Welding parameter correction method for real-time molten pool morphology feedback

Abstract

The invention relates to the technical field of welding parameter correction, and provides a welding parameter correction method for real-time molten pool morphology feedback, which comprises the following steps of: capturing a molten pool image, an arc spectrum and an infrared radiation signal in real time, accurately extracting geometrical characteristics and contour vibration frequency of a molten pool, and dynamically evaluating arc stability and a thermal attenuation state. And when defects are detected, welding parameter correction is generated based on a process knowledge graph and Bayesian reasoning, collaborative correction of parameters such as welding current, welding voltage and welding speed is achieved through reinforcement learning optimization weight distribution, and the change trend of morphology characteristics of a molten pool is continuously monitored after correction. According to the method, the depth belief network is judged and optimized by combining the descending rate of the morphology defect probability, and the corrected successful case is supplemented to the process knowledge graph, so that the welding quality stability is remarkably improved, the method is particularly suitable for automatic production of a high-precision welding scene, the defect rate can be effectively reduced, and the production efficiency can be effectively improved.

Description

Technical Field

[0001] This application relates to the field of welding parameter correction technology, and in particular to a welding parameter correction method with real-time molten pool morphology feedback. Background Technology

[0002] In modern manufacturing, welding is a key processing technology widely used in aerospace, automobile manufacturing and shipbuilding industries. With the development of industrial automation and intelligent manufacturing, the traditional method of adjusting welding parameters based on human experience can no longer meet the production requirements of high precision and high efficiency. Existing technologies mainly collect information about the molten pool through a single sensor, such as using a visual sensor to obtain a two-dimensional image of the molten pool, or using a spectrometer to monitor the arc signal. However, such methods have obvious limitations. Single image analysis is difficult to accurately obtain three-dimensional information such as the depth of the molten pool and cannot fully reflect the morphology of the molten pool. Relying solely on the arc signal cannot correlate the thermal state and geometric characteristics of the molten pool, resulting in a lack of overall coordination in the adjustment of welding parameters.

[0003] Furthermore, while some existing technologies employ multi-sensor fusion strategies, their data processing methods are simplistic, often involving direct feature stitching without in-depth exploration of the potential relationships between features. This makes it difficult to accurately predict welding defects. In the parameter adjustment stage, adjustments lag under complex working conditions. Machine learning models suffer from insufficient training data and weak generalization ability, and lack dynamic evaluation and optimization mechanisms for model performance, failing to achieve adaptive correction of welding parameters. Most of these technologies fail to address the problem of how to achieve molten pool instability monitoring and decoupled correction of multiple welding parameters under highly dynamic working conditions. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides a welding parameter correction method based on real-time molten pool morphology feedback. The method includes: capturing a surface image of the molten pool in real time, simultaneously acquiring the arc spectral signal and the infrared radiation signal of the molten pool during the welding process, processing the surface image of the molten pool to extract the geometric features and contour vibration frequency of the molten pool, analyzing the Stark broadening of the arc spectral signal and measuring the plasma electron density to determine the arc stability, and obtaining the longitudinal temperature gradient of the molten pool based on the infrared radiation signal of the molten pool, and determining the thermal attenuation factor of the molten pool in combination with the thermal diffusivity of the material.

[0005] The geometric features, profile vibration frequency, and thermal attenuation factor of the molten pool are integrated into the molten pool morphology features. The molten pool morphology features are processed by a deep belief network to output the morphology defect probability. When the probability of any morphology defect is greater than the probability threshold, the welding parameters are corrected, and the process knowledge graph is called to generate the initial deviation of the welding parameters, which include welding current, welding voltage, and welding speed.

[0006] Welding parameters are comprehensively corrected based on arc stability, molten pool profile vibration frequency and thermal attenuation factor, and initial deviation of welding parameters. After the welding parameters are corrected, the changing trend of molten pool morphology is monitored. The average decrease rate of morphology defect probability within multiple consecutive sampling periods is used to determine whether the depth confidence network should be optimized, and the successfully corrected welding parameters are added to the process knowledge graph.

[0007] As an optional implementation, the geometric feature extraction logic includes:

[0008] The surface image of the molten pool is preprocessed by histogram equalization, and the preprocessed surface image is segmented by a deep neural network to identify the molten pool region.

[0009] The edge contour of the molten pool region is obtained by using an edge detection algorithm, and the coordinates of the edge contour are monitored to obtain the molten pool width.

[0010] Sinusoidal fringe structured light is projected onto the surface of the molten pool region to capture deformed fringe images. The deformed fringe images are processed by a phase unwrapping algorithm to obtain the degree of fringe distortion, and the surface coordinates of the molten pool region are obtained to extract the vertical distance from the bottom of the molten pool to the surface as the molten pool depth.

[0011] As an optional implementation, the logic for extracting the contour vibration frequency includes:

[0012] The surface images of the molten pool are continuously captured to form an image sequence. The image sequence is stacked in the time dimension to generate a spatiotemporal image cube to reflect the changes in the edge contour.

[0013] The dynamic time warping algorithm is used to match and align the changes in the edge contour to obtain the contour change data of the molten pool.

[0014] The contour variation data of the molten pool is decomposed into multiple intrinsic mode functions by empirical mode decomposition method, and the spectral analysis of the decomposed intrinsic mode functions is performed to extract the contour vibration frequency of the molten pool.

[0015] As an optional implementation, the logic for determining arc stability includes:

[0016] The arc spectral signal during the welding process was acquired by a spectrometer, and the arc spectral signal was decomposed into multiple scales by wavelet packet decomposition to extract spectral energy features in different frequency bands.

[0017] The spectral energy characteristics of Stark broadening are automatically identified by convolutional neural networks, and the plasma electron density is measured based on the full width at half maximum (FWHM) of Stark broadening.

[0018] Configure the fluctuation threshold range of electron density and compare it with the fluctuation of plasma electron density in multiple monitoring sampling periods to determine the stability of the arc.

[0019] As an optional implementation, the logic for determining the thermal attenuation factor includes:

[0020] The infrared radiation signal from the molten pool is spatially filtered, and the energy distribution characteristics of the infrared radiation signal at different spatial frequencies are extracted based on the frequency domain analysis method of Fourier transform.

[0021] The energy distribution characteristics of infrared radiation signals at different spatial frequencies are simulated by finite element analysis to inversely deduce the longitudinal temperature gradient of the molten pool.

[0022] Based on Fourier's law, the thermal decay factor of the molten pool is determined according to the longitudinal temperature gradient of the molten pool and the thermal diffusivity of the material.

[0023] As an optional implementation, the logic for generating the initial deviation of the welding parameters includes:

[0024] When the probability of any morphological defect exceeds the probability threshold, welding parameter correction is triggered, and the process knowledge graph is invoked to extract parameter rules associated with the morphological defect.

[0025] Based on case-based reasoning, the deviation of welding parameters is generated according to the parameter rules associated with morphological defects;

[0026] The deviation of welding parameters is used as the prior distribution of a Bayesian network. The adjustment direction and magnitude of the welding parameters are determined by the Bayesian network to generate the initial deviation of the welding parameters.

[0027] As an optional implementation, the integration sub-logic of the melt pool morphology features includes:

[0028] Data preprocessing is performed on the geometric characteristics, profile vibration frequency, and thermal attenuation factor of the molten pool;

[0029] A spatiotemporal attention fusion network is constructed to determine the relational weights of different features to the morphology of the melt pool. The spatiotemporal attention fusion network includes a spatial attention module and a temporal attention module.

[0030] The features, weighted by relational weights, are input into a graph convolutional network to construct a feature association graph. The feature association graph is then continuously processed through graph convolution operations to integrate it into the molten pool morphology features.

[0031] As an optional implementation, the output sub-logic of the morphological defect probability includes:

[0032] Deep belief networks are used to process molten pool morphology features to extract the correlation patterns between molten pool morphology and defects.

[0033] The deep belief network is adjusted through supervised learning, and the parameters of the deep belief network are optimized by using historical welding defect data as labels. The deep belief network outputs the probability of morphological defects including undercut, hump and porosity.

