An intelligent detection method for metal welding processing based on industrial vision

Abstract

This invention relates to the field of metal welding inspection technology, specifically to an intelligent inspection method for metal welding processing based on industrial vision. The method includes: constructing a multimodal welding process tensor characterizing the coupling relationship between welding energy, material state, and metallurgical behavior; inputting this tensor into a reasoning computation network based on physical information constraints to forward deduce the evolution of the molten pool morphology and temperature gradient field; simultaneously executing a state inversion algorithm using multispectral fused images as observation data and welding process parameters as boundary conditions to optimize the intrinsic parameters of the welding heat source model; and fusing the forward deduction and inverse optimization results to construct a dynamic digital twin of the welding process. The method calculates the solidification rate, thermal cycle curve, and latent heat of phase transformation at each point on the welding line in real time, compares each point with the standard values ​​of the qualified process window, identifies abnormal locations, and outputs a structured inspection report containing coordinates and deviation types. This invention achieves coupled characterization and dynamic modeling of the welding process, improving the accuracy and real-time performance of the inspection.

Description

Technical Field

[0001] This invention relates to the field of metal welding inspection technology, and in particular to an intelligent inspection method for metal welding processing based on industrial vision. Background Technology

[0002] Current metal welding inspection relies heavily on industrial vision to acquire single-modal or multispectral images. Traditional image processing or pure data-driven neural networks are used to extract molten pool features and identify defects. Some welding process models use single physical simulation software to solve heat conduction and fluid dynamics equations, or rely on process parameters to perform static calibration of heat source models. The inspection process focuses on the morphological identification of defects after welding, and the integration of process modeling and visual observation data is limited to the feature-level stitching level.

[0003] Pure visual inspection cannot characterize the intrinsic coupling relationship between welding energy, material state, and metallurgical behavior. The inspection results can only reflect surface morphology features and are difficult to correlate with changes in internal metallurgical parameters. Physical simulation to solve partial differential equations has high computational latency and cannot be adapted to real-time detection scenarios. Heat source model parameters rely on offline calibration, resulting in low matching with real-time welding conditions. There are deviations between the model's predicted state and the actual observed state, making it impossible to achieve dynamic and accurate characterization of the welding process. Point-by-point calculation and comparative analysis of metallurgical parameters such as solidification rate and thermal cycle curves are difficult to implement. There are technical bottlenecks in the accurate location of abnormal positions and the output of structured reports.

[0004] It is necessary to achieve a unified data representation of the coupling relationship between welding energy, material state and metallurgical behavior. This requires combining forward inference through networks constrained by physical information with inverse optimization based on state inversion using multispectral fusion images to construct a digital twin of the dynamic welding process. This will enable real-time calculation and point-by-point comparison of metallurgical parameters, and achieve accurate identification and structured report output of welding anomalies. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent inspection method for metal welding processes based on industrial vision.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent detection method for metal welding processing based on industrial vision, comprising: A multimodal welding process tensor is constructed, which is used to characterize the coupling relationship between welding energy, material state and metallurgical behavior; The tensor of the multimodal welding process is input into a reasoning computation network based on physical information constraints. The reasoning computation network forward derives the theoretical evolution of the molten pool morphology and temperature gradient field by solving the simplified form of the partial differential equations of welding heat conduction and fluid mechanics. A state inversion algorithm is executed synchronously. The state inversion algorithm uses real-time acquired multispectral fusion images as observation data and welding process parameters as boundary conditions to inversely optimize the endogenous parameters of the welding heat source model, so as to minimize the difference between the model's predicted state and the observed state. By integrating the forward inference results of the inference computing network with the inverse optimization parameters of the state inversion algorithm, a digital twin of the dynamic welding process is obtained. Using the aforementioned digital twin of the dynamic welding process, the solidification rate, thermal cycle curve, and latent heat of phase transformation at each point on the welding line can be calculated in real time. The calculated solidification rate, thermal cycle curve, and latent heat of phase change are compared and analyzed point by point with the standard values ​​in the preset qualified process window. Based on the results of the point-by-point comparison analysis, abnormal locations in the welding area that deviate from the standard value are identified, and a structured inspection report containing the coordinates of the abnormal locations and the type of deviation is output.

[0007] As a further aspect of the present invention, the construction of the multimodal welding process tensor includes: The original optical image sequence of the welding area is acquired by a multispectral imaging system. The original optical image sequence includes synchronous image data of the visible light band, near-infrared band and thermal infrared band. The original optical image sequence is subjected to image registration and fusion processing to generate a multispectral fused image with spatial alignment. From the multispectral fusion image, the molten pool region, the heat-affected region, and the base material region are segmented, and the spectral intensity distribution, temperature field distribution, and texture feature distribution of each region are extracted; Based on the spectral intensity distribution, the temperature field distribution, and the texture feature distribution, a multimodal welding process tensor is constructed; The original optical image sequence is subjected to image registration and fusion processing to generate a multispectral fused image with spatial alignment, specifically including: Feature point detection is performed on the visible light band image, the near-infrared band image, and the thermal infrared band image respectively, and scale-invariant feature transform descriptors are extracted. By utilizing the extracted scale-invariant feature transformation descriptor, feature point matching and spatial transformation estimation are performed on the near-infrared band image, the thermal infrared band image, and the visible light band image to achieve pixel-level alignment between images of different bands. For each registered band image, a weight coefficient based on its physical meaning is assigned. The weight coefficient of the visible light band image focuses on surface morphology, the weight coefficient of the near-infrared band image focuses on internal structure, and the weight coefficient of the thermal infrared band image focuses on temperature information. A weighted average fusion algorithm is used to superimpose the registered images of each band according to their corresponding weight coefficients to generate a multispectral fusion image containing multidimensional physical information. The multispectral fused image is subjected to contrast enhancement and noise suppression processing to improve the robustness of subsequent feature extraction.

[0008] As a further aspect of the present invention, the molten pool region, the heat-affected region, and the base material region are segmented from the multispectral fusion image, including: On the multispectral fused image, a set of pixels higher than the melting temperature threshold of the base material is defined based on thermal infrared band information, and the set of pixels constitutes the initial molten pool region; Starting from the boundary of the initial molten pool region, regional growth is carried out in the surrounding area. The growth criterion combines the spectral gradient of the near-infrared band and the texture abrupt change of the visible light band. Growth stops at the boundary where the spectral and texture features tend to be stable. The area covered by the growth is the transition zone between the molten pool region and the heat-affected zone. A temperature range is defined around the transition zone. The upper limit of the temperature range is below the material phase transition point, and the lower limit is above the ambient temperature. The connected area that falls within this temperature range and is connected to the transition zone is identified as the heat-affected zone. The remaining portion of the multispectral fusion image that does not belong to the molten pool region or the heat-affected zone is classified as the base material region.

