Laser-electric arc hybrid welding method and system

Abstract

The invention relates to the technical field of laser-electric arc hybrid welding, and discloses a laser-electric arc hybrid welding method and system. According to the method, molten pool vision, plasma spectrum, electric signals and temperature field data are collected in real time through multi-modal sensing, and after depth features are extracted and fused, a quality evaluation model based on Stacking ensemble learning is used for recognizing weld defects and confidence in real time; and when the confidence coefficient is insufficient or a defect tendency is detected, triggering a parameter decision optimized by multiple targets such as fusion depth, strength and the like, generating an adjustment instruction set, and executing dynamic fine adjustment after verification of a process knowledge base. The system correspondingly comprises a pose deviation compensation module, an online quality evaluation module, an optimization decision module, a strategy verification module and a parameter execution module. Intelligent sensing and self-adaptive accurate control in the welding process are achieved, and the welding quality stability and the defect suppression capacity are effectively improved.

Description

Technical Field

[0001] This application relates to the field of welding technology, and in particular to a laser-arc hybrid welding method and system. Background Technology

[0002] Laser-arc hybrid welding technology integrates the advantages of high energy density and deep penetration of laser welding with the strong bridging ability and wide process adaptability of arc welding, and has become a key technology in advanced manufacturing. To improve the intelligence level and quality consistency of the welding process, existing technologies are committed to introducing multi-sensor information fusion and artificial intelligence algorithms to build a fully closed-loop adaptive control system of "perception-evaluation-decision-adjustment".

[0003] Chinese Patent, Publication No. CN115971663B, Publication Date: July 11, 2025, discloses a laser-arc hybrid welding method and a laser-arc hybrid welding system. The laser-arc hybrid welding method includes the following steps: operating an arc welding machine under a set pulse current; acquiring the real-time welding current during the welding process; and controlling the power of the laser beam output from the laser according to the real-time welding current so that the laser beam power is at a first set value before the molten droplet falls into the molten pool, and at a second set value after the molten droplet falls into the molten pool, wherein the first set value is less than the second set value. The laser is in a low-power state during the formation and detachment of the molten droplet, which reduces the impact on the droplet detachment process and allows the droplet to fall stably into the molten pool. The laser is in a high-power state after the droplet falls into the molten pool, increasing heat input and facilitating the penetration of the droplet into the molten pool, thus increasing the weld depth. Synchronizing the laser and pulsed arc cycles achieves coordinated control, improving the energy utilization rate of the laser and enhancing welding quality.

[0004] The shortcomings of the above technical solutions are as follows: 1. They may not fully utilize the features of multi-source information in the welding process, and the depth and standardization of feature extraction and fusion need to be strengthened. This limits the accuracy and generalization ability of the quality assessment model in early identification of complex and hidden defects. 2. The structure of its quality assessment model may be relatively simple. When facing the highly nonlinear and multimodal process signals of laser-arc hybrid welding, there is still room for improvement in the accuracy and stability of identifying various defects (especially coupled defects). Summary of the Invention

[0005] The purpose of this application is to provide a laser-arc hybrid welding method and system, which can solve the problem of insufficient control over welding quality stability caused by insufficient utilization of process information and limited performance of quality assessment models in the prior art.

[0006] To achieve the above objectives, this application adopts the following technical solution: This application provides a laser-arc hybrid welding method, which includes the following steps: S101, acquire the three-dimensional point cloud data and contact measurement data of the workpiece to be welded, perform fusion processing on the three-dimensional point cloud data and the contact measurement data to generate a real-time positional deviation signal between the welding gun and the center line of the weld, and drive the laser-arc composite welding head to perform initial positioning according to the real-time positional deviation signal so that the positional deviation is not greater than the preset positional deviation threshold. S102, the laser-arc composite welding head is started to perform welding. During the welding process, the visual image of the molten pool, the composite plasma spectrum, the welding current and voltage waveforms, and the temperature field distribution image of the weld area are acquired in real time and synchronously. The acquired raw signals are preprocessed by noise reduction and synchronous alignment to obtain the preprocessed multimodal signal. S103, Deep feature extraction is performed on the preprocessed multimodal signal obtained in step S102, and standardization and feature-level fusion are performed to form a high-dimensional fusion feature vector for evaluating the stability and defect tendency of the welding process. S104, The high-dimensional fusion feature vector constructed in step S103 is input into the pre-trained quality assessment model. The quality assessment model is built based on the Stacking ensemble learning framework. The base learners of the Stacking ensemble learning framework include random forest, support vector machine and K-nearest neighbor learner. The meta-learner of the Stacking ensemble learning framework is a deep forest model based on attention mechanism, which is used to weight and integrate the outputs of the base learners. The quality assessment model outputs the identification results and confidence scores of defect states including weld formation, porosity, incomplete penetration, undercut and hump in real time. S105, when the confidence level of the identification result output in step S104 is lower than the preset confidence threshold or a specific defect tendency is detected, a welding parameter optimization decision is triggered: First, based on the visual image of the molten pool, the relative distance between the laser beam and the arc is monitored in real time and controlled in a closed loop to stabilize the relative distance within the preset optimal coupling range; then, with the current quality status, workpiece material, bevel size, and the real-time defect confidence level output by the quality assessment model as inputs, and with weld penetration depth, tensile strength, and intergranular corrosion resistance as multiple objectives, online multi-objective optimization is performed under process constraints to generate a comprehensive parameter adjustment instruction set including laser power, arc current, welding speed, and defocusing amount in real time; S106, the comprehensive parameter adjustment instruction set generated in step S105 is matched and verified with the built-in welding process knowledge base, and the final executable fine-tuning strategy is generated based on the matching and verification results. The executable fine-tuning strategy includes at least a parameter gradient curve within the length of the next welding unit or an instant compensation pulse parameter for a specific defect. S107, Receive and execute the executable fine-tuning strategy generated in step S106, and dynamically adjust the output parameters of the laser-arc composite welding head; S108, Repeat steps S102 to S107 until the entire weld seam is completed.

