time interval of each stage is stored in the process node database.

[0033] Time-series data acquisition and extraction: Extracting the acquisition time of each defect from the defect feature dataset, such as the timestamp t for the acquisition of a certain defect by an ultrasonic testing device. def Extract the timestamps of each sampling point from the dynamic trajectory dataset.

[0034] Time alignment and stage label binding: Using a timestamp matching tool, find the time sequence with the t tag. def The closest trajectory sampling timestamp t tra Defect acquisition time t def With trajectory sampling time t tra The association is used to achieve the time correspondence between "defect acquisition" and "trajectory sampling"; at the same time, the time t is determined. def If t is in the process node range def If ∈[T0,T1], then the defect data is labeled as "foundational stage". def If ∈[T1,T2], then label it "filling stage". def If the values ​​are in [T2,T3], then the "covering stage" label is applied to strengthen the correspondence between the two datasets in the welding time sequence.

[0035] III. Feature Association Mapping Implementation Operation Feature extraction and rule construction: Process-temporal features are extracted from the trajectory dynamic dataset, including welding torch speed fluctuation (Δv, representing the deviation between the welding torch speed and the reference speed in a certain spatiotemporal region) and welding current fluctuation amplitude (ΔI, representing the deviation between the current and the set current in a certain spatiotemporal region); morphological-attribute features are extracted from the defect feature dataset, including the linear length of the defect (L). def ), cross-sectional area (S) def ), ultrasonic echo signal intensity (I echo Based on sample data accumulated from previous welding experiments, association rules are established. "When Δv > Δv0 (Δv0 is the velocity fluctuation threshold symbol) and ΔI > ΔI0 (ΔI0 is the current fluctuation threshold symbol) in a certain spatiotemporal region of the trajectory dynamic data, and the region is in the filling stage, the corresponding defect feature data L..." def >L0 (L0 is the length threshold symbol), I echo The morphological-attribute features of ∈[I1,I2] (I1 and I2 are the symbols of the echo intensity interval) are determined as 'linear crack risk features', and the rule is entered into the feature association rule library.

[0036] Process Anomaly-Defect Risk Mapping Execution: During the welding process, the process-spatiotemporal characteristics of the trajectory dynamic data are monitored in real time. When it is detected that Δv=Δv1>Δv0 and ΔI=ΔI1>ΔI0 in a certain spatiotemporal region, and the stage label of the region is "filling stage", the system automatically calls the feature association rule library to match the corresponding "linear crack risk feature". At the same time, the system retrieves the defect data corresponding to the spatiotemporal region in the defect feature dataset. If a defect L is found... def =L1>L0、I echoIf I3∈[I1,I2], then the "welding gun speed-current fluctuation" process anomaly feature of the trajectory is bound to the "linear crack risk" feature of the defect, and the mapping result is presented in the data report in the form of "process anomaly: Δv1+ΔI1→defect risk: linear crack".

[0037] Through the above operations, a dual-dataset benchmark mapping result with "spatial coordinate correspondence, temporal sequence alignment, and feature risk binding" is finally formed, providing unified and related dataset support for subsequent data preprocessing and defect localization.

[0038] Specifically, the method for constructing the coordinate drift correction model is as follows: The vibration signal is spectrally decomposed using a modal separation algorithm to distinguish between the rigid displacement component and the flexible deformation component of the mechanical vibration; combined with the three-dimensional temperature field distribution acquired by an infrared thermal imager, the material temperature-deformation coefficient matrix is ​​introduced to calculate the coordinate offset caused by the arc thermal disturbance. A dynamic coupling correction model is constructed. Based on the time-varying coupling law of mechanical vibration and thermal disturbance during welding, a three-dimensional coordinate drift function containing a time decay factor is established to transform the decoupled disturbance components into correction parameters for the trajectory coordinates. Perform reverse verification of process parameters, extract transient fluctuation characteristics of welding current and voltage to construct interference-process correlation model, compare the model-predicted drift with the actual measurement residual, iteratively optimize the coefficient matrix of the coupled correction model, and dynamically adapt the coordinate correction accuracy to the process fluctuation state.