[0009] As a further aspect of the present invention, the tensor of the multimodal welding process is input into a reasoning computation network based on physical information constraints. This reasoning computation network, by solving a simplified form of the partial differential equations relating welding heat conduction and fluid dynamics, forward deduces the theoretical evolution of the molten pool morphology and the temperature gradient field, including: The inference computing network employs an encoder structure to map the multimodal welding process tensor into a high-dimensional latent space vector; The implicit space vector is embedded with a process parameter vector consisting of welding current, voltage, and speed. Construct a physical information decoder, the forward propagation process of which is equivalent to discretizing and solving a simplified three-dimensional transient heat conduction equation and the Navier-Stokes equation; The physical information decoder uses the latent space vector and the process parameter vector as initial and boundary conditions to perform iterative calculations at multiple time steps. Each iterative calculation outputs the predicted values ​​of the molten pool fluid velocity field, temperature field, and pressure field for the current time step; By combining the predicted values ​​from multiple consecutive time steps in chronological order, the theoretical evolution of the molten pool morphology and the temperature gradient field are obtained.

[0010] As a further aspect of the present invention, the synchronous execution state inversion algorithm, which uses real-time acquired multispectral fusion images as observation data and welding process parameters as boundary conditions, inversely optimizes the intrinsic parameters of the welding heat source model, including: Define a parameterized welding heat source model, which includes heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters; The actual geometric dimensions and temperature field distribution of the molten pool extracted from the multispectral fusion image are used as the observation data. Establish a positive simulation model of the welding process with the aforementioned welding process parameters as fixed boundary conditions; Construct a loss function that calculates the mean square error between the output of the positive simulation model of the welding process and the observed data; The gradient descent optimization algorithm is used to iteratively adjust the heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters of the welding heat source model in order to minimize the loss function; When the value of the loss function is lower than a preset threshold or the maximum number of iterations is reached, the optimization stops, and the parameters of the heat source model at this time are the optimized endogenous parameters.

[0011] As a further aspect of the present invention, the forward deduction results of the inference computing network and the inverse optimization parameters of the state inversion algorithm are integrated to obtain a digital twin of the dynamic welding process, including: The heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters obtained by the state inversion algorithm are used as updated boundary conditions and input into the physical information decoder of the inference computing network. Using a physical information decoder with updated boundary conditions, the forward inference of multiple time steps starting from the current moment is re-executed; During the simulation, the theoretical melt pool geometric parameters calculated by the physical information decoder at each time step are compared with the actual melt pool geometric parameters segmented from the multispectral fusion image at the corresponding time. Based on the residuals generated by the comparison, the values ​​of the fluid viscosity parameters in the physical information decoder are dynamically fine-tuned using a proportional-integral controller. After multiple rounds of feedback adjustments, the simulation system's output remains synchronized and consistent with the actual observation data in time and space, and the simulation system is established as the digital twin of the dynamic welding process.

[0012] As a further aspect of the present invention, the digital twin of the dynamic welding process is used to calculate the solidification rate, thermal cycle curve, and latent heat of phase transformation at each point on the welding line in real time, including: In the digital twin of the dynamic welding process, the movement and thermal history of a series of material points evenly distributed along the weld centerline are tracked. For each of the material points, record the data of its temperature change over time to form the original temperature time series of the material points; By performing a differential operation on the original temperature-time series, a curve of the cooling rate of the material point changing with time is obtained, and the negative value at a specific stage represents the solidification rate. From the original temperature time series, the time and temperature path experienced by the material point from the heating peak temperature to the specific phase transition temperature is extracted, and the time and temperature path constitutes the thermal cycle curve; Based on the shape of the thermal cycle curve and the chemical composition of the material point, the latent heat of phase change released or absorbed by the material point during the solid-state phase change during the cooling process is calculated by querying the material phase diagram database.

[0013] As a further aspect of the present invention, the step of comparing and analyzing the calculated solidification rate, the thermal cycling curve, and the latent heat of phase change with the standard values ​​in a preset qualified process window point by point includes: For each type of material and joint being welded, a qualified process window database is pre-defined; From the qualified process window database, retrieve the solidification rate standard range, thermal cycle curve standard envelope, and phase change latent heat standard threshold that match the current process. On the same cross-section of the weld, the actual solidification rate of each calculated point is compared with the standard range of solidification rates, and points that are below the lower limit of the range or above the upper limit of the range are marked. Overlay the actual thermal cycle curve of each point onto the standard envelope of the thermal cycle curve, and calculate the area of ​​the actual curve deviating from the standard envelope as the thermal cycle deviation of the point. The actual latent heat of phase change calculated at each point is compared with the standard threshold of latent heat of phase change to determine whether it meets the expected requirements of metallurgical microstructure transformation.

[0014] As a further aspect of the present invention, based on the results of the point-by-point comparison analysis, abnormal locations deviating from the standard value in the welding area are identified, including: Points marked with abnormal actual solidification rates, points with thermal cycling deviations exceeding allowable errors, and points with actual latent heat of phase change failing to meet requirements are merged into a set of potential anomalies. Spatial clustering analysis is performed on the set of potential outliers to aggregate adjacent outliers into an anomalous region; Calculate the geometric features of each identified anomalous region, including the region's area, equivalent diameter, centroid coordinates, and average deviation within the region; Based on the geometric characteristics of the abnormal regions and their corresponding deviation types, a defect risk level code is assigned to each abnormal region.

[0015] As a further aspect of the present invention, it also includes a step of visualizing the detection results and performing process correlation analysis: Establish a two-dimensional coordinate system on the surface of the welded workpiece, and map all identified abnormal areas and their centroid coordinates onto the two-dimensional coordinate system to generate an abnormal location distribution map; On the abnormal location distribution map, different colors and fill patterns are used to distinguish different defect risk level codes and deviation types; Extract welding process parameter records corresponding to abnormal areas in time during the welding process, including current, voltage, wire feed speed and welding torch posture; The welding process parameters are recorded and the abnormal location distribution map are spatiotemporally correlated to analyze the correlation patterns between fluctuations in specific process parameters and the occurrence of specific types of abnormal areas. The abnormal location distribution map, detailed geometric features of each abnormal area, risk level, associated process parameters and their correlation analysis patterns are integrated into a structured inspection report.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: A multimodal welding process tensor is constructed to characterize the coupling relationship between welding energy, material state, and metallurgical behavior. This transforms multi-source data from the welding process into a unified tensor form, enabling the integrated representation of welding energy, material state, and metallurgical behavior-related data. This avoids feature matching errors caused by heterogeneous multi-source data and eliminates the information fragmentation problem in the fusion process of data from different dimensions. As a result, the coupling characteristics of the welding process can be accurately quantified and represented, providing a standardized data input format for subsequent deduction and optimization.