[0034] The confidence interval of the shape defect probability is obtained based on the distribution of the shape defect probability output by the deep belief network, and the relationship weight is adjusted based on the confidence interval of the shape defect probability.

[0035] As an optional implementation, the optimization judgment logic of the deep belief network includes:

[0036] Calculate the average decrease rate of the probability of morphological defects within multiple consecutive sampling periods, and determine the fluctuation entropy of the probability of morphological defects within multiple consecutive sampling periods;

[0037] The optimal trigger probability is determined by combining the average decrease rate of the probability of morphological defects and the fluctuation entropy through Bayesian inference.

[0038] When the optimization trigger probability is greater than the probability threshold, the deep belief network is optimized. After optimizing the deep belief network, the rate of decrease in the probability of morphological defects is monitored to roll back and update the optimization trigger probability.

[0039] As an optional implementation, the welding parameter correction sub-logic includes:

[0040] Arc stability, molten pool profile vibration frequency, thermal attenuation factor, and initial deviation of welding parameters are used as the state space for reinforcement learning to optimize the weight allocation strategy for welding parameter correction.

[0041] Welding parameters are adjusted by weighted summation based on a weighted allocation strategy.

[0042] After the corrected welding parameters are input into the welding equipment for adjustment, the change trend of the molten pool morphology is used to determine whether a secondary correction is triggered and to determine the correction step size of the welding parameters.

[0043] Compared with existing technologies, the beneficial effects of this application are as follows: by capturing molten pool images, arc spectra, and infrared radiation signals in real time, it accurately extracts multi-dimensional parameters such as geometric features and contour vibration frequency of the molten pool, dynamically evaluates arc stability and thermal decay status, and then predicts the probability of morphological defects such as undercut and porosity. When a defect is detected, welding parameter correction is triggered based on the process knowledge graph and Bayesian inference. Through reinforcement learning, the weight allocation is optimized to achieve coordinated correction of parameters such as welding current, welding voltage, and welding speed. After correction, the changing trend of molten pool morphological features is continuously monitored. The reduction rate of morphological defect probability is combined to determine the optimization of the deep belief network, and successful correction cases are added to the process knowledge graph, thereby significantly improving the stability of welding quality. It is especially suitable for automated production in high-precision welding scenarios, and can effectively reduce the defect rate and improve production efficiency. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0045] Figure 1 A flowchart illustrating a welding parameter correction method based on real-time molten pool morphology feedback provided in an embodiment of this application;

[0046] Figure 2 The logic diagram for extracting the contour vibration frequency of a welding parameter correction method for real-time molten pool morphology feedback provided in the embodiments of this application;

[0047] Figure 3 An integrated sub-logic diagram of the weld pool morphology features of a welding parameter correction method with real-time weld pool morphology feedback provided in an embodiment of this application;

[0048] Figure 4 The output sub-logic diagram of the morphology defect probability of a welding parameter correction method with real-time molten pool morphology feedback provided in the embodiments of this application is shown. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0050] like Figure 1 The diagram shown is a flowchart of a welding parameter correction method based on real-time molten pool morphology feedback, provided in an embodiment of this application. The method includes:

[0051] S1. Real-time capture of the surface image of the molten pool, simultaneously acquiring the arc spectral signal and the infrared radiation signal of the molten pool during the welding process, processing the surface image of the molten pool to extract the geometric features and contour vibration frequency of the molten pool, analyzing the Stark broadening of the arc spectral signal and measuring the plasma electron density to determine the arc stability, and obtaining the longitudinal temperature gradient of the molten pool based on the infrared radiation signal of the molten pool, in combination with the thermal diffusivity of the material to determine the thermal attenuation factor of the molten pool.

[0052] Specifically, the logic for extracting geometric features includes:

[0053] The surface image of the molten pool is preprocessed by histogram equalization, and the preprocessed surface image is segmented by a deep neural network to identify the molten pool region.

[0054] The edge contour of the molten pool region is obtained by using an edge detection algorithm, and the coordinates of the edge contour are monitored to obtain the molten pool width.

[0055] Sinusoidal fringe structured light is projected onto the surface of the molten pool region to capture deformed fringe images. The deformed fringe images are processed by a phase unwrapping algorithm to obtain the degree of fringe distortion, and the surface coordinates of the molten pool region are obtained to extract the vertical distance from the bottom of the molten pool to the surface as the molten pool depth.

[0056] During welding, the strong radiation of the electric arc light, interference from metal spatter particles, and fluctuations in ambient light can cause uneven brightness distribution and low contrast in the surface image of the molten pool. The central area of ​​the arc may lose detail due to overexposure, while the edge areas may become blurred due to insufficient light. Directly identifying the molten pool area under these conditions can lead to boundary misjudgment. Therefore, preprocessing is needed to enhance image contrast, making the grayscale difference between the molten pool and the background more significant. Histogram equalization divides the surface image of the molten pool into multiple overlapping sub-blocks, independently calculating the grayscale distribution and expanding the dynamic range for each sub-block. Compared to global equalization, this method avoids false edge problems caused by amplified local noise. The preprocessed surface image is segmented using a deep neural network. This deep neural network extracts semantic features of the molten pool, including shape and texture, through an encoder structure, recovers spatial details through a decoder structure, and hop connections fuse shallow edge information with deep semantic information. The preprocessed surface image is input into the deep neural network and outputs a binary mask, where the white area represents the identified molten pool area. The grayscale distinction between the molten pool edge and the background is significantly improved in the preprocessed surface image, and the originally blurry solid-liquid interface becomes clear. Through end-to-end learning, the deep neural network can accurately segment the molten pool area, providing a clean target area for subsequent edge detection.

[0057] Weld width is a key geometric parameter for measuring weld formation, and its fluctuations directly reflect the stability of welding heat input. As the boundary between the solid and liquid phases, the edge of the weld pool can intuitively reflect the dynamic evolution of the weld width through coordinate changes. Therefore, it is necessary to extract the precise boundary of the weld pool through edge detection algorithms, and then convert the boundary features into quantified weld width values ​​through coordinate statistics, providing a basis for real-time judgment of welding quality. The edge detection algorithm obtains the edge contour of the weld pool region. First, Gaussian smoothing is applied to the segmented weld pool region to suppress false edges caused by noise. Then, potential edge points are determined by calculating the gradient magnitude and direction. Finally, non-maximum suppression and double threshold screening are used to retain continuous and real edges. This edge detection algorithm, through multi-level processing, can reduce noise interference while preserving edge details, thereby obtaining a continuous edge contour with a single pixel width.

[0058] Coordinate sampling is performed on the edge contour. Each pixel column is traversed along the horizontal axis of the image, and the highest and lowest ordinates of the edge contour of the molten pool region in that column are recorded. The difference between the two is the local molten width at that location. The average value of the local molten widths corresponding to all columns is taken as the molten pool width. The edge detection algorithm can extract continuous edges with a single pixel width. The edge positioning accuracy meets the requirements of molten width measurement. Through coordinate monitoring and statistics, the dynamic change trend of molten width can be tracked in real time. When the welding current fluctuates, the corresponding change in molten width can be captured in time, providing real-time feedback for welding parameter correction.

[0059] Two-dimensional images cannot directly reflect the depth information of the molten pool, while structured light measurement technology can recover the height distribution of the object's surface through the deformation of the fringe projection. Sinusoidal fringes, due to their periodic phase characteristics, are easy to calculate the height difference through phase changes, but the phase value has a 2π periodic ambiguity, which needs to be converted into a continuous phase through an unwrapping algorithm in order to accurately obtain the three-dimensional coordinates. By combining a semiconductor laser and a grating, a sinusoidal fringe structured light with a fixed period is generated and projected vertically onto the surface of the molten pool region. When the sinusoidal fringe structured light is projected onto the uneven surface of the molten pool region, the surface of the molten pool region will deform, and the image of this deformed fringe is captured in real time by a camera.

