An automated welding quality identification system and method based on visual perception

Abstract

The application relates to the technical field of automatic welding identification, in particular to an automatic welding quality identification system and method based on visual perception. A space-time molten pool reconstruction module is used to perform flow field alignment on a heterogeneous exposure image group, and a molten pool dynamic evolution graph is output; a structure tends to be stable constraint is used to deconstruct the evolution graph, and a steady-state topological layer and a transient interference layer are stripped; a spatial gain field which is gradiently attenuated towards a molten pool center is constructed, amplitude suppression is performed on the transient layer to generate a physical coupling vector; an orthogonal projection mechanism is used to take the amplitude deviation ratio as a weight to output a multi-dimensional correlation vector; and distance-direction collaborative discrimination logic is adopted to perform consistency determination and defect identification according to a coordinate mapping interval. The application solves the problem that smooth signals and random fluctuation signals are difficult to separate, and realizes accurate detection and quantitative identification of welding quality.

Description

Technical Field

[0001] This invention relates to the field of automated welding identification technology, specifically to an automated welding quality identification system and method based on visual perception. Background Technology

[0002] Automated welding equipment is a core unit in high-end manufacturing, and its welding quality directly determines the service life of products. Currently, welding quality control mainly adopts rule-based detection technology based on image thresholds, which involves acquiring weld images and calculating geometric features, triggering a defect alarm only when the values ​​exceed a preset static threshold.

[0003] However, the visual environment at the welding site exhibits strong time-varying coupling characteristics. Visual data contains both linear trend components based on normal welding trajectories and strong nonlinear random fluctuations caused by arc and spatter interference. Existing technologies based on single threshold judgment are essentially qualitative decisions and cannot decouple these two characteristics. When stable forming trends and transient noise interference intertwine, simple threshold logic either becomes overly sensitive to spatter noise, leading to false alarms, or lags in responding to subtle internal defects, resulting in missed detections. It is impossible to find a dynamic balance between sensitivity and stability.

[0004] While deep learning algorithms are capable of processing nonlinear data, a contradiction exists between model convergence and generalization ability in strong arc light scenarios. If undecomposed mixed image data is directly input into the network, the model struggles to capture both long-term linear trends and short-term sudden fluctuations, easily getting trapped in local optima and leading to decreased recognition accuracy. Furthermore, relying solely on the model's black-box output makes it difficult to construct an interpretable optimization space that conforms to the welding mechanism.

[0005] In summary, the technical problem that existing technologies need to solve is: how to address the difficulty in separating stable signals from random fluctuation signals in raw visual data, overcome the limitation of single threshold technology in quantitative optimization, and achieve accurate detection and quantitative identification of welding quality.

[0006] To address this, an automated welding quality identification system and method based on visual perception is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide an automated welding quality recognition system and method based on visual perception. It utilizes the total variational sparsity condition of the image to perform energy deconstruction and achieves quantitative recognition of quality through distance-direction collaborative discrimination logic, aiming to achieve accurate detection and quantitative recognition of welding quality under complex working conditions.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An automated welding quality identification system based on vision perception, comprising: The spatiotemporal molten pool reconstruction module acquires a group of molten pool images containing arc noise and different exposure characteristics, performs flow field alignment processing on the molten pool image group, and outputs a dynamic evolution map of the molten pool. The weld energy deconstruction module receives the dynamic evolution diagram of the molten pool, performs deconstruction processing on the dynamic evolution diagram of the molten pool using structural stabilization constraints, extracts the steady-state topology layer that characterizes the macroscopic physical boundary of the weld, and extracts the transient interference layer that characterizes spatter and arc light fluctuations. The molten pool oscillation modulation module extracts the edge contour of the steady-state topology layer, constructs a spatial gain field that decays with a gradient towards the center of the molten pool, and uses the spatial gain field to perform nonlinear amplitude suppression on the transient interference layer to generate a physical coupling vector. The working condition feature integration module integrates the steady-state topology layer features and physical coupling vectors through orthogonal projection, and uses the magnitude deviation ratio of the transient interference layer relative to the steady-state topology layer as the weight value to output a multi-dimensional correlation vector. The forming semantic determination module receives multi-dimensional association vectors, uses distance-direction collaborative discrimination logic to map the multi-dimensional association vectors to the defect semantic index library, and outputs the identification results including welding forming consistency and defect level.

[0009] Preferably, the specific steps of performing the flow field alignment processing include: during the welding operation, acquiring single-frame images with alternating long and short exposure times under non-uniform sampling intervals to form the molten pool image group; assigning pixel grayscale weights according to the exposure gradient of each single-frame image to synthesize a range image sequence; using a sub-pixel displacement compensation algorithm to calculate the pixel grayscale centroid displacement at corresponding feature points of adjacent frames in the range image sequence to obtain the sub-pixel offset; substituting the sub-pixel offset into the affine transformation matrix representing the geometric deformation in the image plane, performing coordinate remapping and resampling on the molten pool image group to eliminate the relative motion vectors caused by camera vibration and molten pool displacement, completing the flow field alignment processing, and outputting a spatiotemporally continuous dynamic evolution map of the molten pool.

[0010] Preferably, the specific steps of performing deconstruction processing based on structural stabilization constraints include: defining the structural stabilization constraints as the total variational sparsity condition of the image, and constructing an energy functional model containing a fidelity term and a gradient energy regularization term; calculating the gradient component of the fidelity term and the divergence component of the regularization term during the iterative stepping process; causing the energy functional model to converge by performing gradient descent updates in the original pixel domain and projection updates in the dual gradient domain, filtering out random noise in the dynamic evolution graph of the molten pool, extracting structural components with piecewise smoothing characteristics, and determining the structural components with piecewise smoothing characteristics as the steady-state topology layer characterizing the macroscopic physical boundary of the weld bead; performing pixel-by-pixel subtraction between the dynamic evolution graph of the molten pool and the steady-state topology layer to determine the residual pulse feature component as the transient interference layer characterizing spatter and arc fluctuations.

[0011] Preferably, the specific steps for constructing the spatial gain field include: performing edge detection processing on the steady-state topology layer to locate the outer boundary contour of the molten pool; performing Euclidean distance transformation on the outer boundary contour using a fast traversal algorithm to generate a distance field map of each pixel coordinate within the steady-state topology layer relative to the nearest boundary point; normalizing the distance field map using a nonlinear exponential mapping function to generate a weight matrix; and spatially mapping the weight matrix to the pixel coordinates to establish a spatial gain field in which the gain intensity decreases at a gradient from the outer boundary contour to the center of the molten pool.

[0012] Preferably, the specific steps for generating the physical coupling vector by performing nonlinear amplitude suppression include: using the spatial gain field as a spatial mask, performing element-wise multiplication with the transient interference layer, and combining the non-uniform distribution characteristics of the spatial gain field in the spatial domain to attenuate the noise signal intensity in the transient interference layer far from the center of the molten pool, thereby performing the nonlinear amplitude suppression; extracting the residual oscillation frequency and oscillation amplitude features after suppression by using a nonlinear activation operator; introducing curvature constraints on the edge contour of the steady-state topology layer, and recombining the extracted oscillation frequency and oscillation amplitude into the physical coupling vector characterizing the dynamic fluctuations of the molten pool surface deformation.