[0007] As a preferred technical solution, step S102 includes: Visual images of the molten pool and keyhole area are acquired using a combination of a high-speed CMOS camera and a narrow-band filter. The emission spectra of the laser-arc composite plasma in the ultraviolet to near-infrared bands are simultaneously acquired using a fiber optic spectrometer coupled with a six-channel high-speed beam splitter, serving as the composite plasma spectrum. Instantaneous waveforms of welding current and voltage are acquired using a Hannover welding quality analyzer, serving as the welding current and voltage waveforms, and the short-time Fourier transform spectrum of these instantaneous waveforms is calculated. Temperature field distribution images of the weld and its heat-affected zone are acquired using a mid-wave infrared thermal imager, and real-time temperature calibration is performed based on colorimetric thermometry. The acquired molten pool visual image is processed using a hybrid algorithm combining adaptive weighted median filtering and wavelet thresholding to remove spatter and plasma scintillation noise. The composite plasma spectrum is smoothed using a moving window Savitzky-Golay filter, and baseline correction is performed using iterative polynomial fitting to eliminate continuous background radiation. The welding current and voltage waveforms are first processed using a power frequency interference cancellation technique based on linear prediction, and then processed by a finite-length unit impulse response low-pass filter designed with a Kaiser window to separate and smooth high-frequency noise. The temperature field distribution image is processed using a two-point non-uniformity correction method based on a reference blackbody and a bad pixel compensation algorithm based on neighborhood weighted interpolation. Synchronous alignment preprocessing forces all acquisition devices to start synchronously at the welding start moment by using a global hardware trigger signal based on the output of a precision programmable delay pulse generator. During continuous acquisition, each data frame is marked with a high-precision, low-drift absolute timestamp provided by a rubidium atomic clock module. Based on the absolute timestamp, a dynamic time warping algorithm is used to perform time series alignment and interpolation compensation on the cross-modal data, generating time-series multimodal signal data blocks that are strictly synchronized in the time domain and aligned in the feature dimension, which serve as the preprocessed multimodal signal.

[0008] As a preferred technical solution, step S103 includes: A dual-branch convolutional neural network based on an attention mechanism is used to process the molten pool visual image in the preprocessed multimodal signal obtained in step S102. One branch extracts the static morphological features of the molten pool, including area, aspect ratio, contour roundness, and trailing angle. The other branch calculates the dynamic motion vector field of the molten pool region using optical flow method, extracts the main frequency, amplitude, and flow velocity distribution features of the keyhole oscillation, and the flow velocity distribution features at the tail of the molten pool. The features of the two branches are then weighted and fused in the channel dimension to form a molten pool depth visual feature vector. For the welding current and voltage waveforms in the preprocessed multimodal signal obtained in step S102, statistical features including probability density distribution skewness and kurtosis, short-circuit transition frequency, coefficient of variation, and waveform factors are extracted in the time domain. In the frequency domain, wavelet packet transform decomposition is performed to calculate the Shannon entropy of the energy proportion of each sub-band. Waveform nonlinear dynamic features based on recursive quantitative analysis are constructed. The waveform nonlinear dynamic features include recursion rate and determinism, which together constitute a multidimensional feature set of the electrical signal. For the composite plasma spectrum in the preprocessed multimodal signal obtained in step S102, the characteristic spectral lines of argon, iron, silicon, manganese and chromium at specific wavelengths are identified, the net intensity is calculated, and the electron temperature and electron density of the plasma are inverted in real time. The intensity ratio, intensity ratio change rate, electron temperature and electron density of each element are used as key spectral physical features. For the temperature field distribution image in the preprocessed multimodal signal obtained in step S102, extract the isotherm of the molten pool profile and the width gradient of the heat-affected zone, and calculate the thermal cycling curve at a specific location based on the time series image, and then extract thermodynamic features including the residence time above the phase change point, t8 / 5 cooling time and peak temperature. The visual feature vector of the molten pool depth, the multi-dimensional feature set of the electrical signal, the key spectral physical features and the thermodynamic features are respectively normalized by Z-score based on a sliding window to eliminate the influence of dimensions and short-term fluctuations. The maximum correlation minimum redundancy algorithm based on mutual information is used to filter all standardized features, remove redundant features, and then the filtered features are concatenated according to a preset time window. A learnable feature weight vector is introduced to assign appropriate weights to features of different modalities, and finally a high-dimensional fusion feature vector is formed.