[0017] The tensor of the multimodal welding process is input into a reasoning computation network based on physical information constraints to solve the simplified form of the partial differential equations of welding heat conduction and fluid dynamics. The evolution of the molten pool morphology and temperature gradient field are deduced in the forward direction. Simultaneously, the state inversion algorithm is executed with real-time acquired multispectral fusion images as observation data and welding process parameters as boundary conditions. The intrinsic parameters of the welding heat source model are optimized in the reverse direction. The forward deduction results and the inverse optimization parameters are fused to construct a dynamic digital twin of the welding process. This reduces the computational delay of solving the partial differential equations, adapts to the timing requirements of real-time detection, narrows the difference between the model's predicted state and the observed state, improves the dynamic adaptability of welding process modeling, realizes real-time digital mapping of the welding process, and supports point-by-point calculation of solidification rate, thermal cycle curve and latent heat of phase change, as well as anomaly location identification. Attached Figure Description

[0018] Figure 1 This is a flowchart of an intelligent inspection method for metal welding based on industrial vision, as described in this invention. Figure 2 A flowchart for dividing the molten pool region, the heat-affected zone, and the base material region; Figure 3 This is a curve comparing the predicted and measured values ​​of the molten pool width. Figure 4 Spatial distribution of abnormal points in the weld section; Figure 5 This is a comparative analysis chart of latent heat of phase change standards. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0020] See Figure 1A multimodal welding process tensor is constructed to characterize the coupling relationship between energy, material state, and metallurgical behavior during welding. This multimodal welding process tensor is input into a physical information-constrained inference computation network. This network forward-derives the theoretical evolution of the molten pool morphology and temperature gradient field by solving simplified forms of the partial differential equations of welding heat conduction and fluid dynamics. A state inversion algorithm is executed, using real-time acquired multispectral fusion images as observation data and welding process parameters as boundary conditions to inversely optimize the intrinsic parameters of a welding heat source model, aiming to minimize the difference between the model's predicted state and the observed state. The forward derivation results of the inference computation network and the inverse optimization parameters of the state inversion algorithm are fused to obtain a dynamically updated digital twin of the welding process. Using this dynamic digital twin, the solidification rate, thermal cycle curve, and latent heat of phase change at each point on the welding line are calculated in real time. The calculated solidification rate, thermal cycle curve, and latent heat of phase change are then compared point-by-point with the standard values ​​in a preset qualified process window. Based on the results of point-by-point comparative analysis, abnormal locations deviating from the standard values ​​in the welding area are identified, and a structured inspection report containing the coordinates of the abnormal locations and the type of deviation is output.

[0021] In one embodiment of the present invention, see [reference] Figure 2 The original optical image sequence of the welding area was acquired using a multispectral imaging system. This sequence included synchronized image data in the visible, near-infrared, and thermal infrared bands. The multispectral imaging system had a sampling frequency of 200 Hz, the visible light camera had a resolution of 2048 x 1536 pixels, and the near-infrared and thermal infrared cameras had resolutions aligned with the visible light camera. In a specific example, gas metal arc welding (GMAW) was performed on a 10 mm thick low-carbon steel plate. The multispectral imaging system was fixed behind the welding torch and simultaneously captured images in three bands at an angle perpendicular to the welding direction. The welding current was 280 amperes, the voltage was 30 volts, and the welding speed was 350 mm / min. In the acquired original optical image sequence, the thermal infrared image showed a maximum temperature of approximately 2100 degrees Celsius at the center of the molten pool, the near-infrared image showed the fluid flow characteristics inside the molten pool, and the visible light image clearly presented the oxide color and ripple morphology of the weld surface.

[0022] Image registration and fusion processing is performed on the original optical image sequence to generate a multispectral fused image with spatial alignment. In specific implementations, feature point detection is performed on the visible light, near-infrared, and thermal infrared images respectively, and scale-invariant feature transform descriptors are extracted. Using the extracted scale-invariant feature transform descriptors, feature point matching and spatial transformation estimation are performed on the near-infrared and thermal infrared images and the visible light image, achieving pixel-level alignment between images of different bands. In some embodiments, feature point matching uses a random sampling consensus algorithm to eliminate mismatched points, and then solves the homography transformation matrix, mapping the coordinates of each pixel in the thermal infrared and near-infrared images to the coordinate system of the visible light image. The mapped image is spatially perfectly aligned with the visible light image. For each registered band image, a weight coefficient based on its physical meaning is assigned: the weight coefficient for the visible light image focuses on surface morphology, the weight coefficient for the near-infrared image focuses on internal structure, and the weight coefficient for the thermal infrared image focuses on temperature information. A weighted average fusion algorithm is used to superimpose the registered images of each band according to their corresponding weight coefficients, generating a multispectral fused image containing multidimensional physical information. The formula for the weighted average fusion algorithm is expressed as:

[0023] in: This represents the pixel value at coordinates (x, y) in the merged image. This represents the pixel values ​​of the registered visible light band image. This represents the pixel values ​​of the registered near-infrared band image. This represents the pixel values ​​of the registered thermal infrared image. This represents the weighting coefficient for the visible light band. This represents the weighting coefficient for the near-infrared band. This represents the weighting coefficient for the thermal infrared band. In a specific example, the weighting coefficient is set to... , , This emphasizes the dominant role of temperature field information. In another specific example, if a greater emphasis is needed on weld morphology, the weighting coefficient can be adjusted to... , , Different weighting coefficients can lead to varying degrees of prominence of different physical information in the multispectral fused image, thus affecting the boundary accuracy of subsequent region segmentation. Contrast enhancement and noise suppression are applied to the multispectral fused image to improve the robustness of subsequent feature extraction. Optionally, a contrast-limited adaptive histogram equalization method can be used to enhance the overall and local contrast of the multispectral fused image, followed by a nonlocal mean filtering algorithm to suppress Gaussian noise and speckle noise in the image.