[0013] Preferably, the specific steps for outputting the multidimensional correlation vector include: extracting weld width fluctuation, centerline deviation, and contour continuity parameters from the steady-state topology layer to form a static geometric feature vector; extracting surface ripple energy distribution and oscillation frequency parameters from the physical coupling vector to form a dynamic physical feature vector; calculating the statistical feature variance of the transient interference layer within the local window, and performing a ratio operation between the feature variance and the global background brightness of the steady-state topology layer to determine the dynamically allocated weight value; performing decorrelation processing on the static geometric feature vector and the dynamic physical feature vector using an orthogonal basis decomposition algorithm, and performing linear superposition on each feature component after decorrelation based on the weight value to output the multidimensional correlation vector.

[0014] Preferably, the specific steps of mapping the multidimensional correlation vector to the defect semantic index using distance-direction co-discrimination logic include: pre-setting cluster center points representing different welding quality templates in the defect semantic index; calculating the Mahalanobis distance deviation of the multidimensional correlation vector relative to each cluster center point, and performing inverse covariance weighting on the vector dimension based on the Mahalanobis distance deviation to obtain a numerical deviation index; calculating the cosine of the direction angle between the multidimensional correlation vector and the cluster center point to obtain a direction offset index; determining the feature intensity deviation through the numerical deviation index, and determining the feature evolution trend deviation in combination with the direction offset index, and establishing distance-direction co-discrimination logic; executing the distance-direction co-discrimination logic to determine the coordinate mapping interval of the multidimensional correlation vector in the defect semantic index.

[0015] Preferably, the specific steps for outputting the result include: receiving a coordinate mapping interval, calculating the Euclidean distance dispersion of each sample point within the interval relative to the center point of the defect semantic index library; determining the welding formation consistency, which reflects the stability of the macroscopic morphology of the weld, based on the distribution stability of the Euclidean distance dispersion over time and the convergence determination result of a preset fluctuation threshold; determining the overlap state between the coordinate mapping interval and the preset defect mapping subspace in the defect semantic index library through spatial inclusion detection, and outputting the defect level based on the overlap ratio and the semantic label of the subspace.

[0016] An automated welding quality identification method based on visual perception includes: A set of molten pool images containing arc noise and different exposure characteristics is acquired. Flow field alignment processing is performed on the molten pool image set to output a dynamic evolution map of the molten pool. The dynamic evolution map of the molten pool is received, and deconstruction processing is performed on the dynamic evolution map of the molten pool using structural stabilization constraints. A steady-state topology layer representing the macroscopic physical boundary of the weld bead is extracted, and a transient interference layer representing spatter and arc fluctuations is extracted. The edge contour of the steady-state topology layer is extracted, and a spatial gain field with gradient decay towards the center of the molten pool is constructed. The spatial gain field is used to perform nonlinear amplitude suppression on the transient interference layer to generate a physical coupling vector. The steady-state topology layer features and the physical coupling vector are weighted and integrated through orthogonal projection, and the amplitude deviation ratio of the transient interference layer relative to the steady-state topology layer is used as the weight value to output a multidimensional correlation vector. The multidimensional correlation vector is received, and distance-direction collaborative discrimination logic is used to map the multidimensional correlation vector to a defect semantic index library to output the identification results of weld formation consistency and defect level.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention acquires single-frame images with alternating exposure times of varying lengths, performs flow field alignment processing, and constructs an energy functional model based on the total variational sparsity condition of the images. This model extracts a steady-state topological layer and a transient interference layer with piecewise smoothing characteristics from the molten pool image set. This technique achieves a substantial separation between macroscopic forming trends and microscopic random noise at the physical level. Compared to the black-box deep learning models mentioned in the background, this solution significantly improves the interpretability of feature extraction and resolves the technical contradiction that the strong coupling between linear trends and nonlinear fluctuations makes the model difficult to converge.

[0018] 2. This invention employs a fast traversal algorithm to perform Euclidean distance transformation on the external boundary contour to generate a distance field map. This establishes a spatial gain field with a gradient attenuation distribution from the external boundary contour towards the center of the molten pool. This gain field is then used as a spatial mask to perform element-wise multiplication with the transient interference layer, achieving nonlinear amplitude suppression. This scheme accurately fits the physical law of the welding interference field distribution with spatial distance, preserving the core characteristics of the molten pool while directionally attenuating the noise intensity far from the center. This processing logic overcomes the limitation of the previous technology's single threshold detection being overly sensitive to spatter interference, achieving accurate signal capture in interference field environments.

[0019] 3. This invention determines weight values ​​by calculating the ratio of the local window variance of the transient interference layer to the global background brightness. It then uses an orthogonal basis decomposition algorithm to decorrelate the static geometric feature vector and the dynamic physical feature vector, and outputs a multidimensional correlation vector based on the linear superposition of the weight values. This method eliminates logical redundancy between macroscopic morphology and microscopic oscillation indicators, ensuring the independence of each feature dimension. The system dynamically adjusts the sensing sensitivity using the weight values, and can adaptively optimize the feature contribution ratio according to the time-varying coupled environment mentioned in the background technology, significantly reducing the risk of false alarms and missed detections caused by environmental fluctuations.

[0020] 4. This invention establishes a distance-direction collaborative discrimination logic that includes numerical deviation and directional offset indices. It determines the overlap between the coordinate mapping interval and the preset defect mapping subspace through spatial inclusion detection. This logic integrates feature intensity deviation and evolution trend deviation, transforming qualitative decision-making into quantitative identification. The solution projects abstract high-dimensional data into clear defect levels and weld formation consistency scores, providing a standardized decision-making basis for automated welding monitoring and solving the technical bottleneck of existing technologies that cannot achieve a dynamic balance between detection sensitivity and stability. Attached Figure Description

[0021] Figure 1 The flowchart of the overall logical architecture of an automated welding quality recognition system based on vision perception is provided in Example 1. Figure 2 The flowchart illustrates the coupling of various modules in an automated welding quality identification system based on visual perception, as provided in Embodiment 1. Figure 3 A flowchart illustrating the gradient distribution logic and physical mapping relationship of the spatial gain field provided in Example 2; Figure 4 This is a graph showing the quality results of automated testing. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figures 1 to 2 This invention provides an automated welding quality identification system and method based on visual perception, the technical solution of which is as follows: An automated welding quality identification system based on vision perception, comprising: The spatiotemporal molten pool reconstruction module acquires a group of molten pool images containing arc noise and different exposure characteristics, performs flow field alignment processing on the molten pool image group, and outputs a dynamic evolution map of the molten pool. The weld energy deconstruction module receives the dynamic evolution diagram of the molten pool, performs deconstruction processing on the dynamic evolution diagram of the molten pool using structural stabilization constraints, extracts the steady-state topology layer that characterizes the macroscopic physical boundary of the weld, and extracts the transient interference layer that characterizes spatter and arc light fluctuations. The molten pool oscillation modulation module extracts the edge contour of the steady-state topology layer, constructs a spatial gain field that decays with a gradient towards the center of the molten pool, and uses the spatial gain field to perform nonlinear amplitude suppression on the transient interference layer to generate a physical coupling vector. The working condition feature integration module integrates the steady-state topology layer features and physical coupling vectors through orthogonal projection, and uses the magnitude deviation ratio of the transient interference layer relative to the steady-state topology layer as the weight value to output a multi-dimensional correlation vector. The forming semantic determination module receives multi-dimensional correlation vectors, uses distance-direction collaborative discrimination logic to map the multi-dimensional correlation vectors to the defect semantic index library, and outputs the identification results of welding forming consistency and defect level.