[0060] A three-step phase-shifting process is performed on the deformed fringe image to obtain a wrapped phase map. This involves projecting fringes with three different phases, followed by phase unwrapping using a quality-guided path tracing algorithm. The smoothness of the phase gradient is used as a quality metric, and unwrapping proceeds gradually from high-quality regions to low-quality regions. High-quality regions are areas with gradual phase changes, while low-quality regions are areas with phase jumps, avoiding unwrapping errors caused by noise and thus obtaining a continuous absolute phase map. Then, based on the camera and laser calibration parameters, including baseline distance and intrinsic / extrinsic parameter matrices, a mapping relationship between phase values ​​and actual height is established, transforming the absolute phase map into a continuous absolute phase map. The phase value is converted into the surface coordinates of the molten pool region. The vertical distance from the bottom of the molten pool to the surface is taken as the weld depth. Preferably, the surface coordinates of the molten pool region are (x, y, z), then the bottom of the molten pool is the point with the minimum z value, and the surface is the point with the maximum z value, where the z-axis direction is perpendicular to the direction of the welded workpiece. The structured light projection technology can acquire the three-dimensional morphology of the molten pool surface in real time. Even when the molten pool is dynamically fluctuating, it can capture instantaneous height changes through a high-speed camera. The phase unwrapping algorithm eliminates periodic ambiguity, ensures the continuity of the surface coordinates, and enables the weld depth measurement to accurately reflect the actual penetration state of the molten pool.

[0061] Specifically, such as Figure 2 As shown, the logic for extracting the contour vibration frequency includes:

[0062] The surface images of the molten pool are continuously captured to form an image sequence. The image sequence is stacked in the time dimension to generate a spatiotemporal image cube to reflect the changes in the edge contour.

[0063] The dynamic time warping algorithm is used to match and align the changes in the edge contour to obtain the contour change data of the molten pool.

[0064] The contour variation data of the molten pool is decomposed into multiple intrinsic mode functions by empirical mode decomposition method, and the spectral analysis of the decomposed intrinsic mode functions is performed to extract the contour vibration frequency of the molten pool.

[0065] The vibration of the molten pool profile is a dynamic process. A single frame image can only record the profile state at a certain time and cannot reflect its evolution over time. Therefore, it is necessary to continuously capture image sequences to transform the changes in the time dimension into features in the spatial dimension. The generation of a spatiotemporal image cube can transform the dynamic vibration of the edge profile into a spatiotemporal pattern in three-dimensional data, which is convenient for subsequent analysis of the periodicity and frequency characteristics of the vibration. By continuously capturing surface images of the molten pool at a constant frame rate with a camera to form an image sequence, it is ensured that key vibration details are not lost during the welding process. Each frame of surface image is preprocessed and the molten pool is segmented in sequence to obtain the coordinates of the edge profile of each frame. The edge profile coordinates of multiple consecutive frames are stacked in chronological order to generate a spatiotemporal image cube. The spatiotemporal image cube transforms the dynamic changes of the edge profile of the molten pool region into computable three-dimensional data, so that the vibration pattern that originally changed over time can be observed and analyzed in the spatial dimension.

[0066] During welding, the vibration of the molten pool is affected by factors such as arc stability and droplet transfer, exhibiting non-uniform motion characteristics. Directly analyzing it in chronological order would lead to phase shifts in vibrations at different periods, thus affecting the accuracy of frequency calculations. Therefore, this method selects the central profile contour along the width of the molten pool within the spatiotemporal image cube as the analysis object, forming a time series, i.e., the change of contour coordinates over time. Then, a distance matrix for a dynamic time warping algorithm is constructed, where matrix elements represent the degree of difference in edge contour states at different time points, obtainable through Euclidean distance. Next, dynamic programming is used to solve for the optimal path, which satisfies the continuity and monotonicity constraints of time warping, ensuring that the time-warped sequence is preserved. The original temporal relationship of vibrations is preserved. When the vibration speed increases, the path will compress the corresponding interval on the time axis, and when the speed decreases, the path will stretch the corresponding interval, thereby homogenizing the vibration period of the entire sequence. Finally, the normalized sequence is subjected to differential processing to obtain the contour change amount between adjacent time points, i.e., contour change data. This data eliminates the influence of non-uniform velocity on the time axis and only retains the true displacement change of the contour vibration. The dynamic time warping algorithm can effectively handle the time alignment problem of non-uniform vibration, transforming the originally phase-disordered vibration sequence into a regular sequence with equal time intervals, providing reliable time series data for subsequent frequency analysis and avoiding misjudgment of frequency components due to time offset.

[0067] Molten pool profile vibration typically contains multiple frequency components, including the arc's own oscillation frequency, periodic disturbances caused by droplet transition, and low-frequency fluctuations caused by thermal stress. Empirical Mode Decomposition (EMD) is an adaptive signal decomposition method that can decompose complex signals into multiple intrinsic mode functions (EMFs). Each EMF corresponds to an independent vibration mode and has a specific frequency range. By performing spectral analysis on the EMFs, the energy distribution of each frequency component can be separated and extracted, thereby identifying the frequency characteristics of the dominant vibration mode. The upper and lower envelopes of the aligned profile change data are determined by cubic spline interpolation. The mean of the envelope is calculated and subtracted from the original data to obtain candidate EMFs. The screening process is repeated until the candidate EMFs satisfy the conditions that the number of extrema and zero-crossings is similar and the envelope is locally symmetrical, thus obtaining the first EMF. Then, the above steps are repeated for the remaining data to sequentially decompose multiple EMFs until the remaining data is a monotonic trend term.

[0068] Finally, a Fourier transform is performed on each intrinsic mode function to convert the time-domain vibration signal into a frequency-domain energy distribution, generating a spectrum. By analyzing the peak positions in the spectrum, the frequency values ​​of each vibration mode are determined, and the dominance of the frequency component is judged based on the magnitude of the peak energy, which serves as the profile vibration frequency of the molten pool. Empirical mode decomposition does not require preset basis functions and can adaptively separate vibration modes at different time scales, making it particularly suitable for processing non-stationary signals such as molten pool vibration. Through spectrum analysis, the frequency characteristics corresponding to different vibration sources can be accurately identified, including the high-frequency components of arc oscillation and the low-frequency components of droplet transition, providing a basis for judging the stability of the welding process. The extracted profile vibration frequency, combined with the geometric features and thermal attenuation factor of the molten pool, constitutes the dynamic characteristic dimension of the molten pool morphology.

[0069] Specifically, the logic for judging arc stability includes:

[0070] The arc spectral signal during the welding process was acquired by a spectrometer, and the arc spectral signal was decomposed into multiple scales by wavelet packet decomposition to extract spectral energy features in different frequency bands.

[0071] The spectral energy characteristics of Stark broadening are automatically identified by convolutional neural networks, and the plasma electron density is measured based on the full width at half maximum (FWHM) of Stark broadening.

[0072] Configure the fluctuation threshold range of electron density and compare it with the fluctuation of plasma electron density in multiple monitoring sampling periods to determine the stability of the arc.

[0073] During welding, the arc spectral signal is affected by factors such as arc length variation, droplet transfer behavior, and welding parameter fluctuations, exhibiting complex non-stationary characteristics. Direct analysis of the original spectral signal makes it difficult to effectively separate the frequency components corresponding to different physical processes such as arc plasma oscillation and droplet transfer. Therefore, after acquiring the arc spectral signal through a spectrometer, it is necessary to perform multi-scale decomposition of the arc spectral signal using wavelet packet decomposition. In practice, the spectrometer probe is kept at a fixed working distance from the arc to ensure that the complete spectral band containing Stark broadening information can be acquired. Considering the time-frequency characteristics of the arc spectral signal, wavelet basis functions are selected for three-level wavelet packet decomposition, which divides the original arc spectral signal into multiple different frequency bands. This allows the separation of high-frequency components generated by the arc's own oscillation from low-frequency components caused by droplet transfer. After decomposition, the spectral energy features within each frequency band are extracted. These spectral energy features can intuitively reflect the dynamic changes of the arc plasma, providing a structured frequency domain input for subsequent Stark broadening feature identification and effectively avoiding feature extraction errors caused by signal aliasing.