[0024] From the multispectral fusion image, the molten pool region, heat-affected zone, and base material region are segmented, and the spectral intensity distribution, temperature field distribution, and texture feature distribution of each region are extracted. In specific implementation, on the multispectral fusion image, a set of pixels with a temperature higher than the base material melting temperature threshold is defined based on thermal infrared band information; this set of pixels constitutes the initial molten pool region. The base material melting temperature threshold is set according to the melting point of the material being welded; for low-carbon steel, the melting temperature threshold is set to 1500 degrees Celsius. In the multispectral fusion image, all pixels with a temperature higher than 1500 degrees Celsius are classified as pixels in the initial molten pool region. Starting from the boundary of the initial molten pool region, region growth is performed outwards to the surrounding area. The growth criterion combines the spectral gradient of the near-infrared band and the texture abrupt change in the visible light band. Growth stops at the boundary where the spectral and texture features tend to stabilize, and the area covered by growth is the transition zone between the molten pool region and the heat-affected zone. In some embodiments, the criteria for region growth are defined as follows: if the absolute value of the near-infrared intensity gradient of an adjacent pixel is less than a threshold T_grad and the local binary mode texture feature distance in the visible light band is less than a threshold T_text, then the pixel is included in the growth region; otherwise, growth stops. A temperature range is defined around the transition zone, with the upper limit below the material phase transition point and the lower limit above the ambient temperature. Connected regions falling within this temperature range and connected to the transition zone are identified as heat-affected regions. For low-carbon steel, the temperature range can be defined as between 800 degrees Celsius and the Ac3 phase transition point temperature of the material, for example, 800 degrees Celsius to 1200 degrees Celsius. The remaining portion of the multispectral fused image that does not belong to either the molten pool region or the heat-affected region is classified as the base material region. A multimodal welding process tensor is constructed based on the spectral intensity distribution, temperature field distribution, and texture feature distribution. In practice, the average value and variance of pixels in the molten pool region, heat-affected region, and base material region in each channel of the multispectral fusion image are calculated as spectral intensity distribution features. The maximum value, minimum value, mean value, and standard deviation of the temperature field are extracted from the thermal infrared channel as temperature field distribution features. The local binary mode histogram is extracted from the visible light channel as texture feature distribution. These features are arranged into a three-dimensional array according to their spatial position relationship, namely the multimodal welding process tensor.

[0025] In one embodiment of the present invention, the inference computing network employs an encoder structure to map the multimodal welding process tensor into a high-dimensional latent space vector. The encoder structure can be composed of multiple three-dimensional convolutional layers and pooling layers connected in series, used to compress and abstract the spatial and feature dimension information contained in the multimodal welding process tensor. In a specific example, the input multimodal welding process tensor has a dimension of 256 x 256 x 8, representing a 256-pixel x 256-pixel image and its physical features in 8 channels. After processing by the encoder, a one-dimensional latent space vector of length 512 is output. Embedded in the latent space vector is a process parameter vector composed of welding current, voltage, and speed. For example, the process parameter vector could be [280, 30, 5.83], corresponding to a current of 280 amperes, a voltage of 30 volts, and a welding speed of 350 millimeters per minute (approximately 5.83 millimeters per second), respectively. It can be understood that the process parameter vector, after being transformed by a fully connected layer, is concatenated or added to the latent space vector, fusing them into an enhanced latent vector containing the initial state and boundary conditions.

[0026] A physical information decoder is constructed, whose forward propagation process is equivalent to discretizing and solving a simplified three-dimensional transient heat conduction equation and the Navier-Stokes equations. The physical information decoder can be composed of a series of fully connected layers and residual blocks, and its mathematical operations are designed as an implicit expression of the discrete form of the governing equations. The physical information decoder performs iterative calculations across multiple time steps using the latent space vector and process parameter vector as initial and boundary conditions. Each iterative calculation outputs the predicted values ​​of the molten pool fluid velocity field, temperature field, and pressure field for the current time step. In some embodiments, the time step is set to 0.001 seconds. Starting from the output state of the previous time step (or the initial enhanced latent vector), the physical information decoder calculates through its internal network and outputs a tensor characterizing the fluid velocity field, temperature field, and pressure field distribution of the molten pool region in the next time step. Combining the predicted values ​​from multiple consecutive time steps in chronological order yields the theoretical molten pool morphology evolution and temperature gradient field. For example, by continuously extrapolating 100 time steps and stacking the temperature field tensors output at each time step along the time dimension, a four-dimensional spatiotemporal temperature field evolution sequence can be obtained, from which the morphology of the melt pool and the temperature gradient at any time can be extracted.

[0027] The state inversion algorithm defines a parameterized welding heat source model, which includes heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters. In specific implementations, the welding heat source model can adopt a double ellipsoidal heat source model. Its energy distribution shape parameters include geometric parameters such as the major axis, minor axis, and depth of the front and rear hemispheres. The heat input efficiency parameter represents the ratio of arc energy to workpiece heat energy, and the heat loss coefficient parameter integrates convection and radiation heat loss. The actual geometric dimensions and temperature field distribution of the molten pool extracted from the multispectral fusion image are used as observation data. The observation data includes the length, width, maximum temperature, and spatial distribution of the temperature field of the molten pool. A welding process simulation positive model is established with welding process parameters as fixed boundary conditions. The welding process simulation positive model is a numerical simulation program based on the finite volume method. Its governing equations are consistent with the simplified equations used by the physical information decoder, but the solution is more accurate. A loss function is constructed to calculate the mean square error between the output of the welding process simulation positive model and the observation data. A gradient descent optimization algorithm is used to iteratively adjust the heat source energy distribution shape parameters, thermal input efficiency parameters, and heat loss coefficient parameters of the welding heat source model to minimize the loss function. Optionally, the Adam optimizer can be used with a learning rate of 0.01. The specific form of the loss function can be expressed as:

[0028] in: This indicates a dependence on the set of parameters of the heat source model. The loss value, where N represents the number of observed data points. Indicates the first On each data point, using parameters The output values ​​(such as temperature or weld pool size) of the positive model for simulating the welding process. Indicates the first The actual observed values ​​corresponding to each data point. In each iteration, the loss function is calculated. Regarding parameters gradient Then update the parameters according to the optimizer rules. When the value of the loss function falls below a preset threshold or the maximum number of iterations is reached, optimization stops, and the parameters of the heat source model at this point are the optimized endogenous parameters. In some embodiments, the preset threshold is set to 0.1 degrees Celsius squared, and the maximum number of iterations is set to 500. It can be understood that the optimized endogenous parameters make the predicted output of the welding process simulation positive model statistically closest to the actual observation data extracted from the multispectral fusion image, thereby calibrating the model.