[0024] Example 1: This embodiment demonstrates the implementation method of the vision-based automated welding quality recognition system provided by the present invention in online quality inspection of automated gas shielded welding in industrial robots. (See also...) Figures 1 to 2 The system includes a spatiotemporal molten pool reconstruction module, a weld energy deconstruction module, a molten pool oscillation modulation module, a working condition feature integration module, and a forming semantic determination module. The specific steps are as follows: Further, the specific steps of performing the flow field alignment processing include: during the welding operation, acquiring single-frame images with alternating long and short exposure times under non-uniform sampling intervals to form the molten pool image group; assigning pixel grayscale weights according to the exposure gradient of each single-frame image to synthesize a range image sequence; using a sub-pixel displacement compensation algorithm to calculate the pixel grayscale centroid displacement at corresponding feature points of adjacent frames in the range image sequence to obtain the sub-pixel offset; substituting the sub-pixel offset into the affine transformation matrix representing the geometric deformation in the image plane, performing coordinate remapping and resampling on the molten pool image group to eliminate the relative motion vectors caused by camera vibration and molten pool displacement, completing the flow field alignment processing, and outputting a spatiotemporally continuous dynamic evolution map of the molten pool.

[0025] Specifically, the non-uniform sampling interval refers to the image acquisition device being triggered by the welding power pulse signal, dynamically adjusting the sampling frequency according to the peak value and base period of the welding current pulse to ensure that the acquisition time is synchronized with the physical instantaneous position of the molten droplet transition. The specific method for the alternating long and short exposure times is as follows: control three sensors to continuously perform three shutter opening operations with stepped spans at the same physical sampling point. Specifically, the first frame is set to a short exposure time of 15 microseconds to capture the liquid surface texture at the center of the high-brightness arc; the second frame is set to a medium exposure time of 150 microseconds to capture details of the molten pool transition area; and the third frame is set to a long exposure time of 1500 microseconds to capture the dark information of the background substrate and the bevel edge. The acquired images constitute the single-frame image and are summarized to form the molten pool image group. The exposure gradient refers to the layered distribution characteristics of local brightness and contrast in the image caused by the difference in exposure time. The pixel grayscale weight allocation refers to calculating the contrast, saturation, and exposure goodness index of each pixel, using a multi-resolution Laplacian pyramid decomposition structure, assigning higher weight coefficients to high-quality pixels in the middle grayscale range, and performing multi-scale fusion. The range image sequence is a linear data stream with a bit depth of 16 bits generated after fusion.

[0026] Specifically, the subpixel displacement compensation algorithm refers to using bicubic interpolation to increase the image spatial resolution to four times the original size, locating the image within a virtual high-density sampling grid. Adjacent frame images refer to two consecutive frames in a time-series composite image. Corresponding feature points refer to at least three non-collinear geometrically significant points extracted from each frame, such as molten pool edge corners or background texture points of the parent material. Pixel grayscale centroid displacement refers to the weighted average position displacement over time within a square window in the neighborhood of each feature point, using pixel grayscale values ​​as quality weights; the system determines the subpixel offset of each point by summarizing the displacement vector sets of all feature points between adjacent frames.

[0027] Specifically, the affine transformation matrix is ​​constructed as a second-order extended matrix containing translation coefficients, sine and cosine components of rotation angles, and shear coefficients. The system constructs a system of statically indeterminate equations from the extracted feature point coordinates and their displacement vectors, and solves this system using the least squares method to accurately calculate the translation, rotation, and shear parameters characterizing the overall geometric deformation within the image plane. The calculated parameter matrices are then used to perform coordinate remapping and resampling on the molten pool image set, eliminating the relative motion vectors caused by camera vibration and molten pool displacement, completing the flow field alignment process, and ensuring that the output molten pool dynamic evolution map maintains a high degree of geometrical center consistency on the time axis.

[0028] In a preferred embodiment, the specific steps of performing flow field alignment processing further include: introducing an imaging enhancement unit based on narrowband near-infrared spectroscopy. Specifically, while performing cyclic exposure operations, the image acquisition device is driven to acquire a near-infrared feature image of the molten pool through a narrowband filter located in the 850 nm to 950 nm wavelength band; the near-infrared feature image is aligned with the pixel coordinates of the range image sequence using a feature registration algorithm; and the molten pool contour with lower arc interference in the near-infrared image is extracted as guiding information. The system uses this guiding information to perform spatial weight correction on the feature point distribution in the image sequence, ensuring that the calculation of sub-pixel offsets avoids the arc saturation region. This embodiment reduces the interference of extremely bright arc light on feature extraction through physical spectral isolation, providing a more stable original reference benchmark for flow field alignment processing at the underlying hardware level.

[0029] This invention utilizes stepped exposure fusion and subpixel displacement compensation techniques to eliminate motion vector deviations caused by external vibrations and macroscopic displacement of the molten pool, ensuring high alignment of the molten pool dynamic evolution map in the spatiotemporal dimensions. Flow field alignment processing overcomes the challenge of limited imaging dynamic range caused by extremely bright arcs and extremely dark backgrounds during welding, providing a high-fidelity raw data foundation for subsequent precise quantitative analysis.

[0030] Further, the specific steps of performing deconstruction processing based on structural stabilization constraints include: defining the structural stabilization constraints as the total variational sparsity condition of the image, and constructing an energy functional model containing a fidelity term and a gradient energy regularization term; calculating the gradient component of the fidelity term and the divergence component of the regularization term during the iterative stepping process; causing the energy functional model to converge by performing gradient descent updates in the original pixel domain and projection updates in the dual gradient domain, filtering out random noise in the dynamic evolution graph of the molten pool, extracting structural components with piecewise smoothing characteristics, and determining the structural components with piecewise smoothing characteristics as the steady-state topology layer characterizing the macroscopic physical boundary of the weld bead; performing pixel-by-pixel subtraction between the dynamic evolution graph of the molten pool and the steady-state topology layer, and determining the residual pulse feature components as the transient interference layer characterizing spatter and arc fluctuations.

[0031] Specifically, the energy functional model is represented as a linear weighted sum of the squared L2 norm of the deviation between the image to be solved and the original evolution graph, and the L1 norm of the image gradient (i.e., the total variational term). The fidelity term constrains the numerical approximation in pixel brightness, while the regularization term induces spatial sparsity of the gradient, with the regularization parameter set between 0.05 and 0.2. This model, through a hierarchical mapping of mathematical criteria and optimization logic, constructs a multi-level coupled decoupled architecture to extract physically meaningful feature components from the complex dynamic evolution graph of the melt pool.

[0032] Specifically, in this architecture, the first layer is the criterion mapping layer, which receives the dynamic evolution graph of the molten pool as input, performs partial derivative operations with respect to the pixel to be solved using fidelity term units and outputs gradient components, and simultaneously performs gradient sparsity constraints using regularization term units and outputs divergence components. The second layer is the cross-domain dual optimization layer, which receives the gradient components and divergence components through a dual coupling mechanism, uses the divergence operator in the neighborhood of the pixel using dual variables, performs gradient descent updates in the original pixel domain, and simultaneously performs projection updates of the gradient vector with unit disk constraints in the dual gradient domain. A penalty factor of 1.5 is used between the two domains to perform step synchronization and dynamic weight exchange, achieving edge-preserving smoothing until the residual energy is below a preset threshold and a structural component with piecewise smoothing characteristics is output. The third layer is the residual feature stripping layer, which receives the structural components and performs pixel-to-pixel subtraction arithmetic coupling with the original evolution diagram. Through component difference logic, the output results are respectively determined as the steady-state topology layer characterizing the macroscopic physical boundary of the weld bead, and the transient interference layer carrying the instantaneous signals of spatter particles and arc light.