[0074] Stark broadening is caused by the collision of charged particles in plasma, resulting in the broadening of spectral lines. The full width at half maximum (FWHM) of the Stark broadening is positively correlated with the plasma electron density. However, traditional manual identification methods struggle to accurately capture Stark broadening characteristics in complex spectral backgrounds. Therefore, a convolutional neural network (CNN) is used to automatically identify the spectral energy characteristics of Stark broadening. This CNN takes the frequency band energy sequence obtained from wavelet packet decomposition as input data, extracts local features through multiple convolutional layers, and then maps these local features to the plasma electron density via fully connected layers. During the training phase of the CNN, spectral data under different welding currents and voltages are acquired, and the corresponding electron densities are measured using a Langmuir probe as labels. By learning the distortion characteristics of the spectral line profiles, including peak position shifts and edge blurring, the CNN can automatically locate the Stark broadening region and calculate its FWHM. Then, based on the FWHM of the Stark broadening, the plasma electron density is measured. Compared to traditional spectral analysis, this method can more stably obtain electron density in the face of spatter interference or severe arc fluctuations, providing key parameter support for arc stability assessment.

[0075] Arc stability is essentially a reflection of the smoothness of the plasma state. Electron density, as a key parameter of plasma, directly reflects the stability of the arc. To quantify arc stability, it is necessary to first determine the normal fluctuation threshold range of electron density under different welding processes based on a large amount of historical welding data. This fluctuation threshold range is obtained based on the mean and standard deviation of historical welding data. For a certain type of welding process, the fluctuation threshold range can be set to a certain multiple of the standard deviation above and below the mean. During real-time monitoring, the plasma electron density is processed through a sliding window mechanism to calculate the fluctuation index of the mean offset and the difference between the maximum and minimum values ​​of electron density within the window. To avoid misjudgment caused by accidental interference, the fluctuation of multiple windows needs to be continuously monitored. Only when more than a certain proportion of the window data exceeds the normal fluctuation threshold range is the arc considered to be in an unstable state. When a short-circuit transition causes a drastic change in electron density, multi-cycle comparison can accurately identify the true abnormal fluctuation. The result of the arc stability judgment will directly affect the correction of subsequent welding parameters. When the arc is determined to be unstable, parameters such as welding voltage will be adjusted first to stabilize the arc length and ensure the smooth progress of the welding process.

[0076] Specifically, the logic for determining the thermal decay factor includes:

[0077] The infrared radiation signal from the molten pool is spatially filtered, and the energy distribution characteristics of the infrared radiation signal at different spatial frequencies are extracted based on the frequency domain analysis method of Fourier transform.

[0078] The energy distribution characteristics of infrared radiation signals at different spatial frequencies are simulated by finite element analysis to inversely deduce the longitudinal temperature gradient of the molten pool.

[0079] Based on Fourier's law, the thermal decay factor of the molten pool is determined according to the longitudinal temperature gradient of the molten pool and the thermal diffusivity of the material.

[0080] During the acquisition of infrared radiation signals from the molten pool, interference from thermal noise and measurement noise in the welding environment is easily induced, resulting in the inclusion of radiation characteristics from non-molten pool areas in the original signal. To obtain an effective signal that accurately reflects the temperature distribution of the molten pool, spatial filtering of the infrared radiation signal is necessary. In practice, the molten pool is observed in real time using an infrared thermal imager. The imager is installed at a certain angle to the workpiece surface to avoid direct arc light. First, a Gaussian low-pass filter is used to smooth the infrared image, suppressing the influence of high-frequency noise such as spatter particles. Then, a band-pass filter is used to extract the signal that matches the size of the molten pool. The spatial frequency components are extracted, preserving the temperature gradient characteristics at the edge of the molten pool. Then, a two-dimensional Fourier transform is performed on the filtered infrared image to convert the spatial temperature distribution into a frequency energy distribution. Different spatial frequencies in the frequency domain correspond to different characteristics of the molten pool temperature field. High-frequency energy distribution reflects regions of abrupt temperature gradient changes, including the solid-liquid interface of the molten pool, while low-frequency energy distribution corresponds to regions with relatively gentle temperature changes, such as the center of the molten pool. Through this series of processing steps, noise interference is effectively removed, enabling the frequency domain characteristics to accurately reflect the spatial variations of the molten pool temperature field, providing reliable input data for subsequent finite element analysis.

[0081] There is a close relationship between the frequency domain characteristics of infrared radiation signals and the spatial distribution of the molten pool temperature field. However, this relationship is difficult to directly deduce the temperature gradient using simple analytical methods. Therefore, the finite element method is used to invert the longitudinal temperature gradient. When constructing the transient heat conduction model, the molten pool region is divided into a fine tetrahedral mesh to ensure accurate capture of subtle temperature changes. Then, the thermal properties of the workpiece material, including the thermal conductivity versus temperature curve, are imported from a material database. Regarding boundary condition settings, a double ellipsoidal heat source model is used to simulate the arc heat input, and a convective heat dissipation boundary is set on the bottom surface of the workpiece. In the inversion... The process employs an iterative optimization using a genetic algorithm. First, an initial longitudinal temperature gradient is preset. The corresponding infrared radiation frequency domain characteristics are calculated using a finite element model. Then, the simulated frequency domain characteristics are compared with the actual measured characteristics, and the error between the two is calculated. Based on the error result, the longitudinal temperature gradient is adjusted, and the simulation is repeated. This iterative process continues until the error between the simulated and measured characteristics meets the preset requirements. This fully considers the influence of material thermophysical parameters and boundary conditions, accurately capturing the abrupt temperature gradient change at the bottom of the molten pool caused by intense heat dissipation, and providing key longitudinal temperature gradient parameters for calculating the thermal decay factor.

[0082] The thermal decay process of the molten pool is essentially a process of heat conduction from the workpiece to the surroundings. Its decay rate is determined by both the temperature gradient and the thermophysical properties of the material. Fourier's law reveals the basic law of heat conduction, indicating that the heat flux density is proportional to the temperature gradient, with the proportionality coefficient being the thermal diffusivity of the material. In order to comprehensively reflect the thermal decay characteristics of the molten pool, the concept of a thermal decay factor is introduced. This thermal decay factor needs to consider both the temperature gradient and the thermal diffusivity of the material. The thermal diffusivity of the material reflects the material's ability to conduct and store heat. When calculating the thermal decay factor, the heat flux density is first calculated based on the longitudinal temperature gradient and the thermal diffusivity of the material obtained by reverse calculation. Then, it is processed by dimensionless processing. The thermal decay factor = heat flux density / (molten pool surface temperature - initial workpiece temperature), and the heat flux density = - thermal diffusivity of the material × longitudinal temperature gradient.

[0083] Because the material temperature changes during welding, leading to fluctuations in thermal properties, it is necessary to update the material's thermal diffusivity in real time. The thermal decay factor quantifies the heat dissipation rate of the molten pool from a physical perspective, effectively distinguishing the degree of thermal decay under different welding parameters. Specifically, the thermal decay factor increases with increasing welding speed, indicating a faster cooling rate of the molten pool. As a key feature of the molten pool's thermal state, the thermal decay factor, when integrated with parameters such as the molten pool's geometric features and contour vibration frequency, will have a significant impact on the depth confidence network's judgment of the probability of morphological defects. An excessively high thermal decay factor can lead to rapid solidification of the molten pool, thereby increasing the risk of defects such as porosity. This provides important thermophysical basis for subsequent welding quality assessment and parameter correction.

[0084] S2. Integrate the geometric features, contour vibration frequency, and thermal attenuation factor of the molten pool into molten pool morphology features. Process the molten pool morphology features through a deep belief network to output the probability of morphology defects. When the probability of any morphology defect is greater than the probability threshold, trigger the correction of welding parameters and call the process knowledge graph to generate the initial deviation of welding parameters. Welding parameters include welding current, welding voltage, and welding speed.

[0085] Furthermore, such as Figure 3 As shown, the integration sub-logic for the molten pool morphology features includes:

[0086] Data preprocessing is performed on the geometric characteristics, profile vibration frequency, and thermal attenuation factor of the molten pool;

[0087] A spatiotemporal attention fusion network is constructed to determine the relational weights of different features to the morphology of the melt pool. The spatiotemporal attention fusion network includes a spatial attention module and a temporal attention module.