[0039] In this embodiment, the dynamic data calibration of the trajectory of the butt weld of thick-walled Q345 steel is taken as the application scenario. Strictly following the core logic of "multi-physics interference decoupling - dynamic coupling modeling - process reverse verification" in the invention, the construction of the coordinate drift correction model is implemented. The specific process is as follows: First, for the mechanical vibration signals collected during the welding process, a variational mode decomposition (VMD) algorithm was used to perform spectral decomposition. Referring to the "Technical Guide for Mode Decomposition of Mechanical Vibration Signals" JB / T 14035-2021, the core parameters of the VMD algorithm—number of modes K, penalty factor α, and noise tolerance β—were preset. The algorithm decomposed the original vibration signal into three types of modal components. The function of each component was distinguished based on its center frequency characteristics: the component with a center frequency of 5-10Hz was defined as the rigid displacement component ΔX. v (In the X direction, corresponding to the overall displacement of the welding torch and clamping mechanism due to transmission clearance and foundation vibration), the 15-25Hz component is defined as the flexible deformation component ΔX. f(In the X direction, corresponding to the micro-deformation caused by local thermal stress in the welding area of ​​the workpiece), components greater than 30Hz are judged as environmental noise and directly eliminated; when calculating the interference in the Y (weld section normal) and Z (molten pool depth) directions, the same parameters K, α, and β are used to obtain ΔY simultaneously. v ΔY f With ΔZ v ΔZ f To achieve decoupling and separation of vibration interference in all directions: Next, the coordinate shift caused by the thermal disturbance of the electric arc is calculated: First, the linear expansion coefficient α of Q345 steel in different temperature ranges is obtained by referring to the "Handbook of Thermal Expansion Coefficients of Metallic Materials" GB / T 4339-2008. t —The temperature range of 20~300℃ is α t1 The α range is 300~600℃. t2 The α range is 600~800℃. t3 The second step is to determine the basic parameters: Assume the ambient temperature T0 = 25℃ (meeting the standard for normal industrial environments), select the initial coordinates (x0, y0, z0) of any sampling point in the trajectory dynamic data, and obtain the real-time temperature T(x0, y0, z0, t) of the sampling point (which changes dynamically with the welding process); The third step is to substitute the values ​​into the formula to calculate the thermal disturbance offset, taking the X direction as an example: , In the formula, ΔX t Let α be the thermal disturbance offset in the X direction (in mm). If T(x0,y0,z0,t) = 450℃ (within the range of 300~600℃), then α is selected. t2 Substitute the values ​​into the calculation; using the same formula logic in the Y and Z directions, we obtain ΔY respectively. t =(T(x0,y0,z0,t)-T0)×α t2 ×y0、ΔZ t =(T(x0,y0,z0,t)-T0)×α t ×z0, and maintain α throughout. t The consistency of T0 values ​​ensures the continuity of thermal disturbance offset calculations; Based on the vibration disturbance component (ΔX) obtained from the previous decoupling v ΔX f etc.) and thermal disturbance offset (ΔX) t(etc.), combined with the interference change law of the welding process, a coupled model is constructed: it is clear that as the welding time t goes by, the influence of mechanical vibration gradually decreases and the influence of thermal disturbance gradually increases. Therefore, a time decay factor λ(t) is introduced (refer to the "Technical Specification for Dynamic Interference Control of Welding Process" JB / T 13866-2020). Let λ(t)=e^(-kt), where k is the decay coefficient - k1 in the initial stage of welding (high proportion of mechanical vibration), k2 in the middle stage (equilibrium of vibration and thermal disturbance), and k3 in the later stage (high proportion of thermal disturbance). The value of k is applied synchronously in the three directions of X, Y and Z throughout the process.

[0040] Taking the X-direction as an example, a three-dimensional coordinate drift function is constructed to transform the decoupled disturbance components into correction parameters: , In the formula, ΔX total (t) represents the total X-direction coordinate drift at time t (i.e., the X-direction correction parameter of the trajectory coordinates, in mm), where ΔX v ΔX f ΔX t All symbols are reused from previous calculations; as welding progresses into the later stages, λ(t) decreases, (1-λ(t)) increases, and the thermal disturbance offset ΔX... t The increased weighting in the total drift is consistent with actual disturbance patterns. Drift functions of the same form are constructed synchronously in the Y and Z directions: ΔY total (t)=λ(t)・(ΔY v +ΔY f )+(1-λ(t))・ΔY t ΔZ total (t)=λ(t)・(ΔZ v +ΔZ f )+(1-λ(t))・ΔZ t This ensures that the calculation logic for all-directional correction parameters is consistent.