[0029] In one embodiment of the present invention, the heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters obtained by the state inversion algorithm are used as updated boundary conditions and input into the physical information decoder of the inference computing network. In a specific example, after optimization by the state inversion algorithm, the length parameters of the front hemisphere of the double ellipsoidal heat source model are 3.2 mm, the length parameter of the rear hemisphere is 6.8 mm, the heat input efficiency parameter is 0.78, and the heat loss coefficient parameter is 15 watts per square meter Kelvin. These optimized endogenous parameters will replace the original default boundary condition parameters in the physical information decoder. Using the physical information decoder with updated boundary conditions, the forward inference of multiple time steps starting from the current moment is re-executed. For example, starting from the second second of the welding process, the latent space vector obtained by the tensor encoding of the multimodal welding process at the current moment is used as the initial state, and the inference calculation of 500 time steps in the next 0.5 seconds is performed in combination with the updated boundary conditions.

[0030] During the simulation, the theoretical melt pool geometry parameters calculated by the physical information decoder at each time step are compared with the actual melt pool geometry parameters segmented from the multispectral fusion image at the corresponding time. The theoretical melt pool geometry parameters include the melt pool length and width predicted by the physical information decoder, while the actual melt pool geometry parameters are the melt pool length and width measured in real-time from the synchronously acquired multispectral fusion image through image processing. For example, at time point 2.1 seconds, the physical information decoder predicts a melt pool length of 8.5 mm and a width of 5.1 mm, while the measured melt pool length from the image at the same time is 8.7 mm and a width of 4.9 mm. Based on the residual generated by the comparison, a proportional-integral (PI) controller dynamically fine-tunes the fluid viscosity parameter value in the physical information decoder. The PI controller outputs the adjustment amount for the fluid viscosity parameter based on the error between the predicted and measured values ​​of the geometry parameters and the integral of the error. The output formula of the PI controller is expressed as:

[0031] in: Indicates the first The fluid viscosity parameter needs to be applied within each control cycle. The adjustment amount, Represents the proportional gain coefficient. Represents the integral gain coefficient. Indicates the first The error between the predicted and measured values ​​of the molten pool geometric parameters within a control cycle. Indicates from the start of control to the number The cumulative sum of errors over several cycles. In some embodiments, the proportional gain coefficient... Set to 0.05, integral gain coefficient The value was set to 0.001, ensuring the control cycle remained consistent with the time step of the physical information decoder. After multiple rounds of feedback adjustments, the simulation system's output remained synchronized and consistent with the actual observation data in time and space. This simulation system was thus established as a digital twin of the dynamic welding process. It can be understood that a digital twin of the dynamic welding process is a computational model capable of reflecting the real state of the weld pool in real time and predicting its short-term evolution.

[0032] Using a digital twin of the dynamic welding process, the solidification rate, thermal cycle curve, and latent heat of phase transformation at each point on the welding line are calculated in real time. Within the digital twin, the movement and thermal history of a series of material points evenly distributed along the weld centerline are tracked; the spacing between these points can be set to 0.1 mm. For each material point, its temperature change over time is recorded, forming the original temperature-time series. The sampling interval of the original temperature-time series is the same as the calculation step size of the digital twin. Differential operations are performed on the original temperature-time series to obtain the curve of the cooling rate of the material point over time; negative values ​​at specific stages represent the solidification rate. For example, in the stage where the material point's temperature cools from the liquidus to the solidus, the average value of the cooling rate curve in that temperature range is negative, representing the average solidification rate of that point. From the original temperature-time series, the time and temperature path experienced by the material point from the peak heating temperature to the specific phase transformation temperature are extracted; this time and temperature path constitute the thermal cycle curve. Optionally, for low-carbon steel materials, a specific phase transformation temperature of 800 degrees Celsius can be selected. The thermal cycling curve characterizes the process of the material point cooling from its peak temperature to 800 degrees Celsius. Based on the shape of the thermal cycling curve and the chemical composition of the material point, a material phase diagram database is consulted to calculate the latent heat of phase transformation released or absorbed by the material point during the solid-state phase transformation during cooling. In some embodiments, the material phase diagram database stores the critical temperatures and latent heat values ​​of various phase transformations under different alloy compositions. By matching the actual thermal cycling curve of the material point with the standard phase transformation curve in the database and considering the influence of the cooling rate on the phase transformation kinetics, the value of the latent heat of phase transformation is calculated through interpolation.

[0033] See Figure 3In the comparative analysis of predicted and measured geometric parameters of the molten pool (2.0–2.5 seconds), the evolution of the molten pool width reflects the synergistic optimization effect of the physical information constraint inference network and the state inversion algorithm. From the curve shape, the predicted width (solid line) exhibits a "U-shaped" trend of first decreasing and then increasing: reaching a local minimum of approximately 4.95 mm near 2.2 seconds, and then continuously increasing to approximately 5.25 mm at 2.5 seconds; the measured width (dashed line) shows a symmetrical "V-shaped" change: dropping to a trough of approximately 4.80 mm at 2.2 seconds, and then gradually rising back to approximately 5.00 mm at 2.5 seconds. The phase and amplitude differences between the two curves reflect the calibration logic of the digital twin of the dynamic welding process: initial deviation: at 2.0 seconds, there is an initial offset of approximately 0.23 mm between the predicted value (approximately 5.16 mm) and the measured value (approximately 4.93 mm), stemming from the uncalibrated heat source model parameters and simplified physical field assumptions. Trend Synchronization: Both reached their extreme points at 2.2 seconds, indicating that the inference network accurately captured the physical mechanism of molten pool contraction-expansion, verifying the effectiveness of the simplified form of the heat conduction and fluid dynamics equations. Deviation Convergence: As time progressed, the absolute deviation between the predicted and measured values ​​gradually decreased from 0.23 mm at 2.0 seconds to approximately 0.25 mm at 2.5 seconds (due to the faster increase in the predicted value). This demonstrates the inverse optimization effect of the state inversion algorithm on the endogenous parameters of the heat source model, keeping the residual between the predicted and observed states within an acceptable range. The comparative results provide reliable morphological boundary constraints for subsequent calculations of solidification rate, thermal cycling curves, and latent heat of phase transformation, serving as fundamental data support for identifying welding quality anomalies.