[0088] The features, weighted by relational weights, are input into a graph convolutional network to construct a feature association graph. The feature association graph is then continuously processed through graph convolution operations to integrate it into the molten pool morphology features.

[0089] The geometric features, contour vibration frequencies, and thermal attenuation factors of the molten pool are obtained from different devices and processing methods, resulting in issues such as inconsistent data types, large differences in dimensions, and noise interference. Directly using the raw data would lead to instability in subsequent network training and poor feature fusion performance. Therefore, preprocessing is necessary to standardize and reduce noise, ensuring comparability between different features. For the geometric features of the molten pool, median filtering is used to remove outliers, i.e., instantaneous size changes caused by welding spatter. The contour vibration frequencies are normalized to map the data to the same value range, eliminating differences in frequency magnitudes. For the thermal attenuation factor, which involves material properties and temperature field calculations, missing values ​​are first imputed using a weighted average of adjacent time data to fill the gaps, and then the data curve is smoothed using a moving average filter to reduce noise caused by temperature fluctuations. In addition, all feature data are subjected to one-hot encoding or standardization transformation to meet the network input requirements. Data preprocessing effectively improves data quality, eliminates the influence of dimensions and outliers between different features, makes the geometric features more reflective of the true size of the molten pool, smooths the fluctuations of the contour vibration frequency, and ensures the stability and reliability of the thermal attenuation factor, laying a solid foundation for subsequent feature fusion.

[0090] The geometric features, contour vibration frequency, and thermal attenuation factor of the molten pool have varying degrees of influence on morphological defects. Furthermore, the importance of each feature dynamically changes at different stages and under different working conditions during the welding process. Directly splicing features cannot reflect these differences. Therefore, a spatiotemporal attention fusion network is needed to adaptively learn and assign weights to different features in the spatial and temporal dimensions, highlighting the influence of key features on the molten pool morphology. The spatiotemporal attention fusion network consists of a spatial attention module and a temporal attention module. The spatial attention module generates a spatial attention weight matrix by calculating the correlation of features at different locations. In specific implementations, in the edge region of the molten pool, the weld width and weld depth among the geometric features have a greater impact on the morphology. The force module assigns higher weights to the region; the time attention module analyzes the changing trends of features over time. For sudden abnormal fluctuations in the contour vibration frequency during welding, the time attention module increases the weight of the feature corresponding to that time. The outputs of the two modules are weighted and fused to obtain the final feature weight allocation, thereby assessing the importance of different features in the spatiotemporal dimension. The spatiotemporal attention fusion network can dynamically capture key information of different features in the spatiotemporal dimension, including automatically increasing the weight of the thermal attenuation factor during the rapid solidification stage of the molten pool, and increasing the weight of the contour vibration frequency when arc instability causes molten pool oscillation. This adaptive weight allocation makes the network more focused on the core factors affecting the morphology of the molten pool.

[0091] The weighted features remain dispersed, and the potential correlations between features are not fully explored. Graph convolutional networks can model the relationships between features in a graph structure. Through information transmission between nodes and edges, the correlations between different types of features are made explicit, thus integrating them into a feature vector that comprehensively describes the morphology of the melt pool, providing rich information for predicting the probability of morphological defects. The weighted geometric features, contour vibration frequency, and thermal attenuation factor are used as node features in the graph convolutional network. Edges are constructed based on the physical relationships between features. This includes a causal relationship between melt depth and thermal attenuation factor, with strong connections established between their nodes. Contour vibration frequency and melt width are correlated due to arc stability, and corresponding connections are established. The graph convolutional network, through multi-layer convolution operations, allows the nodes to... Information between points is continuously propagated and aggregated. Each convolutional layer updates node features and fuses information from adjacent nodes. After multiple iterations, the final output set of node features becomes the integrated melt pool morphology feature, which includes each original feature and its interrelationship information. The graph convolutional network effectively mines the potential correlation between features. When it is found that the melt depth is deep and the thermal attenuation factor is large, the change in the contour vibration frequency has a greater impact on the edge defect. The integrated melt pool morphology feature is comprehensive and in-depth, and can more accurately describe the true state of the melt pool. The integrated melt pool morphology feature serves as the input of the deep belief network, providing high-quality data for it to extract the correlation pattern between the melt pool morphology and defects, thereby improving the accuracy and reliability of defect probability prediction.

[0092] Furthermore, such as Figure 4 As shown, the output sub-logic for the probability of morphological defects includes:

[0093] Deep belief networks are used to process molten pool morphology features to extract the correlation patterns between molten pool morphology and defects.

[0094] The deep belief network is adjusted through supervised learning, and the parameters of the deep belief network are optimized by using historical welding defect data as labels, so that the probability of morphological defects including undercut, hump and porosity can be output through the deep belief network.

[0095] The confidence interval of the shape defect probability is obtained based on the distribution of the shape defect probability output by the deep belief network, and the relationship weight is adjusted based on the confidence interval of the shape defect probability.

[0096] There is a complex nonlinear relationship between molten pool morphology features and defects such as undercut, humps, and porosity, which traditional methods struggle to model accurately. Deep belief networks (DBNs) possess powerful feature learning and nonlinear mapping capabilities, automatically extracting defect-related patterns from molten pool morphology features, providing a basis for predicting morphological defect probabilities. A DBN consisting of multiple stacked restricted Boltzmann machines is constructed. The integrated molten pool morphology features are input into the DBN, which is then trained layer by layer through unsupervised learning to learn hierarchical representations of the features. The bottom layer learns the basic features of the molten pool geometry, while the higher layers learn... This method extracts abstract features related to defects in one step. During training, the parameters of the deep belief network are quickly updated by a contrastive divergence algorithm to reduce computation. Through feature transformation of a multi-layer network, the molten pool morphology features are mapped to feature representations related to defects, capturing correlation patterns such as the tendency for undercutting to occur when the melt width is too large and the contour vibration is severe. The deep belief network can efficiently mine the complex hidden relationship between molten pool morphology and defects, and discover the correlation between feature combinations that are difficult to detect manually and defects, including the influence of molten pool thermal attenuation factors and specific frequency vibration combinations on porosity formation, providing more comprehensive information for morphology defect prediction.

[0097] While unsupervised learning-derived deep belief networks (DPNs) extract feature patterns, they lack a correspondence with actual defects. Supervised learning, by introducing historical welding defect data as labels, allows the DPN's learned feature patterns to match real-world defect conditions. Adjusting the DPN's parameters makes its output morphological defect probabilities more consistent with actual welding situations. A large amount of historical welding defect data, containing molten pool morphology features and corresponding defect labels, is divided into training, validation, and test sets. Defect labels include the presence of undercut, humps, and porosity. During the training phase, the molten pool morphology features from the training set are input into the DPN. The network outputs predicted values ​​of morphological defect probabilities. The difference between the predicted values ​​and defect labels is calculated using the cross-entropy loss function, and the error is backpropagated using the stochastic gradient descent algorithm to update the parameters of the deep belief network. During training, the network performance is periodically evaluated using a validation set, and hyperparameters such as the learning rate and number of iterations are adjusted to prevent overfitting. After multiple rounds of training, the parameters of the deep belief network are optimized, enabling it to accurately output morphological defect probabilities based on the input molten pool morphological features. Supervised learning closely links the deep belief network with actual welding defects, making the morphological defect probabilities output by the deep belief network more practical and accurate, providing a reliable reference for welding quality control.