[0033] Specifically, the original pixel domain refers to the brightness representation space of image data in a two-dimensional coordinate system, and the dual gradient domain refers to the auxiliary space generated by transformation that carries the image gradient vector. The gradient descent update refers to updating the brightness value along the energy descent direction in the pixel domain, and the projection update refers to performing a unit disk constraint mapping operation on the gradient vector in the dual gradient domain. The model is considered converged when the residual energy of the functional model is lower than a preset threshold of 0.0001. The structural components with piecewise smoothness obtained by deconstruction exhibit a physical morphology with uniform gray-level distribution and clear boundaries, which can effectively extract the macroscopic physical boundary of the weld bead, and thus determine it as the steady-state topology layer. The system performs pixel-by-pixel subtraction, that is, subtracts the steady-state component from the evolution graph, and extracts the remaining signal components with high-frequency pulse characteristics as the transient interference layer, representing the gray-level response generated by the instantaneous flicker of spatter particles and arc light.

[0034] This invention utilizes image total variational sparsity constraints to perform deconstruction processing, achieving substantial separation between the steady-state topology layer and the transient interference layer. This processing method preserves the sharpness of the macroscopic physical boundaries of the weld bead while enabling the directional extraction of spatter particles and arc pulse signals, thus avoiding logical aliasing of feature signals at different scales from a physical level.

[0035] Furthermore, the specific steps for constructing the spatial gain field include: performing edge detection processing on the steady-state topology layer to locate the outer boundary contour of the molten pool; performing Euclidean distance transformation on the outer boundary contour using a fast traversal algorithm to generate a distance field map of each pixel coordinate within the steady-state topology layer relative to the nearest boundary point; normalizing the distance field map using a nonlinear exponential mapping function to generate a weight matrix; and spatially mapping the weight matrix to the pixel coordinates to establish a spatial gain field in which the gain intensity decreases at a gradient from the outer boundary contour to the center of the molten pool.

[0036] Specifically, the process of constructing the spatial gain field begins with edge detection processing of the steady-state topology layer. This processing calls an edge operator with gradient-sensitive characteristics to identify gray-level jump points in the image, thereby accurately locking the outer boundary contour of the region enclosing the molten metal pool. Subsequently, using this outer boundary contour as the starting wavefront, the system performs numerical solution of partial differential equations using the fast travel algorithm to simulate the physical process of the wavefront gradually advancing from the boundary to the center of the molten pool. During this process, the system performs the Euclidean distance transformation to calculate the minimum geometric distance from each pixel coordinate point in the steady-state topology layer to the nearest boundary point in the set of outer boundary contours. The result of this calculation is recorded in scalar matrix form, forming the distance field map, which quantifies the physical span of each pixel point in spatial distribution affected by edge arc source interference.

[0037] Specifically, after acquiring the distance field map, the system performs numerical transformation on the distance values ​​in the image using the nonlinear exponential mapping function. This function employs a monotonically decreasing exponential structure logic, mapping larger physical distance values ​​to smaller values, and, in conjunction with the normalization process, limits all results to the range of zero to one, thereby generating the weight matrix. Next, the system performs spatial mapping on the coefficients in the weight matrix and the corresponding pixel physical coordinates. The resulting spatial gain field exhibits a non-uniform distribution characteristic, with its weight coefficients decreasing centripetally along the outer boundary contour towards the geometric center of the molten pool. In this gain field, the edge regions adjacent to interference sources have higher weight assignments, while the central region of the molten pool has lower weight assignments, thus constructing a spatial mask corresponding to the spatial evolution law of the welding environment interference intensity. The specific implementation process is as follows: the system uses the distance field map generated by the fast travel algorithm as a numerical benchmark, and performs nonlinear exponential mapping to transform the Euclidean distance from the pixel to the boundary. Convert to corresponding gain weights The generated gain weight array maintains a high amplitude at the edge coordinates near the arc and spatter sources, and converges towards zero as the coordinates move closer to the geometric center of the molten pool, thus establishing a physical mapping matrix that performs spatially differentiated amplitude adjustments for the transient interference layer.

[0038] As a preferred implementation, the specific steps for constructing the spatial gain field further include: introducing a dynamic region of interest (ROI) locking unit based on spatial prior guidance. Specifically, the steps are: obtaining the centroid coordinates of the outer boundary contour and the direction of the weld bead extension axis; and constructing a dynamic ROI matrix in the steady-state topology layer based on the centroid coordinates and axis direction. The system limits the execution range of the fast traversal algorithm to the local space defined by this matrix, performing Euclidean distance transformation only on the pixel coordinates within the ROI. The generated distance field map retains the core physical features while eliminating redundant values ​​in non-welded areas of the background. The system uses this local distance distribution to generate a weight matrix and performs subsequent spatial gain field construction. This embodiment reduces the data processing load of the system under large-size welding conditions by reducing the computational operator space of the algorithm, achieving directional optimization of spatial gain construction.

[0039] This invention utilizes a fast-moving algorithm to construct a non-uniform spatial gain field with low center gain and high edge gain, achieving accurate fitting of the physical distribution characteristics of welding interference fields. By performing differentiated gain allocation on different spatial regions, the scheme establishes a signal enhancement mechanism with spatial prior constraints, improving the system's sensitivity to complex edge clutter.

[0040] Furthermore, the specific steps for generating the physical coupling vector by performing nonlinear amplitude suppression include: using the spatial gain field as a spatial mask, performing element-wise multiplication with the transient interference layer, and combining the non-uniform distribution characteristics of the spatial gain field in the spatial domain to attenuate the noise signal intensity in the transient interference layer far from the center of the molten pool, thereby performing the nonlinear amplitude suppression; extracting the residual oscillation frequency and oscillation amplitude features after suppression through a nonlinear activation operator; introducing the curvature constraint of the edge contour of the steady-state topology layer, and recombining the extracted oscillation frequency and oscillation amplitude into the physical coupling vector characterizing the dynamic fluctuations of the molten pool surface deformation.

[0041] Specifically, the process of performing nonlinear amplitude suppression begins by defining a pre-constructed spatial gain field as a spatial mask. The system drives this mask to perform element-wise multiplication with the transient interference layer output by the deconstruction module. That is, according to the principle of one-to-one correspondence of spatial coordinates, the weight coefficient of each position in the mask is multiplied with the gray value of the corresponding pixel in the interference layer. Since the spatial gain field is designed with a non-uniform distribution characteristic of large edge gain and small center gain, this operation mode realizes differentiated adjustment of signals at different spatial positions in the transient interference layer. In terms of physical distribution, the multiplication operation focuses on performing significant numerical compression on pixels far from the center of the melt pool, realizing directional attenuation of the noise signal intensity generated by arc light diffuse reflection in this region. This complete process of changing the amplitude distribution of the interference signal through the spatial weight matrix is ​​the nonlinear amplitude suppression.