[0041] The reverse verification of process parameters first extracts the transient fluctuation characteristics of welding process parameters: real-time welding current I(t) and voltage U(t) are obtained from the process parameters associated with the trajectory dynamic data. Preset process reference values—referring to the Q345 steel butt welding process standard JB / T 4709-2000, set current reference I0 and voltage reference U0, and use wavelet transform to calculate the fluctuation amplitude: ΔI=|I(t)-I0| (current fluctuation amount) and ΔU=|U(t)-U| (voltage fluctuation amount). In subsequent verification processes, the reference values ​​of I0 and U0 remain unchanged, and the process fluctuation state is reflected by the changes in ΔI and ΔU.

[0042] Next, a disturbance-process correlation model is constructed and iteratively optimized: with ΔI and ΔU as input variables, and the total drift ΔX in the X direction as the input variable. totalLet (t) be the output variable, and establish a multiple linear regression model: Specifically, the method for performing cross-modal feature purification is as follows: , In the formula, ΔX pred (t) represents the total drift in the X direction predicted by the model (using ΔX). total The physical meaning of (t), unit mm), a, b, c are model coefficients, the initial values ​​are obtained by fitting pre-experimental data. Then, the actual measurement residual ε(t) = |ΔX pred (t)-ΔX meas (t)|(ΔX meas (t) represents the actual drift of the trajectory coordinates), with a preset residual threshold ε0 = 5μm (refer to the accuracy standard of high-precision measurement equipment). If ε(t) > ε0, then a, b, and c are iteratively updated using the gradient descent algorithm until ε(t) ≤ ε0, ensuring the model's adaptability to highly fluctuating working conditions.

[0043] Finally, dynamic adaptation of correction accuracy and process fluctuations is achieved: the process level is divided according to the fluctuation range of ΔI—ΔI∈[I0,I1]A is low fluctuation, ΔI∈(I2,I3]A is medium fluctuation, and ΔI∈(I3,I4]A is high fluctuation. A dedicated coefficient matrix is ​​matched for different levels: low fluctuation corresponds to (a1, b1), medium fluctuation corresponds to (a2, b2), and high fluctuation corresponds to (a3, b3). The correlation model in the Y and Z directions also follows the coefficient matrix rules. The coefficients are automatically matched through the real-time values ​​of ΔI and ΔU to ensure that the coordinate correction accuracy in all directions is dynamically synchronized with the process status.

[0044] The coordinate drift correction model constructed through the above operations ultimately outputs ΔX. total (t), ΔY total (t), ΔZ total (t) can be directly used for coordinate calibration of trajectory dynamic data, ensuring that the deviation between the calibrated data and the actual weld position is controlled within the threshold of ε0=5μm, providing reliable support for the accuracy of the spatial reference for subsequent defect location.

[0045] Variational mode decomposition (VMD) algorithm is used to separate defect vibration signals from structural noise in ultrasonic echo signals. Intrinsic vibration modes determined solely by the physical properties of the defect are selected based on modal energy ratio. Multi-scale edge detection and morphological thinning are combined to extract the topological skeleton and edge connectivity features of the defect region from the X-ray images, while simultaneously suppressing pseudo-contours formed by welding spatter. The grayscale stretching range of the X-ray images is optimized based on the frequency characteristics of the ultrasonic vibration modes. At the same time, the spatial boundary of the X-ray topology is used to constrain the effective analysis range of the ultrasonic modes, forming a mutual reinforcement mechanism of the two modal features and weakening the detection blind zone of a single mode.