[0098] The probability of morphological defects output by deep belief networks (DBNs) has a certain degree of uncertainty. Judging defect risk based solely on a single probability value is not reliable enough. By calculating confidence intervals, the fluctuation range and reliability of morphological defect probabilities can be assessed. When the confidence interval shows that the defect risk has high uncertainty or exceeds the acceptable range, the relation weights are adjusted to reassess the importance of each feature, thereby improving the prediction accuracy of the deep belief network. Through resampling, multiple sample sets are generated by sampling with replacement from the original predicted morphological defect probabilities. The mean and standard deviation of the morphological defect probabilities are calculated for each sample set. Confidence intervals are constructed based on statistical principles. Preferably, a 95% confidence interval is constructed. When the actual morphological defect probability falls within or exceeds the confidence interval, it indicates that the deep belief network has a positive view of the defect. When predictions have significant uncertainty or high risk, the confidence interval information is fed back to the spatiotemporal attention fusion network, triggering adjustments to the relation weights. The spatiotemporal attention fusion network then recalculates the weights of each feature in the spatiotemporal dimension based on the new feedback, optimizing the feature fusion method. The introduction of confidence intervals makes the assessment of morphological defect probability more reliable and comprehensive, enabling timely detection of instability in deep confidence network predictions. By triggering relation weight adjustments, the deep confidence network can dynamically adapt to changes in the welding process. When changes in welding materials lead to alterations in molten pool characteristics, the feature weights are quickly adjusted to maintain prediction accuracy. After the relation weights are adjusted, the re-integrated molten pool morphological features are input into the deep confidence network to further optimize defect probability prediction, providing a more accurate basis for welding parameter correction.

[0099] Specifically, the logic for generating the initial deviation of the welding parameters includes:

[0100] When the probability of any morphological defect exceeds the probability threshold, welding parameter correction is triggered, and the process knowledge graph is invoked to extract parameter rules associated with the morphological defect.

[0101] Based on case-based reasoning, the deviation of welding parameters is generated according to the parameter rules associated with morphological defects;

[0102] The deviation of welding parameters is used as the prior distribution of a Bayesian network. The adjustment direction and magnitude of the welding parameters are determined by the Bayesian network to generate the initial deviation of the welding parameters.

[0103] When the probability of any morphological defect exceeds a probability threshold, it indicates that the current welding parameters cannot guarantee welding quality and adjustments are necessary. The process knowledge graph stores a large amount of historical welding experience and expert knowledge, including association rules between different morphological defects and welding parameters. Calling the process knowledge graph can quickly obtain the direction and range of parameter adjustments for specific defects, avoiding blind adjustments. A real-time comparison mechanism between the probability of morphological defects and the probability threshold is established. Once the probability of any defect such as undercut, hump, or porosity exceeds the probability threshold, the welding parameter correction process is immediately triggered. Nodes related to the defect are retrieved from the process knowledge graph, and the associated parameter rules are extracted. For undercut defects, the parameter rules stored in the process knowledge graph are that when undercut occurs, the welding current can be reduced or the welding speed increased. These parameter rules are stored in the form of a semantic network, and the rule content is parsed using natural language processing technology and transformed into computer-executable instructions. Based on the parameter rule extraction from the process knowledge graph, historical experience and expert knowledge are fully utilized to provide a reliable reference for welding parameter adjustments, avoiding parameter adjustment errors caused by lack of experience and improving the efficiency and accuracy of parameter correction.

[0104] Parameter rules only provide the direction of adjustment; the specific adjustment range needs to be determined based on the actual welding conditions. Case-based reasoning, by searching historical similar cases and drawing on past successful parameter adjustment experiences, can generate welding parameter deviations that fit the current situation, avoiding the need to re-explore adjustment ranges and improving adjustment efficiency. A case library is built to store working condition information, original welding parameters, adjusted welding parameters, and adjustment effects when various morphological defects occurred in historical welding. When parameter rules related to the current defect are obtained, similar cases are retrieved from the case library. Similarity assessment comprehensively considers multiple dimensions such as material, welding process, and defect type. For stainless steel materials... For undercut defects in stainless steel, priority is given to retrieving cases where the same stainless steel material exhibits undercut defects. Based on successful parameter adjustment experience in similar cases, appropriate modifications are made according to the current actual working conditions, generating deviations in parameters such as welding current, welding voltage, and welding speed. If similar cases have solved the undercut problem by reducing the welding current, a suggested reduction in welding current deviation is generated based on the current welding equipment and material characteristics. This case-based reasoning method can quickly and accurately generate deviations in welding parameters, making full use of historical experience, reducing the trial-and-error costs of parameter adjustments, and improving the efficiency and effectiveness of welding parameter correction, especially suitable for complex and variable welding conditions.

[0105] The deviations in welding parameters generated by case-based reasoning have a certain degree of uncertainty and do not fully consider real-time data and uncertainties in the current welding process. Bayesian networks can integrate prior knowledge and real-time observation data, quantify uncertainty through probabilistic reasoning, and more accurately determine the direction and magnitude of welding parameter adjustments, generating reliable initial deviations. A Bayesian network is constructed by inputting deviations in parameters such as welding current, welding voltage, and welding speed as prior distributions, while real-time monitored data such as arc stability and the contour vibration frequency of the molten pool are input as evidence nodes. The Bayesian network then uses Bayes' theorem to perform probabilistic reasoning based on the prior distributions and evidence node information. The process involves reasoning and calculating the posterior probabilities of welding parameters under different adjustment directions and magnitudes. By comparing the magnitudes of the posterior probabilities, the adjustment scheme with the highest posterior probability is selected to determine the final adjustment direction and magnitude of the welding parameters, generating an initial deviation. A Bayesian network effectively integrates prior knowledge and real-time data to quantify the uncertainty of welding parameter adjustments, making the generated initial deviation more consistent with the current welding situation. This improves the scientific rigor and accuracy of welding parameter adjustments, reduces the risk of welding quality deterioration due to improper parameter adjustments, and uses the generated initial deviation for subsequent welding parameter corrections. Simultaneously, it provides data support for reinforcement learning to optimize weight allocation strategies, continuously improving the welding parameter correction system.

[0106] S3. Based on arc stability, molten pool profile vibration frequency and thermal attenuation factor, and initial deviation of welding parameters, the welding parameters are comprehensively corrected. After the welding parameters are corrected, the changing trend of molten pool morphology is monitored. Based on the average decrease rate of morphology defect probability within multiple consecutive sampling periods, it is determined whether the depth confidence network is optimized, and the successfully corrected welding parameters are added to the process knowledge graph.

[0107] Furthermore, the modification logic for welding parameters includes:

[0108] Arc stability, molten pool profile vibration frequency, thermal attenuation factor, and initial deviation of welding parameters are used as the state space for reinforcement learning to optimize the weight allocation strategy for welding parameter correction.

[0109] Welding parameters are adjusted by weighted summation based on a weighted allocation strategy.

[0110] After the corrected welding parameters are input into the welding equipment for adjustment, the change trend of the molten pool morphology is used to determine whether a secondary correction is triggered and to determine the correction step size of the welding parameters.

[0111] During the welding process, the arc stability, the molten pool profile vibration frequency, the heat attenuation factor, and the initial deviation of the welding parameters are interrelated. A single adjustment strategy is difficult to cope with complex working conditions. If the welding current is increased rashly based solely on the instability of the arc, it will aggravate the molten pool vibration and lead to welding defects. Therefore, these factors are jointly constructed into a reinforcement learning state space. Through intelligent algorithms, the weight relationship of each factor on the welding quality is dynamically explored to achieve adaptive optimization of welding parameter correction.

[0112] In practical implementation, the arc stability is first classified into different levels, including stable, fluctuating, and violently fluctuating. The molten pool profile vibration frequency is distinguished into low-frequency, medium-frequency, and high-frequency ranges. The thermal attenuation factor is determined to increase, stabilize, and decrease. The initial deviation is clearly defined to increase, decrease, and remain unchanged, and these are all discretized and encoded. When designing the reward function, the degree of improvement in morphology defects, changes in welding efficiency, and energy consumption constraints are comprehensively considered. For example, the rate of decrease in morphology defect probability, the magnitude of parameter adjustment, and welding energy consumption are used as reward functions. When the probability of morphology defects increases or the magnitude of parameter adjustment is too large, a negative reward is given. The agent is trained using a deep deterministic policy gradient algorithm, where the experience replay pool stores training samples, the Actor network outputs the weight allocation strategy, and the Critic network evaluates the strategy. This method, through gradient descent to update the network parameters of reinforcement learning, allows the agent to continuously experiment with different weight allocation combinations during the welding process. When the heat attenuation factor increases rapidly, it prioritizes adjusting the welding speed rather than blindly changing the current. Training continues until the reward function converges, gradually learning the optimal weight allocation strategy for different working conditions. This enables the agent to automatically adjust the weight ratio of each factor in welding parameter correction based on the actual welding situation. Specifically, in thick and thin plate welding, the correction weights of welding current, welding voltage, and welding speed are dynamically changed to address differences in material properties, avoiding the poor correction effect caused by fixed weight allocations. This significantly improves the adaptability of welding parameter correction to complex working conditions. The optimized weight allocation strategy provides a precise decision-making basis for subsequent weighted summation correction of welding parameters.