[0042] Specifically, the system inputs the signal stream after the aforementioned suppression processing to a nonlinear activation operator. The nonlinear activation operator performs a hard-threshold-based nonlinear mapping to eliminate background perturbation signals with amplitudes below a preset safety threshold. The preset safety threshold is set to 10% to 15% of the global average brightness value of the current image sequence, dynamically suppressing low-energy background clutter to ensure the significance of residual oscillation signals. After activation processing, the system extracts the residual oscillation frequency and amplitude characteristics reflecting the true fluctuations of liquid metal. The residual oscillation frequency is defined as the periodic fluctuation rate of the target pixel brightness value on the sampling time axis; the oscillation amplitude characteristic is defined as the maximum physical deviation of the brightness value from the smoothing reference.

[0043] Specifically, the system reassembles the extracted residual oscillation frequency, oscillation amplitude, and curvature constraints of the steady-state topological layer edge contour into a physical coupling vector characterizing the dynamic fluctuations of the molten pool surface deformation, according to a specific data structure. This reassembly process involves constructing a three-dimensional feature vector: its first dimension is the residual oscillation frequency reflecting the rate of brightness fluctuation in the liquid metal; its second dimension is the oscillation amplitude reflecting the maximum deviation of brightness from the smoothing baseline; and its third dimension is the deformation intensity affected by the geometric shape, obtained by calculating the product of the normalized curvature value at the corresponding pixel position and the oscillation amplitude. In this three-dimensional feature vector, the third dimension uses curvature as a weight to perform a nonlinear mapping on the amplitude, quantifying the constraint effect of the molten pool edge geometric topology on surface physical fluctuations. This three-dimensional vector with a clear physical meaning constitutes the physical coupling vector, providing a standardized input baseline for subsequent orthogonal basis decomposition to eliminate logical redundancy between static geometric features and dynamic physical features.

[0044] This invention utilizes the synergistic effect of spatial mask dot product and nonlinear activation operators to attenuate the noise signal intensity in regions far from the center of the molten pool. By introducing edge contour curvature constraints, the generated physical coupling vector can characterize the intrinsic correlation between the deformation of the liquid metal surface and the geometric topological constraints, significantly enhancing the analytical depth of the system for the dynamic evolution of the molten pool surface.

[0045] Further, the specific steps for outputting the multidimensional correlation vector include: extracting weld width fluctuation, centerline deviation, and contour continuity parameters from the steady-state topology layer to form a static geometric feature vector; extracting surface ripple energy distribution and oscillation frequency parameters from the physical coupling vector to form a dynamic physical feature vector; calculating the statistical feature variance of the transient interference layer within the local window, and performing a ratio operation between the feature variance and the global background brightness of the steady-state topology layer to determine the dynamically allocated weight value; performing decorrelation processing on the static geometric feature vector and the dynamic physical feature vector using an orthogonal basis decomposition algorithm, and performing linear superposition on each feature component after decorrelation based on the weight value to output the multidimensional correlation vector.

[0046] Specifically, the process of outputting the multidimensional correlation vector begins with the parallel quantization extraction of both steady-state and dynamic parameters. The system first parses the weld width fluctuation, centerline deviation, and contour continuity parameters from the steady-state topology layer. The weld width fluctuation represents the statistical standard deviation of the maximum lateral span of the molten pool on the time axis; the centerline deviation represents the vertical displacement distance of the geometric center of the molten pool relative to the preset welding trajectory; and the contour continuity parameter represents the topological integrity score of the edge closure path. These parameters together constitute the static geometric feature vector. Simultaneously, the system parses the surface ripple energy distribution and oscillation frequency parameter from the physical coupling vector. The surface ripple energy distribution represents the concentration of the energy spectral density of the coupling vector in the spatial domain; and the oscillation frequency parameter represents the frequency value at which the surface brightness fluctuation of the molten pool is most significant in the frequency domain. These parameters together constitute the dynamic physical feature vector.

[0047] Specifically, the system selects a local window of fixed size on the transient interference layer and calculates the statistical feature variance of the pixel brightness within the window. This variance reflects the dispersion of splash particles or arc flicker. Subsequently, the system performs a ratio calculation between this variance and the global background brightness of the steady-state topology layer. The global background brightness is defined as the average grayscale value of all pixels in the steady-state image. The resulting value is defined as the dynamically assigned weight value, used to quantify the active proportion of transient noise relative to the steady-state structure. Next, the system calls the orthogonal basis decomposition algorithm to perform decorrelation processing on the static geometric feature vector and the dynamic physical features. This processing eliminates the linear correlation between features of different dimensions by calculating the projection of the static vector onto the dynamic vector space and removing the overlapping component. Finally, the system performs linear superposition on each processed feature component based on the dynamically assigned weight value, and the generated output product is the multidimensional correlation vector. Refer to Table 1 for a phased comparison of the preprocessing effects: Table 1: Inter-stage comparison of pretreatment effects

[0048] Table 1 quantifies the processing efficiency of the "weld energy deconstruction module" using objective indicators such as image entropy, signal-to-noise ratio, and edge sharpness. The data clearly demonstrates that after executing the TV total variational energy functional model, the system successfully separates the original high-noise signal into a "steady-state topology layer" with high signal-to-noise ratio and sharp boundaries, and a "transient interference layer" that carries random fluctuations. This substantial separation process proves that the solution can effectively overcome the impact of strong arc and spatter interference on the extraction of underlying features, as mentioned in the background technology.

[0049] This invention utilizes an orthogonal basis decomposition algorithm to perform decorrelation processing, eliminating logical redundancy between static geometric features and dynamic physical features, and ensuring the independence of feature dimensions. By dynamically adjusting weight coefficients based on disturbance activity, the multidimensional correlation vector achieves deep integration of macroscopic forming indicators and microscopic oscillation indicators, enhancing the system's robustness to complex disturbance conditions.

[0050] Further, the specific steps of mapping the multidimensional correlation vector to the defect semantic index using distance-direction co-discrimination logic include: pre-setting cluster center points representing different welding quality templates in the defect semantic index; calculating the Mahalanobis distance deviation of the multidimensional correlation vector relative to each cluster center point, and performing inverse covariance weighting on the vector dimension based on the Mahalanobis distance deviation to obtain a numerical deviation index; calculating the cosine of the direction angle between the multidimensional correlation vector and the cluster center point to obtain a direction offset index; determining the feature intensity deviation through the numerical deviation index, and determining the feature evolution trend deviation in combination with the direction offset index to establish distance-direction co-discrimination logic; executing the distance-direction co-discrimination logic to determine the coordinate mapping interval of the multidimensional correlation vector in the defect semantic index.