[0046] Specifically, the core defect feature consistency verification includes: A multi-dimensional matching framework is constructed, establishing the correlation verification dimensions of core defect features from three levels: spatial location, morphological features, and physical properties. The spatial verification features are based on the positional coincidence of the spatial verification features in the weld coordinate system; the morphological features are based on the consistency of the contour geometric features; and the physical properties are based on the intrinsic matching relationship between ultrasonic vibration characteristics and X-ray attenuation characteristics. Cross-modal cross-validation is performed, constraining the effective analysis boundary of the X-ray topology by the spatial distribution range of the ultrasonic vibration modes, and at the same time verifying the authenticity of the ultrasonic vibration signal source by the defect contour morphology of the X-ray image, forming a mutually corroborating verification logic. A pseudo-signal identification mechanism is established. For features that appear only in a single mode or have contradictory multi-dimensional matching, the mechanism is compared with a feature library of common interference sources in welding processes to identify them as pseudo-defect signals and remove them.

[0047] Specifically, the method for transforming the preprocessed trajectory data and defect feature data into a feature format suitable for fusion requirements is as follows: Based on the welding time sequence, continuous spatiotemporal segments are divided, the changing trend of trajectory coordinates of each segment is extracted, the welding process parameter fluctuation index of the corresponding segment is associated, and spatiotemporal information and process information are coupled through matrix operation to form a process-spatiotemporal coupling feature that characterizes the process state at a specific spatiotemporal location. Extract the core morphological parameters of the defects, associate them with the physical attribute data of the defects, and combine them with the feature weight rules of the defect type to construct morphological-attribute association features that are deeply bound to the morphology and attributes. The transformed features are uniformly mapped to a preset feature space. Through time alignment, the features are bound to welding process stage labels. Dimension normalization is performed synchronously. Through feature scaling, parameters of different physical dimensions are unified to the same numerical range.

[0048] Specifically, the welding process-defect formation mechanism correlation network includes: Using process-spatiotemporal coupling characteristics as the core input, it incorporates welding material parameters, groove structure parameters, and spatial coordinate information in a dynamic coordinate system to construct an input system covering process dynamics and spatial positioning. It has built-in welding thermodynamics and fluid dynamics models to establish a quantitative correlation between process characteristics and the spatial location and formation mechanism of defects, and to perform a deep mapping between process fluctuations and defect generation areas and their underlying mechanisms. Based on the dynamic adjustment of feature weights in the welding stage, it outputs the location results of defects in the dynamic coordinate system, risk level, and appropriate process adjustment suggestions.

[0049] Specifically, the three-dimensional dynamic profile of the weld includes: The spatiotemporal coordinate sequence and process parameter fluctuation characteristics of dynamic trajectory data are analyzed to construct a spatial position-process state mapping relationship, transforming discrete trajectory points into three-dimensional feature points containing forming influencing factors. Based on the spatial topological relationship of feature points and process association rules, three-dimensional contours of each stage are generated and formed into a continuous contour framework through temporal splicing. Combined with contour deviation feedback from real-time sensing, the detailed parameters of the framework are dynamically corrected to keep the reconstructed contour consistent with the actual forming state of the weld in spatiotemporal terms.

[0050] Specifically, the method for performing the corresponding location operation on the defect is as follows: A feature-coordinate dynamic binding mechanism is established. Using the dynamic coordinate system as the spatial reference, the cross-modal purified defect features are associated with the coordinate system spatial parameters, the spatial attributes of the defect features are extracted, and the feature data are mapped to the initial three-dimensional coordinates under the dynamic coordinate system through matrix transformation. The process traceability-assisted positioning correction is performed by calling the process-spatiotemporal coupling features in the welding process-defect formation mechanism association network, comparing the process fluctuation area corresponding to the initial coordinates of the defect, and inferring the possible extension boundary of the defect through the spatiotemporal range of the process anomaly, thus correcting the coordinate deviation caused by the sensor blind zone in the initial positioning, and forming a two-way verification of feature positioning and process traceability. Multi-source feedback closed-loop calibration is performed, and the weld seam three-dimensional dynamic contour is reconstructed in real time. The defect location coordinates are compared with the geometric features of the contour, and the accurate three-dimensional coordinates of the defect in the dynamic coordinate system and the location information of the corresponding welding process stage are output.