[0113] Different welding defects exhibit varying sensitivities to parameters such as welding current, welding voltage, and welding speed. Adjusting a single parameter often fails to effectively address the problem and may even introduce new defects. For instance, simply reducing the voltage to correct undercut defects can lead to insufficient molten pool temperature, resulting in incomplete fusion. Weighted summation can comprehensively consider the influence weights of each parameter, enabling coordinated optimization of multiple parameters. In practice, the weight vector obtained from reinforcement learning is calculated in correspondence with the initial deviation of the welding parameters. Taking specific welding parameters as an example, if the correction weights for welding current, welding voltage, and welding speed are high, medium, and medium, respectively, and the initial deviations are reduced current, increased voltage, and increased speed, then the final correction amount will be weighted according to the initial deviations. To prevent excessive correction from affecting the stability of the welding process, an upper limit for welding parameter adjustment is set.

[0114] By using this weighted summation correction method, the welding parameters are synergistically optimized. When correcting undercut defects, not only is the welding voltage reasonably reduced, but the welding speed is also appropriately increased. The two parameters work together to eliminate defects more efficiently and avoid the negative impact of adjusting a single parameter. The corrected welding parameters will be input into the welding equipment for adjustment, and the effect of the adjustment will be evaluated by monitoring the changing trend of the molten pool morphology to provide data support for secondary correction.

[0115] Welding is a dynamic process influenced by various factors, including material properties and environmental conditions. A single correction of welding parameters cannot completely eliminate welding defects, or the inertia of equipment response may lead to over- or under-adjustment of welding parameters. Therefore, after the corrected welding parameters are input into the welding equipment for adjustment, it is necessary to monitor the changing trend of the molten pool morphology in real time to determine whether a secondary correction is needed and to determine the appropriate correction step size. In practice, for a period of time after the welding parameters are adjusted, the molten pool morphology features are continuously acquired over multiple sampling periods, and the changes in the probability of morphological defects are analyzed. If the probability of morphological defects continues to decrease and approaches a set threshold, it indicates that the current welding... If the adjustment direction of the welding parameters is correct, the current welding parameters can be maintained. If the downward trend of the probability of morphological defects slows down or even increases, it indicates that the current correction effect is not good and a second correction needs to be triggered. The step size of the second correction is dynamically adjusted according to the stability of the trend of the probability of morphological defects. When the probability of morphological defects fluctuates greatly, a smaller correction step size is used for fine adjustment to prevent over-adjustment. When the probability of morphological defects fluctuates less, the correction step size is appropriately increased to speed up the defect elimination speed. In the process of correcting undercut defects, if the probability of morphological defects decreases slowly and fluctuates greatly after the first correction, the welding voltage will be finely adjusted and the adjustment range of the welding speed will be appropriately increased during the second correction.

[0116] This mechanism of secondary correction and step size adjustment based on changes in molten pool morphology enables closed-loop control of welding parameter correction. It can adaptively cope with various uncertainties in the welding process, including local heat input changes caused by material inhomogeneity. At the same time, the evaluation results of the changing trend of molten pool morphology characteristics also provide an important basis for the optimization judgment of the deep belief network, affecting whether the deep belief network needs to be optimized and updated.

[0117] Specifically, the optimization judgment logic of deep belief networks includes:

[0118] Calculate the average decrease rate of the probability of morphological defects within multiple consecutive sampling periods, and determine the fluctuation entropy of the probability of morphological defects within multiple consecutive sampling periods;

[0119] The optimal trigger probability is determined by combining the average decrease rate of the probability of morphological defects and the fluctuation entropy through Bayesian inference.

[0120] When the optimization trigger probability is greater than the probability threshold, the deep belief network is optimized. After optimizing the deep belief network, the rate of decrease in the probability of morphological defects is monitored to roll back and update the optimization trigger probability.

[0121] Evaluating the performance of deep belief networks solely based on the average rate of decrease in the probability of morphological defects has limitations. This is because the average rate of decrease only reflects the overall trend of defect improvement after parameter correction and cannot reflect the stability of the prediction results. Although the probability of morphological defects generally shows a downward trend over a period of time, it fluctuates dramatically, indicating that the prediction accuracy of deep belief networks is unstable under different operating conditions. This may lead to overfitting or insufficient adaptability to complex situations. Therefore, it is necessary to simultaneously calculate the fluctuation entropy of the probability of morphological defects to comprehensively evaluate the network performance from both the effectiveness and stability dimensions.

[0122] In the specific calculation, the sliding window method is used to select the probability data of morphological defects from the most recent consecutive sampling periods. The average decrease rate of the probability of morphological defects within this time period is calculated, which is the ratio of the difference between the first and last probabilities to the time interval. At the same time, the principle of information entropy is used to calculate the fluctuation entropy of the probability sequence. The larger the fluctuation entropy value, the more dispersed the probability distribution, which means that the stability of the prediction results of the deep belief network is worse. In particular, in several consecutive periods, the probability of a certain type of welding morphological defect decreased from a high value to a low value, but fluctuated frequently during the decrease. The calculated fluctuation entropy was high, which indicates that the prediction stability of the deep belief network for this defect needs to be improved. By calculating the average decrease rate and fluctuation entropy, the performance of the deep belief network can be comprehensively evaluated from multiple perspectives, avoiding the one-sided judgment caused by relying on a single indicator. When the average decrease rate is high but the fluctuation entropy is also large, it indicates that the deep belief network has the risk of overfitting and needs to be optimized and adjusted. This provides comprehensive evaluation data for determining the optimization trigger probability through Bayesian inference.

[0123] Judging whether a deep belief network needs optimization based on a fixed threshold is difficult to adapt to complex and variable welding conditions. In different welding tasks, high-speed welding requires high prediction speed from the deep belief network and allows for a certain degree of prediction fluctuation, while precision welding requires extremely high prediction accuracy and cannot tolerate too much fluctuation. Bayesian inference can combine prior knowledge with evidence such as real-time average descent rate and fluctuation entropy to quantitatively evaluate the necessity of deep belief network optimization in a probabilistic form, avoiding blind optimization of the deep belief network and improving the scientific and rational nature of decision-making.

[0124] A Bayesian network is constructed, with the average descent rate and fluctuation entropy set as evidence nodes and the optimization trigger probability as query nodes. The Bayesian network is trained using a large amount of historical welding data to determine the conditional probability distributions for different evidence combinations. When the average descent rate is less than the descent rate threshold and the fluctuation entropy is greater than the entropy threshold, the trained Bayesian network will derive a higher optimization trigger probability. During actual welding, the real-time calculated average descent rate and fluctuation entropy are input into the Bayesian network, and inference calculations are performed using Bayes' theorem to derive the optimization trigger probability of the deep belief network under the current working conditions. This Bayesian inference-based approach realizes the probabilistic and intelligent nature of deep belief network optimization decisions. It can automatically adjust the sensitivity of the deep belief network to optimization according to different welding scenarios and requirements, ensuring that the network always maintains high-precision predictive capabilities and providing a reliable basis for welding parameter correction. The result of the optimization trigger probability directly determines whether to initiate the deep belief network optimization process, effectively avoiding unnecessary waste of computational resources.