[0051] Specifically, the defect semantic index library is obtained by performing offline clustering training on historical welding datasets. Specifically, the historical dataset consists of no fewer than 5000 sets of dynamic weld pool feature samples collected simultaneously by a visual sensing system and a radiographic non-destructive testing device. When constructing the sample set, expert-level semantic annotation of the actual weld quality corresponding to the weld pool image is first performed according to industry standards (such as ISO5817 and corresponding domestic welding quality rating standards) to ensure that each set of multi-dimensional correlation vectors has a clear physical quality label. Subsequently, historical multi-dimensional correlation vectors for four typical working conditions—normal forming, undercut, porosity, and incomplete penetration—are extracted. To eliminate feature drift caused by differences in process parameters, the system performs normalization preprocessing on each dimension of features according to plate thickness and welding current before clustering. The K-means++ algorithm is used to perform initialization in the six-dimensional feature space. Specifically, a vector is randomly selected from the sample set as the first initial cluster center point; for subsequent center points, the shortest Euclidean distance between each sample vector and the current existing center point is calculated. and according to the direct proportion The probability distribution is used to select the remaining three center points sequentially, ensuring that the initial center can effectively cover the heterogeneous distribution of the four typical working conditions in the feature space. During the iteration process, the system assigns each sample vector to the cluster to which the nearest center point belongs based on the Euclidean distance. For each vector within a cluster, its arithmetic mean is calculated in a six-dimensional feature space that includes weld width fluctuation, centerline deviation, contour continuity, surface ripple energy distribution, oscillation frequency, and statistical feature variance, thereby dynamically updating the center point coordinates. When the update displacement of the cluster center is lower than a preset threshold... The iteration stops when the time is right, and the four centroid coordinates obtained by convergence are finally used as the preset cluster center points of the defect semantic index library.

[0052] Specifically, the mapping process using the distance-direction collaborative discrimination logic begins with a call to the defect semantic index. This index is constructed to store a standard feature space corresponding to different welding quality levels, calculated based on a large number of historical welding samples. The system presets multiple cluster centers within this space, each representing a specific defect type or normal forming state. Next, the system calculates the Mahalanobis distance deviation of the input multidimensional correlation vector relative to each cluster center. This calculation quantifies the scale of the current vector's deviation from the standard template by introducing the covariance matrix of the sample set. Based on this, the system performs inverse covariance weighting, using the inverse of the covariance matrix to correct the weights of each dimension of the multidimensional correlation vector. This correction operation eliminates coupling interference between different physical dimensions, and the final output is the numerical deviation index. This index is used at the physical level to determine the feature strength deviation of the current sample relative to the standard quality prototype.

[0053] Specifically, the process of executing the distance-direction collaborative discrimination logic and mapping the multidimensional correlation vector to the coordinate mapping interval refers to defining a two-dimensional polar coordinate mapping domain with clear physical meaning within the local feature space with the cluster center point as the origin, through the joint constraints of the numerical deviation index and the direction offset index. The numerical deviation index quantifies the radial distance of the current vector from the standard prototype based on Mahalanobis distance deviation, and is used to determine the distribution radius of the feature distribution; the direction offset index quantifies the angular displacement of the current vector relative to the cluster center reference vector based on the direction angle cosine, and is used to determine the azimuth angle of the feature distribution. The system defines the local geometric envelope region formed by continuous sampling points satisfying the above distribution radius constraint and azimuth angle constraint in high-dimensional space within a specific time series as the coordinate mapping interval. This interval is a two-dimensional projection domain characterizing the evolution trajectory and steady-state distribution of the current welding condition in the feature space. Through this coordinated locking of radial span and directional orientation, the system can transform abstract multidimensional correlation vectors into geometric entities with searchable and comparable characteristics in the defect semantic index, providing a closed-loop data foundation for subsequent consistency determination and defect level output.

[0054] This invention overcomes the limitations of traditional single distance metrics by utilizing the collaborative discrimination logic of Mahalanobis distance deviation and direction angle cosine. The scheme achieves simultaneous quantitative evaluation of feature intensity deviation and feature evolution trend, effectively distinguishing welding defects of different patterns and improving the logical rigor of semantic space mapping.

[0055] Reference Figure 4 The automated inspection quality result diagram further includes the following specific steps for outputting the result: receiving a coordinate mapping interval, calculating the Euclidean distance dispersion of each sample point within the interval relative to the center point of the defect semantic index library; determining the welding formation consistency, which reflects the stability of the macroscopic morphology of the weld, based on the distribution stability of the Euclidean distance dispersion over time and the convergence determination result of the preset fluctuation threshold; determining the overlap state between the coordinate mapping interval and the preset defect mapping subspace in the defect semantic index library through spatial inclusion detection, and outputting the defect level according to the overlap ratio and the semantic label of the subspace.

[0056] Specifically, the process of outputting the recognition result begins with a deep analysis of the coordinate mapping interval. The system receives the coordinate mapping interval projected by the forming semantic judgment module, which represents a geometric region enclosed by multiple sampling points within a high-dimensional feature space. The system calculates the Euclidean distance dispersion of each sampling point within the coordinate mapping interval relative to the center point of the defect semantic index library. This dispersion value reflects the degree of concentration of the current feature distribution relative to the standard quality prototype. Subsequently, the system introduces the time series dimension, statistically analyzes the variation pattern of the dispersion value within multiple consecutive sampling periods, and quantifies the distribution stability of the feature distribution on the time axis. Next, the system performs a comparison logic between the currently acquired distribution stability index and the preset fluctuation threshold to obtain the convergence judgment result. If the judgment result is within a stable interval within the threshold range, the system determines the macroscopic morphology stability of the weld bead, reflecting the stability of the weld formation state, and the weld formation consistency score accordingly.

[0057] Specifically, while determining the consistency index, the system simultaneously performs spatial inclusion detection on the coordinate mapping interval. This detection process performs a topological intersection operation between the geometric region covered by the current feature point set and the preset defect mapping subspace stored in the defect semantic index. The preset defect mapping subspace refers to a feature distribution region predefined in the index, corresponding to the severity of a specific welding defect. The preset defect mapping subspace is defined as a hypersphere region with the cluster center as the center and a radius of twice the standard deviation of the Mahalanobis distance of the corresponding defect category as the radius. Spatial inclusion detection determines the overlap ratio by calculating the proportion of sample points within the coordinate mapping interval that fall into this hypersphere region. The system determines the overlap state of the two through this operation and quantifies the overlap ratio of the overlapping part to the total volume of the coordinate mapping interval. Based on the magnitude of this ratio, the system retrieves and matches the preset defect mapping subspace with the highest degree of overlap and obtains the semantic label of the corresponding subspace. This label records the physical category information of the defect, and the system finally determines and outputs the recognition result, including the defect level, based on the label content.

[0058] Specifically, the system calculates the overlap ratio based on spatial inclusion detection. Execution level mapping: when When it is determined to be a Level 1 defect, a "Minor Defect" label is output; when When it is determined to be a level 2 defect, a "moderate defect" label is output; when If the overlap ratio is less than 0.6 but its coordinate mapping range shows a continuous deviation trend in the time series, a warning signal is triggered.

[0059] In a preferred embodiment, the output results further include introducing a process parameter control closed-loop unit based on quality score feedback. Specifically, this involves receiving the weld formation consistency score and the defect level identification result; and inputting these results as input components into a preset welding parameter correction model. This model, based on the defect type and severity, sends current adjustment commands or travel speed compensation signals to the welding power source via a fieldbus. The system monitors the change trajectory of the output results in real time throughout the entire welding operation lifecycle and performs dynamic fine-tuning based on real-time feedback of the Euclidean distance dispersion. This embodiment establishes a control closed loop from visual perception to physical execution by converting the identification conclusions into real-time execution commands, achieving online adaptive adjustment of the welding process.