[0051] In this embodiment, the "internal lack of fusion" defect localization in butt welding of thick-walled Q345 steel is taken as the application scenario. Based on the previously constructed dynamic coordinate system (X-axis: welding direction, origin at the weld fusion line starting point; Y-axis: weld cross-section normal direction, bound to workpiece surface geometry calibration; Z-axis: molten pool depth direction, corrected by ultrasonic weld depth data), the defect features purified across modes (ultrasonic vibration modes + ray topology features), and the welding process-defect formation mechanism correlation network, the defect localization operation is implemented step by step. The specific process is as follows: First, determine the core characteristics of the defect after cross-modal purification: extract the inherent vibration mode parameters of the "internal unfused" defect—vibration frequency f—from the ultrasonic testing data. def Modal energy ratio η def (Characterizing the physical properties of defects); Extracting the topological features of defects—linear length L—from X-ray inspection data. def Cross-sectional width W def (Characterizing the morphology of defects), integrated into a defect feature set F def ={f def ,η def ,L def W def}

[0052] Using a dynamic coordinate system as a spatial reference, the spatial attributes of defect features are extracted: through feature association rules, f def With η def Mapping to the depth attribute of the defect relative to the fusion line (Z-axis correlation) – referring to the rule in the correlation network that "a specific frequency and energy ratio range corresponds to a specific depth range below the fusion line", the defect depth attribute Z is determined. attr ; will L def With W def Mapped to the positional attributes of the defect along the welding direction (X-axis) and the cross-section normal direction (Y-axis)—the X-axis distance from the defect initiation point to the origin of the coordinate system in the ray image is converted to X... attr The Y-axis offset of the defect center from the weld center is Y. attr (Because they are not fused, they are mostly close to the bevel sidewall).

[0053] Construct a coordinate transformation matrix M (based on real-time calibration parameters of the dynamic coordinate system, including thermal deformation compensation coefficients) to transform the spatial attributes {X} attr ,Y attr Z attr Substitute into the matrix transformation formula: , The initial three-dimensional coordinates (X0, Y0, Z0) of the defect in the dynamic coordinate system are calculated (matrix transformation corrects the small deviation caused by thermal deformation), thus completing the dynamic binding of the feature and the coordinates.

[0054] The welding process-defect formation mechanism association network is invoked to extract the process-spatiotemporal coupling features corresponding to the initial coordinates (X0,Y0,Z0) of the defect: According to the "mapping relationship between spatiotemporal coordinates and process parameters" in the network, the welding time period corresponding to X0 is located (determined to be the filling stage, because it is between the end of the rooting stage and the beginning of the capping stage). The process parameter fluctuation characteristics of the time period are: the current fluctuation ΔI exceeds the reference value threshold, and the welding gun speed fluctuation Δv exceeds the reference value threshold. That is, the process fluctuation area is the fluctuation of a specific time interval T and the fluctuation of the X-direction interval X.

[0055] By inferring the boundary through the correlation rules between process anomalies and defect extension: In the correlation network, "when ΔI and Δv exceed the threshold during the filling stage, the extension length of the unfused defect along the X-direction is more than the X-ray detection value by a certain proportion", combined with the initial coordinates X0, L def By reverse-engineering the actual X-axis starting and ending points of the defect, the actual X-axis range of the defect is determined (the initial coordinates only reflect the center of the defect and do not cover the extension boundary).

[0056] Correcting initial coordinate deviation: For the sensing blind zone (the ray has insufficient resolution in the depth direction, and the initial Z0 may have a slight deviation), combined with the melt depth data of the process fluctuation area - the average melt depth detected by ultrasonic testing during the T fluctuation period, the non-fusion defects are mostly located in a specific proportion range of the melt depth. The calculated theoretical depth matches the initial Z0 and no correction is needed; the X-axis is corrected to the actual center coordinate X1 of the defect (the original X0 deviates from the actual center), and the Y-axis does not need to be corrected because the position of the bevel sidewall is fixed. The corrected coordinates (X1, Y1, Z1) are obtained, forming a two-way verification of "feature positioning (initial coordinates) - process traceability (fluctuation back-inference)".