[0125] Optimizing deep belief networks (DPRNs) carries certain risks. Insufficient training data and changes in operating conditions can lead to a performance decrease after optimization, potentially even worse than expected. Overfitting can occur with small training datasets, reducing the DPRN's prediction accuracy under new conditions. Therefore, a rollback mechanism is needed to promptly restore the original DPRN parameters when optimization is ineffective. Simultaneously, continuous monitoring of network performance changes after optimization is crucial, and optimization strategies should be dynamically adjusted based on actual results. When the calculated optimization trigger probability exceeds a set threshold, the current DPRN parameters should first be saved as a backup, and then incremental learning should be used to optimize the DPRN, avoiding [further issues]. The time cost of retraining is considered. After optimization, the rate of decrease in the probability of morphological defects before and after optimization is compared. If the rate of decrease increases significantly, it indicates that the optimization is effective. At this time, the optimization trigger probability is updated so as to more accurately determine whether the deep belief network needs optimization in the future. If the rate of decrease does not change significantly or even decreases, the optimization is considered to have failed. The parameters of the previously backed-up deep belief network are immediately rolled back, and the optimization trigger threshold is updated to increase the trigger threshold for subsequent deep belief network optimizations and prevent frequent invalid optimizations. If, after a network optimization, the probability of a certain welding defect decreases instead of increasing, the original deep belief network parameters are restored, and the optimization trigger threshold is adjusted to avoid similar invalid optimizations from happening again.

[0126] This trigger-optimization and rollback update mechanism enables risk control during the deep belief network optimization process. While pursuing improved network prediction accuracy, the deep belief network may experience performance fluctuations due to data changes during the initial stages of welding material replacement or welding process adjustment. The rollback mechanism can effectively suppress these fluctuations, ensuring the smooth progress of the welding process. The optimized deep belief network will output more accurate prediction results of morphological defect probabilities, providing a more reliable reference for subsequent welding parameter correction, thus forming a closed-loop system in which welding parameter correction and network optimization mutually promote and continuously improve each other.

Claims

1. A method for correcting welding parameters based on real-time molten pool morphology feedback, characterized in that, include: The surface image of the molten pool is captured in real time, and the arc spectrum signal and infrared radiation signal of the molten pool during the welding process are acquired. The surface image of the molten pool is processed to extract the geometric features and contour vibration frequency of the molten pool. The Stark broadening of the arc spectrum signal is analyzed and the plasma electron density is measured to determine the arc stability. At the same time, the longitudinal temperature gradient of the molten pool is obtained based on the infrared radiation signal of the molten pool, and the thermal attenuation factor of the molten pool is determined in combination with the thermal diffusivity of the material. The geometric features, profile vibration frequency, and thermal attenuation factor of the molten pool are integrated into the molten pool morphology features. The molten pool morphology features are processed by a deep belief network to output the morphology defect probability. When the probability of any morphology defect is greater than the probability threshold, the welding parameters are corrected, and the process knowledge graph is called to generate the initial deviation of the welding parameters, which include welding current, welding voltage, and welding speed. Welding parameters are comprehensively corrected based on arc stability, molten pool profile vibration frequency and thermal attenuation factor, and initial deviation of welding parameters. After the welding parameters are corrected, the changing trend of molten pool morphology is monitored. The average decrease rate of morphology defect probability within multiple consecutive sampling periods is used to determine whether the depth confidence network should be optimized, and the successfully corrected welding parameters are added to the process knowledge graph.

2. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 1, characterized in that, The extraction logic for the geometric features includes: The surface image of the molten pool is preprocessed by histogram equalization, and the preprocessed surface image is segmented by a deep neural network to identify the molten pool region. The edge contour of the molten pool region is obtained by using an edge detection algorithm, and the coordinates of the edge contour are monitored to obtain the molten pool width. Sinusoidal fringe structured light is projected onto the surface of the molten pool region to capture deformed fringe images. The deformed fringe images are processed by a phase unwrapping algorithm to obtain the degree of fringe distortion, and the surface coordinates of the molten pool region are obtained to extract the vertical distance from the bottom of the molten pool to the surface as the molten pool depth.

3. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 2, characterized in that, The logic for extracting the contour vibration frequency includes: The surface images of the molten pool are continuously captured to form an image sequence. The image sequence is stacked in the time dimension to generate a spatiotemporal image cube to reflect the changes in the edge contour. The dynamic time warping algorithm is used to match and align the changes in the edge contour to obtain the contour change data of the molten pool. The contour variation data of the molten pool is decomposed into multiple intrinsic mode functions by empirical mode decomposition method, and the spectral analysis of the decomposed intrinsic mode functions is performed to extract the contour vibration frequency of the molten pool.

4. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 3, characterized in that, The logic for determining arc stability includes: The arc spectral signal during the welding process was acquired by a spectrometer, and the arc spectral signal was decomposed into multiple scales by wavelet packet decomposition to extract spectral energy features in different frequency bands. The spectral energy characteristics of Stark broadening are automatically identified by convolutional neural networks, and the plasma electron density is measured based on the full width at half maximum (FWHM) of Stark broadening. Configure the fluctuation threshold range of electron density and compare it with the fluctuation of plasma electron density in multiple monitoring sampling periods to determine the stability of the arc.

5. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 4, characterized in that, The logic for determining the thermal decay factor includes: The infrared radiation signal from the molten pool is spatially filtered, and the energy distribution characteristics of the infrared radiation signal at different spatial frequencies are extracted based on the frequency domain analysis method of Fourier transform. The energy distribution characteristics of infrared radiation signals at different spatial frequencies are simulated by finite element analysis to inversely deduce the longitudinal temperature gradient of the molten pool. Based on Fourier's law, the thermal decay factor of the molten pool is determined according to the longitudinal temperature gradient of the molten pool and the thermal diffusivity of the material.

6. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 5, characterized in that, The logic for generating the initial deviation of the welding parameters includes: When the probability of any morphological defect exceeds the probability threshold, welding parameter correction is triggered, and the process knowledge graph is invoked to extract parameter rules associated with the morphological defect. Based on case-based reasoning, the deviation of welding parameters is generated according to the parameter rules associated with morphological defects; The deviation of welding parameters is used as the prior distribution of a Bayesian network. The adjustment direction and magnitude of the welding parameters are determined by the Bayesian network to generate the initial deviation of the welding parameters.

7. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 6, characterized in that, The integration sub-logic of the molten pool morphology features includes: Data preprocessing is performed on the geometric characteristics, profile vibration frequency, and thermal attenuation factor of the molten pool; A spatiotemporal attention fusion network is constructed to determine the relational weights of different features to the morphology of the melt pool. The spatiotemporal attention fusion network includes a spatial attention module and a temporal attention module. The features, weighted by relational weights, are input into a graph convolutional network to construct a feature association graph. The feature association graph is then continuously processed through graph convolution operations to integrate it into the molten pool morphology features.

8. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 7, characterized in that, The output sub-logic for the probability of morphological defects includes: Deep belief networks are used to process molten pool morphology features to extract the correlation patterns between molten pool morphology and defects. The deep belief network is adjusted through supervised learning, and the parameters of the deep belief network are optimized by using historical welding defect data as labels. The deep belief network outputs the probability of morphological defects including undercut, hump and porosity. The confidence interval of the shape defect probability is obtained based on the distribution of the shape defect probability output by the deep belief network, and the relationship weight is adjusted based on the confidence interval of the shape defect probability.

9. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 8, characterized in that, The optimization judgment logic of the deep belief network includes: Calculate the average decrease rate of the probability of morphological defects within multiple consecutive sampling periods, and determine the fluctuation entropy of the probability of morphological defects within multiple consecutive sampling periods; The optimal trigger probability is determined by combining the average decrease rate of the probability of morphological defects and the fluctuation entropy through Bayesian inference. When the optimization trigger probability is greater than the probability threshold, the deep belief network is optimized. After optimizing the deep belief network, the rate of decrease in the probability of morphological defects is monitored to roll back and update the optimization trigger probability.

10. The welding parameter correction method with real-time molten pool morphology feedback as described in claim 9, characterized in that, The correction logic for the welding parameters includes: Arc stability, molten pool profile vibration frequency, thermal attenuation factor, and initial deviation of welding parameters are used as the state space for reinforcement learning to optimize the weight allocation strategy for welding parameter correction. Welding parameters are adjusted by weighted summation based on a weighted allocation strategy. After the corrected welding parameters are input into the welding equipment for adjustment, the change trend of the molten pool morphology is used to determine whether a secondary correction is triggered and to determine the correction step size of the welding parameters.

Images