[0060] Table 2: Comparison of Defect Identification Accuracy

[0061] Table 2, through comparative experimental results, directly demonstrates the significant advantages of this invention in terms of recognition accuracy and false alarm rate reduction. Compared to traditional rule-based detection techniques based on static thresholds, this system utilizes "operating condition feature integration" and "semantic judgment" logic to achieve accurate capture of different welding defects in complex time-varying coupled environments. The substantial improvement in the recognition rate of each typical defect strongly supports the technical conclusion in the specification that this solution can achieve a "dynamic balance between sensitivity and stability."

[0062] Table 3: Semantic Mapping Decision Parameter Table

[0063] Table 3 details how the system transforms abstract high-dimensional multidimensional correlation vectors into physically meaningful quality grade conclusions. By setting key thresholds such as overlap ratio and Mahalanobis distance standard deviation multiple, the system establishes a scientific quantitative grading standard and can provide corresponding process parameter correction actions based on the identified defect level. This quantitative identification process demonstrates the complete closed loop of this invention from low-level visual perception to high-level standardized decision-making, solving the problem that existing technologies cannot achieve quantitative optimization.

[0064] This invention utilizes Euclidean distance dispersion analysis and spatial inclusion detection technology to achieve quantitative evaluation of weld formation consistency and automated identification of defect levels. By performing overlap determination between the coordinate mapping interval and the preset defect mapping subspace, the scheme transforms the abstract high-dimensional feature distribution into standardized quality identification conclusions, providing an objective decision-making basis for real-time quality monitoring in industrial settings.

[0065] Example 2: This embodiment uses automated gas shielded welding with industrial robots as a specific application scenario to elaborate on the automated welding quality identification method based on vision perception provided by the present invention. (See reference...) Figure 3 The specific process is as follows: First, a visual data acquisition and flow field alignment process is executed. During robot operation, a sensor with range imaging capabilities continuously observes the molten pool area. Specifically, during acquisition, the sensor is triggered by the welding power pulse signal and performs three stepped exposures at the same physical location: the first frame uses a 15-microsecond exposure time to capture the bright center details of the molten pool, the second frame uses a 150-microsecond exposure time to capture the deformed edges of the molten pool, and the third frame uses a 1500-microsecond exposure time to capture the texture of the background substrate. After acquiring this group of molten pool images, pixel grayscale weight allocation is performed based on the exposure gradient of each single frame image, and the data is synthesized into a wide dynamic range image sequence with a bit depth of 16 bits. Subsequently, a sub-pixel displacement compensation program is called to extract corresponding feature points between adjacent frames and calculate the grayscale centroid displacement, obtaining a sub-pixel offset with a precision of 0.1 pixels. This offset is assigned to an affine transformation matrix, which is used to perform coordinate remapping and resampling on the image sequence. This process eliminates the relative motion vectors generated by the high-frequency vibration of the camera and outputs a spatiotemporally continuous dynamic evolution map of the molten pool.

[0066] Subsequently, energy deconstruction processing based on structural stabilization constraints is performed on the dynamic evolution map of the molten pool. Using the total variational sparsity condition of the image as a physical criterion, an energy functional model containing a fidelity term and a gradient energy regularization term is constructed. Specifically, the system employs the alternating direction multiplier method for numerical solution, setting a penalty factor of 1.5, and performing alternating step updates between the original pixel domain and the dual gradient domain. By calculating the gradient component of the fidelity term and the divergence component of the regularization term, the energy functional model is made to converge when the residual energy is below 0.0001. This process deconstructs the image into a steady-state topology layer with piecewise smoothing characteristics and a transient interference layer carrying high-frequency pulse characteristics. The steady-state topology layer represents the macroscopic physical morphology of the weld bead, while the transient interference layer centrally characterizes the random response generated by spatter particles and arc flicker.

[0067] Next, the spatial modulation and physical coupling vector generation process begins. First, the outer boundary contour of the steady-state topology layer is extracted using a gradient operator. This contour is then used as the starting wavefront, and a fast marching algorithm is invoked to numerically solve the wave equation. This algorithm calculates the physical distance from each pixel coordinate point in the image to the nearest boundary point by performing an Euclidean distance transformation, generating a distance field map. Subsequently, a nonlinear exponential mapping function is used to normalize this distance field, establishing a spatial gain field with low gain in the central region and high gain in the edge region. This field is then used as a spatial mask and element-wise multiplied with the transient interference layer, performing nonlinear amplitude suppression to attenuate the noise signal intensity far from the center of the molten pool. After truncation by a nonlinear activation operator, the residual oscillation frequency and amplitude features are extracted and recombined with the curvature constraints of the edge contour to generate a physical coupling vector characterizing the dynamic fluctuations of the molten pool surface deformation.

[0068] Subsequently, the feature integration and multidimensional correlation vector output process is executed. Weld width fluctuation, centerline deviation, and contour continuity parameters are extracted from the steady-state topology layer and encapsulated into a static geometric feature vector. Simultaneously, surface ripple energy distribution and oscillation frequency parameters are extracted from the physical coupling vector and encapsulated into a dynamic physical feature vector. Meanwhile, the statistical feature variance of the transient interference layer is calculated within a 16×16 pixel local window, and the ratio of this variance to the global background brightness is calculated to determine the dynamically assigned weight value. Then, the orthogonal basis decomposition algorithm is called to perform decorrelation processing on the above two sets of feature vectors, eliminating overlapping information between dimensions through Schmitt orthogonalization logic. Finally, based on this weight value, linear superposition is performed on each decorrelated feature component to output a multidimensional correlation vector, achieving deep fusion of macroscopic morphology indicators and microscopic oscillation indicators.

[0069] Finally, a quality assessment process based on semantic space mapping is executed. The defect semantic index is invoked, and the Mahalanobis distance deviation of the multidimensional association vector relative to the preset cluster centers in the index is calculated. The inverse covariance matrix is ​​then used to adjust the weights of the feature dimensions, yielding a numerical deviation index. Simultaneously, the cosine of the direction angle between vectors is calculated to obtain a direction offset index. Through distance-direction collaborative discrimination logic, the current feature is projected onto a coordinate mapping interval within the index. Upon receiving this interval, the Euclidean distance dispersion of the sample points relative to the origin is calculated, and convergence is determined by combining the distribution stability over time, quantifying the weld formation consistency. Finally, spatial inclusion detection is performed, calculating the overlap ratio between the coordinate mapping interval and the preset defect mapping subspace. When the overlap ratio is greater than 0.6, the corresponding defect level identification conclusion is output based on the semantic label of the subspace, completing the online identification loop.

[0070] This invention solves the technical challenges of inaccurate molten pool feature extraction and limited quantization dimensions under strong arc light interference by constructing a closed-loop system encompassing spatiotemporal reconstruction, energy decoupling, oscillation modulation, feature integration, and semantic determination. The solution achieves automated mapping of welding quality from underlying physical image features to high-level quality semantic labels, significantly improving the stability and accuracy of automated welding recognition under complex and fluctuating conditions.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automated welding quality identification system based on visual perception, characterized in that, include: The spatiotemporal molten pool reconstruction module acquires a group of molten pool images containing arc noise and different exposure characteristics, performs flow field alignment processing on the molten pool image group, and outputs a dynamic evolution map of the molten pool. The weld energy deconstruction module receives the dynamic evolution diagram of the molten pool, performs deconstruction processing on the dynamic evolution diagram of the molten pool using structural stabilization constraints, extracts the steady-state topology layer that characterizes the macroscopic physical boundary of the weld, and extracts the transient interference layer that characterizes spatter and arc light fluctuations. The molten pool oscillation modulation module extracts the edge contour of the steady-state topology layer, constructs a spatial gain field that decays with a gradient towards the center of the molten pool, and uses the spatial gain field to perform nonlinear amplitude suppression on the transient interference layer to generate a physical coupling vector. The working condition feature integration module integrates the steady-state topology layer features and physical coupling vectors through orthogonal projection, and uses the magnitude deviation ratio of the transient interference layer relative to the steady-state topology layer as the weight value to output a multi-dimensional correlation vector. The forming semantic determination module receives multi-dimensional association vectors, uses distance-direction collaborative discrimination logic to map the multi-dimensional association vectors to the defect semantic index library, and outputs the identification results including welding forming consistency and defect level.