[0057] Retrieve the three-dimensional dynamic contour of the weld seam reconstructed based on trajectory dynamic data: the contour includes the cross-sectional geometric features of the time period corresponding to the filling stage - the groove width Y contour, the molten pool depth Z contour, and the X-axis contour coordinate synchronized with the welding travel direction.

[0058] The corrected coordinates (X1, Y1, Z1) are compared with the geometric features of the contour: Y1 is within the Y contour (consistent with the positional pattern of unfused near the bevel sidewall), Z1 is within the reasonable melt depth range of the Z contour, and X1 is within the filling stage range of the X contour; further comparison of the melt pool boundary features of the contour shows that the melt pool boundary at X1 deviates very little from Z1, thus determining that the positioning coordinates match the contour shape.

[0059] If deviations exist, iterative calibration is performed (in this case, there are no deviations and no adjustment is needed), ultimately determining the precise three-dimensional coordinates of the defect (X). f ,Y f Z f Output location information: Synchronously associate process stage labels (filling stage) and process anomaly causes (current and speed exceeding threshold fluctuations) to form a complete location result—"Defect type: Internal non-fusion; Precise coordinates: (X...)" f ,Y f Z f ); Corresponding process stage: filling stage; Process cause: current and speed exceeding threshold fluctuations; Location reliability: based on feature-process-contour triple verification", to complete the entire defect location operation.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A defect localization method based on the fusion of weld defect features and trajectory tracking data, characterized in that, include: S1: Captures dynamic trajectory data and defect feature data of the weld formation process; Based on the characteristics of arc energy fluctuation, the spatiotemporal resolution of trajectory data is adjusted, and the multidimensional capture density of defect features is enhanced by the characteristics of molten pool morphology changes. A dynamic coordinate system is constructed based on the starting feature points of the weld, and a benchmark mapping relationship between the two datasets is established through real-time calibration of the coordinate system. S2: Perform multi-physics field interference decoupling correction on the trajectory dynamic data, construct a coordinate drift correction model, and at the same time use the process parameter fluctuation law to back-verify and correct the coordinate deviation; Cross-modal feature purification is performed on the defect feature data, and false defect signals are eliminated through core defect feature consistency verification. S3: Convert the preprocessed trajectory data and defect feature data into feature forms that meet the fusion requirements; construct a welding process-defect formation mechanism association network, and adjust the scenario-based fusion weights of the features to generate a fusion feature body containing the causal correlation between process and defect; S4: Reconstruct the three-dimensional dynamic contour of the weld based on the trajectory dynamic data, calculate the contour thermal deformation amount as the positioning reference correction parameter; call the adapted dynamic positioning logic based on the fused feature body, perform the corresponding positioning operation on the defect, and output the defect positioning result carrying the process-defect causal confidence.

2. The method according to claim 1, characterized in that, The method for adjusting the spatiotemporal resolution of the arc energy fluctuation characteristics trajectory data is as follows: The arc energy fluctuation coefficient is calculated by the real-time change rate of arc voltage and the effective value of welding current to quantify the degree of dynamic change of arc energy; a step-by-step adjustment rule is set based on the fluctuation coefficient, and a reference spatiotemporal resolution is used for coefficients below a preset threshold, while the resolution is increased by gradient for coefficients exceeding the threshold. The location of the weld fusion line is synchronously correlated, and the resolution sensitivity is further enhanced in the area near the fusion line to perform dynamic resolution adjustment to adapt to arc energy fluctuations and key areas of the weld.

3. The method according to claim 1, characterized in that, The method for adjusting the spatiotemporal resolution of the arc energy fluctuation characteristics trajectory data is as follows: By simultaneously acquiring the molten pool contour, temperature field distribution, and liquid metal flow trajectory through infrared imaging and visual sensing, quantitative parameters of the molten pool dynamic morphology are extracted, and a quantitative characterization model of molten pool morphology changes is constructed. Morphology-defect correlation rules are established to analyze the mapping relationship between different morphological change characteristics and potential defect types, and to determine the key dimensions and triggering conditions for defect feature capture. Dynamic capture density control is implemented, and based on the detected high-risk morphological change characteristics, the sampling frequency of ultrasonic echo signals and the layered scanning density of X-ray images are simultaneously increased. The capture dimension combination is adapted in combination with the characteristics of the weld formation stage to match the multi-dimensional capture density of defect features with the risk of molten pool morphology.