[0034] In one embodiment of the present invention, a qualified process window database is preset for each type of material and joint to be welded. This database is stored in the system in the form of a lookup table or configuration file, and its content is predetermined based on numerous process experiments and metallographic analysis results. From the qualified process window database, the standard range of solidification rate, the standard envelope of the thermal cycling curve, and the standard threshold of latent heat of phase transformation matching the current process are retrieved. For example, for a butt joint made of Q235B low-carbon steel with a plate thickness of 10 mm and using gas metal arc welding (GMAW), the qualified process parameters retrieved from the database are: welding current 280±20 amperes, voltage 30±2 volts, and welding speed 350±30 mm / min. Refer to Table 1 for the corresponding qualified process window data.

[0035] Table 1: Qualified Process Window Data Table for Q235B Low Carbon Steel Butt Joints (10mm Thickness)

[0036] On the same cross-section of the weld, the calculated actual solidification rate of each point is compared with the standard range of solidification rates, and points below the lower limit or above the upper limit are marked. In a specific example, 100 material points are selected at equal intervals on the weld cross-section for calculation. The actual solidification rate of material point 15 is 10.2 degrees Celsius per second, which is lower than the lower limit of 12.5 degrees Celsius per second in Table 1, and this point is marked as an abnormal solidification rate point; the actual solidification rate of material point 48 is 28.5 degrees Celsius per second, which is higher than the upper limit of 25.0 degrees Celsius per second, and is also marked as an abnormal point. The actual thermal cycle curve of each point is overlaid on the standard envelope of the thermal cycle curve, and the area of ​​the actual curve deviating from the standard envelope is calculated as the thermal cycle deviation of the point. It can be understood that the standard envelope is enclosed by an upper limit curve and a lower limit curve in the temperature-time coordinate system, and the formula for calculating the thermal cycle deviation D is:

[0037] Where: D represents the thermal cycling deviation. and These are the start and end points of the time interval of the thermal cycling curve. Indicates time The actual temperature value, Indicates time The standard envelope boundary temperature value is used. When the actual temperature is inside the envelope, the temperature value of the closest boundary curve is taken. If the actual thermal cycle curve of material point 33 is completely inside the standard envelope during the cooling stage, its thermal cycle deviation D is zero; if the actual thermal cycle curve of material point 67 exceeds the upper limit of the envelope when cooled to around 600 degrees Celsius, its thermal cycle deviation D is a positive value. The calculated actual latent heat of phase transformation for each point is compared with the standard threshold for latent heat of phase transformation to determine whether it meets the expected requirements for metallurgical microstructure transformation.

[0038] Points with abnormal actual solidification rates, deviations from allowable thermal cycling parameters, and actual latent heats of phase change that do not meet requirements are merged into a potential anomaly set. Spatial clustering analysis is then performed on this set to group adjacent anomalies into anomalous regions. Optionally, density-based clustering algorithms can be used, such as setting a neighborhood radius of 0.5 mm and a minimum point threshold of 3; adjacent points meeting these conditions are grouped into the same cluster, forming an anomaly region. The geometric features of each identified anomaly region are calculated, including its area, equivalent diameter, centroid coordinates, and average deviation within the region.

[0039] See Figure 4In the visualization stage of the spatial distribution of abnormal points in the weld cross-section, the identification of abnormal regions relies on the coupled application of multiphysics index deviation analysis and spatial clustering technology. Specifically, the weld centerline serves as the baseline topological skeleton, derived from the digital twin of the dynamic welding process, representing the ideal weld geometry. Normal measurement points (circles) are material points whose solidification rate, thermal cycling curve, and latent heat of phase transformation all fall within the acceptable process window; their spatial distribution closely matches the weld centerline, reflecting stable metallurgical behavior. Abnormal measurement points (pentagons) are material points whose actual indicators deviate from the standard threshold, covering three typical types: excessive solidification rate, excessive thermal cycling deviation, and insufficient latent heat of phase transformation. Density-based spatial clustering analysis is performed on the potential abnormal point set. With a neighborhood radius of 0.5 mm and a minimum point threshold of 3, adjacent abnormal points can be aggregated into independent abnormal regions: the left peak region, the bottom valley region, and the right peak region each form three clusters, corresponding to different defect risk levels. The geometric features of the abnormal region, such as its centroid coordinates, equivalent diameter, and average deviation, can be further mapped to the two-dimensional coordinate system of the welded workpiece, providing a quantitative basis for subsequent defect tracing and process parameter optimization. This visualization method intuitively presents the spatial clustering pattern of abnormal points on the weld cross-section, realizing a cross-scale mapping from microscopic metallurgical index deviations to macroscopic defect region identification.

[0040] In one embodiment of the present invention, a two-dimensional coordinate system is established on the surface of the welded workpiece. All identified abnormal regions and their centroid coordinates are mapped onto the two-dimensional coordinate system to generate an abnormal location distribution map. The two-dimensional coordinate system has the lower left corner of the workpiece as the origin, the positive X-axis along the length of the weld, and the positive Y-axis perpendicular to the weld and pointing to the other side of the workpiece. The coordinate unit is millimeters. In a specific example, three abnormal regions are identified on a weld with a length of 200 millimeters. Their centroid coordinates, after mapping, are (45.2, 10.1), (102.5, 9.8), and (159.0, 10.3), respectively. These coordinate points, along with their abnormal region boundaries, are plotted on the abnormal location distribution map. On the anomaly location distribution map, different colors and filling patterns are used to distinguish different defect risk level codes and deviation types. For example, the area with defect risk level code "223" (representing thermal cycling anomaly, medium area, severe deviation) is filled with red diagonal lines, the area with defect risk level code "112" (representing solidification rate anomaly, small area, moderate deviation) is filled with blue grids, and the area with defect risk level code "311" (representing phase change latent heat anomaly, small area, slight deviation) is filled with green dotted patterns.