2. The automated welding quality identification system based on visual perception according to claim 1, characterized in that, The specific steps for performing the flow field alignment process include: During the welding operation, single-frame images with alternating long and short exposure times under non-uniform sampling intervals are acquired to form the molten pool image group. Pixel grayscale weights are allocated according to the exposure gradient of each single-frame image to synthesize a range image sequence. A sub-pixel displacement compensation algorithm is used to calculate the pixel grayscale centroid displacement at corresponding feature points of adjacent frames in the range image sequence, obtaining the sub-pixel offset. The sub-pixel offset is substituted into the affine transformation matrix representing the geometric deformation in the image plane, and coordinate remapping and resampling are performed on the molten pool image group to eliminate the relative motion vectors caused by camera vibration and molten pool displacement, completing the flow field alignment processing and outputting a spatiotemporally continuous dynamic evolution map of the molten pool.

3. The automated welding quality identification system based on visual perception according to claim 1, characterized in that, The specific steps for performing deconstruction processing based on structural stabilization constraints include: The structural stabilization constraint is defined as the total variational sparsity condition of the image, and an energy functional model containing a fidelity term and a gradient energy regularization term is constructed. During the iterative stepping process, the gradient component of the fidelity term and the divergence component of the gradient energy regularization term are calculated. By performing gradient descent updates in the original pixel domain and projection updates in the dual gradient domain, the energy functional model is made to converge, and random noise in the dynamic evolution graph of the molten pool is filtered out. The structural component with piecewise smoothing characteristics is extracted and determined as the steady-state topology layer representing the macroscopic physical boundary of the weld bead. The dynamic evolution graph of the molten pool and the steady-state topology layer are subjected to pixel-by-pixel subtraction to determine the residual pulse feature component as the transient interference layer representing spatter and arc fluctuations.

4. The automated welding quality identification system based on visual perception according to claim 1, characterized in that, The specific steps for constructing the spatial gain field include: Edge detection processing is performed on the steady-state topology layer to locate the outer boundary contour of the molten pool; a fast traversal algorithm is used to perform Euclidean distance transformation on the outer boundary contour to generate a distance field map of each pixel coordinate within the steady-state topology layer relative to the nearest boundary point; the distance field map is normalized using a nonlinear exponential mapping function to generate a weight matrix; the weight matrix is ​​spatially mapped to the pixel coordinates to establish a spatial gain field with a gradient decay distribution of gain intensity from the outer boundary contour to the center of the molten pool.

5. The automated welding quality identification system based on visual perception according to claim 1, characterized in that, The specific steps for performing nonlinear amplitude suppression to generate the physical coupling vector include: Using the spatial gain field as a spatial mask, element-wise multiplication is performed with the transient interference layer. Combined with the non-uniform distribution characteristics of the spatial gain field in the spatial domain, the noise signal intensity in the transient interference layer, far from the center of the molten pool, is attenuated, and nonlinear amplitude suppression is performed. The residual oscillation frequency and amplitude features after suppression are extracted using a nonlinear activation operator. Curvature constraints on the edge contour of the steady-state topology layer are introduced, and the extracted oscillation frequency and amplitude are recombined into the physical coupling vector characterizing the dynamic fluctuations of the molten pool surface deformation.

6. The automated welding quality identification system based on visual perception according to claim 1, characterized in that, The specific steps for outputting the multidimensional correlation vector include: Weld width fluctuation, centerline deviation, and contour continuity parameters are extracted from the steady-state topology layer to form a static geometric feature vector; surface ripple energy distribution and oscillation frequency parameters are extracted from the physical coupling vector to form a dynamic physical feature vector; the statistical feature variance of the transient interference layer within the local window is calculated, and the ratio of the statistical feature variance to the global background brightness of the steady-state topology layer is used to determine the dynamically allocated weight value; the static geometric feature vector and the dynamic physical feature vector are decorrelated using an orthogonal basis decomposition algorithm, and the decorrelated feature components are linearly superimposed based on the weight value to output the multidimensional correlation vector.

7. The automated welding quality identification system based on visual perception according to claim 1, characterized in that, The specific steps for mapping multidimensional association vectors to the defect semantic index using distance-direction collaborative discrimination logic include: In the defect semantic index, cluster centers representing different welding quality templates are preset; the Mahalanobis distance deviation of the multidimensional correlation vector relative to each cluster center is calculated, and the vector dimension is weighted by inverse covariance based on the Mahalanobis distance deviation to obtain a numerical deviation index; the cosine of the direction angle between the multidimensional correlation vector and the cluster center is calculated to obtain a direction offset index; the feature intensity deviation is determined by the numerical deviation index, and the feature evolution trend deviation is determined by combining the direction offset index to establish a distance-direction co-discrimination logic; the distance-direction co-discrimination logic is executed to determine the coordinate mapping interval of the multidimensional correlation vector in the defect semantic index.

8. The automated welding quality identification system based on visual perception according to claim 1, characterized in that, The specific steps for outputting the results include: Receive the coordinate mapping interval, calculate the Euclidean distance dispersion of each sample point within the interval relative to the center point of the defect semantic index library; based on the distribution stability of the Euclidean distance dispersion over time, and combined with the convergence determination result of the preset fluctuation threshold, determine the welding formation consistency that reflects the macroscopic morphology stability of the weld bead; through spatial inclusion detection, determine the overlap state between the coordinate mapping interval and the preset defect mapping subspace in the defect semantic index library, and output the defect level according to the overlap ratio and the semantic label of the subspace.

9. An automated welding quality identification method based on visual perception, characterized in that, include: A set of molten pool images containing arc noise and different exposure characteristics is acquired. Flow field alignment processing is performed on the molten pool image set to output a dynamic evolution map of the molten pool. The dynamic evolution map of the molten pool is received, and deconstruction processing is performed on the dynamic evolution map of the molten pool using structural stabilization constraints. A steady-state topology layer representing the macroscopic physical boundary of the weld bead is extracted, and a transient interference layer representing spatter and arc fluctuations is extracted. The edge contour of the steady-state topology layer is extracted, and a spatial gain field with gradient decay towards the center of the molten pool is constructed. The spatial gain field is used to perform nonlinear amplitude suppression on the transient interference layer to generate a physical coupling vector. The steady-state topology layer features and the physical coupling vector are weighted and integrated through orthogonal projection, and the amplitude deviation ratio of the transient interference layer relative to the steady-state topology layer is used as the weight value to output a multidimensional correlation vector. The multidimensional correlation vector is received, and distance-direction collaborative discrimination logic is used to map the multidimensional correlation vector to a defect semantic index library to output the recognition results containing weld formation consistency and defect level.

Images