4. The method according to claim 1, characterized in that, The dynamic coordinate system is constructed from the weld initiation feature points using the following method: By combining visual recognition with laser contour scanning, the starting point of the weld fusion line or the starting position of the bevel root is located. Combined with the starting coordinate reference in the welding process document, the physical anchor point of the coordinate system origin is determined. Define the direction of the dynamic coordinate axis, with the welding direction as the X-axis, and update the axial direction in real time by tracking the movement trajectory of the welding torch. Dynamic calibration of workpiece surface geometry is performed using the weld section normal direction as the Y-axis. Using the molten pool depth direction as the Z-axis, the axial perpendicularity is corrected by combining the molten depth data from ultrasonic testing; Real-time coordinate system calibration is performed. By using high-frequency acquired trajectory dynamic data and weld formation feedback, the coordinate offset caused by thermal deformation is calculated, and the origin and axial parameters of the coordinate axes are periodically corrected.

5. The method according to claim 1, characterized in that, The baseline mapping relationship between the two datasets includes: Spatial reference mapping: The spatial coordinates of the defect feature dataset are mapped to the spatial framework of the trajectory dynamic dataset through the coordinate transformation rules of the dynamic coordinate system. This establishes a one-to-one correspondence between the spatial location of the defect features and the spatiotemporal coordinates of the trajectory data, and associates the spatial markers of the key areas of the weld. Time reference mapping: A time alignment reference is established based on welding process nodes, and the acquisition time of defect feature data is associated with the sampling time of trajectory dynamic data. The correspondence between the two in the welding time sequence is strengthened by process stage labels. Feature association mapping: Establish association rules between the process-spatiotemporal features of trajectory dynamic data and the morphological-attribute features of defect feature data, and map the process anomaly features of trajectory data with the risk features of defect data.

6. The method according to claim 1, characterized in that, The specific method for constructing the coordinate drift correction model is as follows: The vibration signal is spectrally decomposed using a modal separation algorithm to distinguish between the rigid displacement component and the flexible deformation component of the mechanical vibration; combined with the three-dimensional temperature field distribution acquired by an infrared thermal imager, the material temperature-deformation coefficient matrix is ​​introduced to calculate the coordinate offset caused by the arc thermal disturbance. A dynamic coupling correction model is constructed. Based on the time-varying coupling law of mechanical vibration and thermal disturbance during welding, a three-dimensional coordinate drift function containing a time decay factor is established to transform the decoupled disturbance components into correction parameters for the trajectory coordinates. Perform reverse verification of process parameters, extract transient fluctuation characteristics of welding current and voltage to construct interference-process correlation model, compare the model-predicted drift with the actual measurement residual, iteratively optimize the coefficient matrix of the coupled correction model, and dynamically adapt the coordinate correction accuracy to the process fluctuation state.

7. The method according to claim 1, characterized in that, The specific method for performing cross-modal feature purification is as follows: Variational mode decomposition (VMD) algorithm is used to separate defect vibration signals from structural noise in ultrasonic echo signals. Intrinsic vibration modes determined solely by the physical properties of the defect are selected based on modal energy ratio. Multi-scale edge detection and morphological thinning are combined to extract the topological skeleton and edge connectivity features of the defect region from the X-ray images, while simultaneously suppressing pseudo-contours formed by welding spatter. The grayscale stretching range of the X-ray images is optimized based on the frequency characteristics of the ultrasonic vibration modes. At the same time, the spatial boundary of the X-ray topology is used to constrain the effective analysis range of the ultrasonic modes, forming a mutual reinforcement mechanism of the two modal features and weakening the detection blind zone of a single mode.