[0041] Extract welding process parameter records corresponding to the abnormal area in time during the welding process, including current, voltage, wire feed speed, and welding torch posture. In some embodiments, the welding torch posture includes the angle between the welding torch and the workpiece normal and the welding torch travel direction angle. For an abnormal area identified at 102.5 mm in the X coordinate of the workpiece coordinate system, its corresponding welding time point can be obtained by back-calculating the welding speed. Assuming the welding speed is 5.83 mm / s, the welding time corresponding to this abnormal area is approximately 17.6 seconds after the workpiece starts welding. Extract the average value of the process parameters within the time window from 17.5 seconds to 17.7 seconds from the welding process data record file, obtaining a current record of 305 amperes, a voltage record of 32 volts, a wire feed speed record of 7.2 meters per minute, and a welding torch posture record of a forward angle of 10 degrees. Spatiotemporally correlate the welding process parameter records with the abnormal location distribution map to analyze the correlation pattern between fluctuations in specific process parameters and the occurrence of specific types of abnormal areas. Optionally, the correlation analysis can use the method of calculating the Pearson correlation coefficient, expressed by the formula:

[0042] in: This represents the Pearson correlation coefficient, used to quantify linear correlation. Indicates the total number of abnormal areas discovered; Indicates the first The values ​​of specific process parameters (such as current) corresponding to each abnormal region; Indicates all The average value of the process parameter corresponding to each abnormal region; Indicates the first A certain characteristic quantity of an abnormal region; This indicates that the feature quantity is in The average value of each abnormal region. The calculated correlation coefficient. The closer the absolute value is to 1, the stronger the linear correlation between the fluctuation of the process parameter and this type of abnormal characteristic.

[0043] See Figure 5In the comparative analysis of latent heat of phase change (LCH) standards, the deviation between the actual LCH value of the welding area and the preset acceptable process window standard range is presented intuitively and quantitatively. Specifically, three abnormal regions identified along the welding path are used as the analysis objects. The actual LCH (J / g) calculated for each region is compared point-by-point with the preset standard threshold range (200J / g~300J / g): Region 1 has an actual value of 210J / g, near the lower limit of the standard range; Region 2 has an actual value of 320J / g, significantly exceeding the upper limit of the standard range, and is marked as a high-risk abnormal point "exceeding the standard"; Region 3 has an actual value of 205J / g, close to the lower limit of the standard range. By mapping the actual LCH data of each region to a two-dimensional coordinate graph, a LCH standard comparison curve is formed, intuitively displaying the LCH deviation characteristics of the abnormal regions: the peak deviation in Region 2 represents excessive release of LCH, possibly corresponding to metallurgical structure abnormalities caused by excessive welding heat input; the low-value deviations in Regions 1 and 3 reflect insufficient LCH release, or are related to excessively fast welding cooling rates and insufficient thermal circulation. This analysis provides direct data support for subsequent defect risk level coding and process parameter tracing. It can further correlate welding current, voltage and other process parameters in the corresponding time window to pinpoint the process root cause of the anomaly.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications 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 protection scope of the present invention.

Claims

1. An intelligent detection method for metal welding processing based on industrial vision, characterized in that, include: A multimodal welding process tensor is constructed, which is used to characterize the coupling relationship between welding energy, material state and metallurgical behavior; The tensor of the multimodal welding process is input into a reasoning computation network based on physical information constraints. The reasoning computation network forward derives the theoretical evolution of the molten pool morphology and temperature gradient field by solving the simplified form of the partial differential equations of welding heat conduction and fluid mechanics. A state inversion algorithm is executed synchronously. The state inversion algorithm uses real-time acquired multispectral fusion images as observation data and welding process parameters as boundary conditions to inversely optimize the endogenous parameters of the welding heat source model, so as to minimize the difference between the model's predicted state and the observed state. By integrating the forward inference results of the inference computing network with the inverse optimization parameters of the state inversion algorithm, a digital twin of the dynamic welding process is obtained. Using the aforementioned digital twin of the dynamic welding process, the solidification rate, thermal cycle curve, and latent heat of phase transformation at each point on the welding line can be calculated in real time. The calculated solidification rate, thermal cycle curve, and latent heat of phase change are compared and analyzed point by point with the standard values ​​in the preset qualified process window. Based on the results of the point-by-point comparison analysis, abnormal locations in the welding area that deviate from the standard value are identified, and a structured inspection report containing the coordinates of the abnormal locations and the type of deviation is output.

2. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 1, characterized in that, The construction of the multimodal welding process tensor includes: The original optical image sequence of the welding area is acquired by a multispectral imaging system. The original optical image sequence includes synchronous image data of the visible light band, near-infrared band and thermal infrared band. The original optical image sequence is subjected to image registration and fusion processing to generate a multispectral fused image with spatial alignment. From the multispectral fusion image, the molten pool region, the heat-affected region, and the base material region are segmented, and the spectral intensity distribution, temperature field distribution, and texture feature distribution of each region are extracted; Based on the spectral intensity distribution, the temperature field distribution, and the texture feature distribution, a multimodal welding process tensor is constructed. The original optical image sequence is subjected to image registration and fusion processing to generate a multispectral fused image with spatial alignment, specifically including: Feature point detection is performed on the visible light band image, the near-infrared band image, and the thermal infrared band image respectively, and scale-invariant feature transform descriptors are extracted. By utilizing the extracted scale-invariant feature transformation descriptor, feature point matching and spatial transformation estimation are performed on the near-infrared band image, the thermal infrared band image, and the visible light band image to achieve pixel-level alignment between images of different bands. For each registered band image, a weight coefficient based on its physical meaning is assigned. The weight coefficient of the visible light band image focuses on surface morphology, the weight coefficient of the near-infrared band image focuses on internal structure, and the weight coefficient of the thermal infrared band image focuses on temperature information. A weighted average fusion algorithm is used to superimpose the registered images of each band according to their corresponding weight coefficients to generate a multispectral fusion image containing multidimensional physical information. The multispectral fused image is subjected to contrast enhancement and noise suppression processing to improve the robustness of subsequent feature extraction.

3. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 2, characterized in that, From the multispectral fused image, the molten pool region, the heat-affected region, and the base material region are segmented, including: On the multispectral fused image, a set of pixels higher than the melting temperature threshold of the base material is defined based on thermal infrared band information, and the set of pixels constitutes the initial molten pool region; Starting from the boundary of the initial molten pool region, regional growth is carried out in the surrounding area. The growth criterion combines the spectral gradient of the near-infrared band and the texture abrupt change of the visible light band. Growth stops at the boundary where the spectral and texture features tend to be stable. The area covered by the growth is the transition zone between the molten pool region and the heat-affected zone. A temperature range is defined around the transition zone. The upper limit of the temperature range is below the material phase transition point, and the lower limit is above the ambient temperature. The connected area that falls within this temperature range and is connected to the transition zone is identified as the heat-affected zone. The remaining portion of the multispectral fused image that does not belong to the molten pool region or the heat-affected zone is classified as the base material region.

4. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 3, characterized in that, The tensor of the multimodal welding process is input into a reasoning computation network based on physical information constraints. This network, by solving a simplified form of the partial differential equations relating welding heat conduction and fluid dynamics, forward derives the theoretical evolution of the molten pool morphology and the temperature gradient field, including: The inference computing network uses an encoder structure to map the multimodal welding process tensor into a high-dimensional latent space vector; The implicit space vector is embedded with a process parameter vector consisting of welding current, voltage, and speed. Construct a physical information decoder, the forward propagation process of which is equivalent to discretizing and solving a simplified three-dimensional transient heat conduction equation and the Navier-Stokes equation; The physical information decoder uses the latent space vector and the process parameter vector as initial and boundary conditions to perform iterative calculations over multiple time steps. Each iterative calculation outputs the predicted values ​​of the molten pool fluid velocity field, temperature field, and pressure field for the current time step; By combining the predicted values ​​from multiple consecutive time steps in chronological order, the theoretical evolution of the molten pool morphology and the temperature gradient field are obtained.

5. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 4, characterized in that, The synchronous execution state inversion algorithm uses real-time acquired multispectral fusion images as observation data and welding process parameters as boundary conditions to inversely optimize the intrinsic parameters of the welding heat source model, including: Define a parameterized welding heat source model, which includes heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters; The actual geometric dimensions and temperature field distribution of the molten pool extracted from the multispectral fusion image are used as the observation data. Establish a positive simulation model of the welding process with the aforementioned welding process parameters as fixed boundary conditions; Construct a loss function that calculates the mean square error between the output of the positive simulation model of the welding process and the observed data; The gradient descent optimization algorithm is used to iteratively adjust the heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters of the welding heat source model in order to minimize the loss function; When the value of the loss function is lower than a preset threshold or the maximum number of iterations is reached, the optimization stops, and the parameters of the heat source model at this time are the optimized endogenous parameters.

6. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 5, characterized in that, By integrating the forward inference results of the inference computation network with the inverse optimization parameters of the state inversion algorithm, a digital twin of the dynamic welding process is obtained, including: The heat source energy distribution shape parameters, heat input efficiency parameters, and heat loss coefficient parameters obtained by the state inversion algorithm are used as updated boundary conditions and input into the physical information decoder of the inference computing network. Using a physical information decoder with updated boundary conditions, the forward inference of multiple time steps starting from the current moment is re-executed; During the simulation, the theoretical melt pool geometric parameters calculated by the physical information decoder at each time step are compared with the actual melt pool geometric parameters segmented from the multispectral fusion image at the corresponding time. Based on the residuals generated by the comparison, the values ​​of the fluid viscosity parameters in the physical information decoder are dynamically fine-tuned by a proportional-integral controller. After multiple rounds of feedback adjustments, the simulation system's output remains synchronized and consistent with the actual observation data in time and space, and the simulation system is established as the digital twin of the dynamic welding process.

7. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 6, characterized in that, Using the aforementioned digital twin of the dynamic welding process, the solidification rate, thermal cycling curve, and latent heat of phase transformation at each point on the welding line are calculated in real time, including: In the digital twin of the dynamic welding process, the movement and thermal history of a series of material points evenly distributed along the weld centerline are tracked. For each of the material points, record the data of its temperature change over time to form the original temperature time series of the material points; By performing a differential operation on the original temperature-time series, a curve of the cooling rate of the material point changing with time is obtained, and the negative value of a specific stage represents the solidification rate. From the original temperature time series, the time and temperature path experienced by the material point from the heating peak temperature to the specific phase transition temperature is extracted, and the time and temperature path constitutes the thermal cycle curve; Based on the shape of the thermal cycle curve and the chemical composition of the material point, the latent heat of phase change released or absorbed by the material point during the solid-state phase change during the cooling process is calculated by querying the material phase diagram database.

8. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 7, characterized in that, The step of comparing and analyzing the calculated solidification rate, the thermal cycling curve, and the latent heat of phase change with the standard values ​​in the preset qualified process window point by point includes: For each type of material and joint being welded, a qualified process window database is pre-defined; From the qualified process window database, retrieve the solidification rate standard range, thermal cycle curve standard envelope, and phase change latent heat standard threshold that match the current process. On the same cross-section of the weld, the actual solidification rate of each calculated point is compared with the standard range of solidification rates, and points that are below the lower limit of the range or above the upper limit of the range are marked. Overlay the actual thermal cycle curve of each point onto the standard envelope of the thermal cycle curve, and calculate the area of ​​the actual curve deviating from the standard envelope as the thermal cycle deviation of the point. The actual latent heat of phase change calculated at each point is compared with the standard threshold of latent heat of phase change to determine whether it meets the expected requirements of metallurgical microstructure transformation.

9. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 8, characterized in that, Based on the results of the point-by-point comparison analysis, abnormal locations deviating from the standard values ​​in the welding area were identified, including: The points marked with abnormal actual solidification rates, points with thermal cycling deviations exceeding the allowable error, and points with actual latent heat of phase change not meeting the requirements are merged into a set of potential anomalies. Spatial clustering analysis is performed on the set of potential outliers to aggregate adjacent outliers into an anomalous region; Calculate the geometric features of each identified anomalous region, including the region's area, equivalent diameter, centroid coordinates, and average deviation within the region; Based on the geometric characteristics of the abnormal regions and their corresponding deviation types, a defect risk level code is assigned to each abnormal region.

10. The intelligent inspection method for metal welding processing based on industrial vision as described in claim 9, characterized in that, It also includes steps for visualizing and mapping the test results and analyzing their correlation with the process: Establish a two-dimensional coordinate system on the surface of the welded workpiece, and map all identified abnormal areas and their centroid coordinates onto the two-dimensional coordinate system to generate an abnormal location distribution map; On the abnormal location distribution map, different colors and fill patterns are used to distinguish different defect risk level codes and deviation types; Extract welding process parameter records corresponding to abnormal areas in time during the welding process, including current, voltage, wire feed speed and welding torch posture; The welding process parameters are recorded and the abnormal location distribution map are spatiotemporally correlated to analyze the correlation patterns between fluctuations in specific process parameters and the occurrence of specific types of abnormal areas. The abnormal location distribution map, detailed geometric features of each abnormal area, risk level, associated process parameters and their correlation analysis mode are integrated into a structured inspection report.

Images