8. The method according to claim 1, characterized in that, The core defect feature consistency verification includes: A multi-dimensional matching framework is constructed, establishing the correlation verification dimensions of core defect features from three levels: spatial location, morphological features, and physical properties. The spatial verification features are based on the positional coincidence of the spatial verification features in the weld coordinate system; the morphological features are based on the consistency of the contour geometric features; and the physical properties are based on the intrinsic matching relationship between ultrasonic vibration characteristics and X-ray attenuation characteristics. Cross-modal cross-validation is performed, constraining the effective analysis boundary of the X-ray topology by the spatial distribution range of the ultrasonic vibration modes, and at the same time verifying the authenticity of the ultrasonic vibration signal source by the defect contour morphology of the X-ray image, forming a mutually corroborating verification logic. A pseudo-signal identification mechanism is established. For features that appear only in a single mode or have contradictory multi-dimensional matching, the mechanism is compared with a feature library of common interference sources in welding processes to identify them as pseudo-defect signals and remove them.

9. The method according to claim 1, characterized in that, The method for transforming the preprocessed trajectory data and defect feature data into a feature format suitable for fusion requirements is as follows: Based on the welding time sequence, continuous spatiotemporal segments are divided, the changing trend of trajectory coordinates of each segment is extracted, the welding process parameter fluctuation index of the corresponding segment is associated, and spatiotemporal information and process information are coupled through matrix operation to form a process-spatiotemporal coupling feature that characterizes the process state at a specific spatiotemporal location. Extract the core morphological parameters of the defects, associate them with the physical attribute data of the defects, and combine them with the feature weight rules of the defect type to construct morphological-attribute association features that are deeply bound to the morphology and attributes. The transformed features are uniformly mapped to a preset feature space. Through time alignment, the features are bound to welding process stage labels. Dimension normalization is performed synchronously. Through feature scaling, parameters of different physical dimensions are unified to the same numerical range.

10. The method according to claim 1, characterized in that, The welding process-defect formation mechanism correlation network includes: Using process-spatiotemporal coupling characteristics as the core input, it incorporates welding material parameters, groove structure parameters, and spatial coordinate information in a dynamic coordinate system to construct an input system covering process dynamics and spatial positioning. It has built-in welding thermodynamics and fluid dynamics models to establish a quantitative correlation between process characteristics and the spatial location and formation mechanism of defects, and to perform a deep mapping between process fluctuations and defect generation areas and their underlying mechanisms. Based on the dynamic adjustment of feature weights in the welding stage, it outputs the location results of defects in the dynamic coordinate system, risk level, and appropriate process adjustment suggestions.

11. The method according to claim 1, characterized in that, The three-dimensional dynamic profile of the weld includes: The spatiotemporal coordinate sequence and process parameter fluctuation characteristics of dynamic trajectory data are analyzed to construct a spatial position-process state mapping relationship, transforming discrete trajectory points into three-dimensional feature points containing forming influencing factors. Based on the spatial topological relationship of feature points and process association rules, three-dimensional contours of each stage are generated and formed into a continuous contour framework through temporal splicing. Combined with contour deviation feedback from real-time sensing, the detailed parameters of the framework are dynamically corrected to keep the reconstructed contour consistent with the actual forming state of the weld in spatiotemporal terms.

12. The method according to claim 1, characterized in that, The specific method for performing the corresponding location operation on the defect is as follows: A feature-coordinate dynamic binding mechanism is established. Using the dynamic coordinate system as the spatial reference, the cross-modal purified defect features are associated with the coordinate system spatial parameters, the spatial attributes of the defect features are extracted, and the feature data are mapped to the initial three-dimensional coordinates under the dynamic coordinate system through matrix transformation. The process traceability-assisted positioning correction is performed by calling the process-spatiotemporal coupling features in the welding process-defect formation mechanism association network, comparing the process fluctuation area corresponding to the initial coordinates of the defect, and inferring the possible extension boundary of the defect through the spatiotemporal range of the process anomaly, thus correcting the coordinate deviation caused by the sensor blind zone in the initial positioning, and forming a two-way verification of feature positioning and process traceability. Multi-source feedback closed-loop calibration is performed, and the weld seam three-dimensional dynamic contour is reconstructed in real time. The defect location coordinates are compared with the geometric features of the contour, and the accurate three-dimensional coordinates of the defect in the dynamic coordinate system and the location information of the corresponding welding process stage are output.

Images