[0009] As a preferred technical solution, the operating mechanism of the quality assessment model built based on the Stacking ensemble learning framework in step S104 includes: The high-dimensional fused feature vector constructed in step S103 is input in parallel to the three base learners: the random forest, the support vector machine, and the K-nearest neighbors. The random forest learner is constructed using an extreme random tree algorithm to evaluate feature importance and output a first preliminary prediction probability for various defects. The support vector machine learner uses a Gaussian radial basis kernel function to construct a classification hyperplane in high-dimensional space and output a second preliminary prediction probability. The K-nearest neighbors learner uses a hybrid metric that combines Euclidean distance and dynamic time curvature distance to perform similarity matching in the historical sample database and output a third preliminary prediction probability. The vector composed of the first preliminary prediction probability, the second preliminary prediction probability, and the third preliminary prediction probability output by the three base learners is concatenated with the preset key physical features in the original fused feature vector formed in step S103 to form a meta-feature vector, which is then input into the LightGBM meta-learner whose hyperparameters have been optimized by the particle swarm algorithm. The LightGBM meta-learner integrates an attention weight allocation module. The attention weight allocation module dynamically calculates the decision reliability weight of each base learner based on the entropy value of the output probability vector of each base learner, and performs weighted fusion of the first preliminary prediction probability, the second preliminary prediction probability, and the third preliminary prediction probability. The LightGBM model then makes the final decision based on the weighted meta-features and outputs the defect category label and confidence score.

[0010] As a preferred technical solution, the welding parameter optimization decision in step S105 includes: A first-layer CNN regression model was constructed and trained to monitor the relative distance between the laser beam and the electric arc. The construction and training process involved conducting calibration experiments, setting a series of known laser-arc relative distances and simultaneously acquiring corresponding visual images of the molten pool, constructing an image-distance pairing dataset, and using the image-distance pairing dataset to train a deep convolutional neural network containing a residual structure. The deep convolutional neural network takes a single frame of molten pool image as input and outputs an estimated value of the relative distance. The trained CNN regression model is deployed on an online system to process the visual image of the molten pool in real time and output a distance estimate. The distance estimate is compared with the preset optimal coupling interval to generate a distance deviation signal. The distance deviation signal is input to a fuzzy adaptive PID controller, which outputs an adjustment command for the pose of the laser head or arc welding gun to achieve closed-loop stable control of the relative distance. A second-layer PSO-BP neural network optimization model is constructed and invoked to generate the comprehensive parameter adjustment instruction set. The construction process involves collecting a historical welding process database, which contains multiple quality target results of process parameter combinations and working conditions with corresponding weld penetration, tensile strength, and intergranular corrosion resistance. A multilayer perceptron neural network is trained using the backpropagation algorithm to establish an initial nonlinear mapping model from input parameters to output quality indicators. An improved particle swarm optimization algorithm is used to globally optimize the connection weights and biases of the initial nonlinear mapping model. The improvements include the introduction of a linear decreasing strategy for inertial weights and a Pareto optimal solution selection mechanism based on crowding distance, thereby obtaining an optimized PSO-BP neural network model. When invoked online, the real-time defect confidence and category, current workpiece material and bevel size output by the quality assessment model are used as partial inputs. Combined with the current welding parameters, these are input to the optimized PSO-BP neural network model. The PSO-BP neural network model aims to maximize the comprehensive quality score. It performs forward calculation and rapid optimization within the process constraint boundary and outputs the comprehensive parameter adjustment instruction set, which includes laser power, arc current, welding speed and defocusing amount, in real time.

[0011] As a preferred technical solution, step S106 includes: Receive the integrated parameter adjustment instruction set generated in step S105, the integrated parameter adjustment instruction set including the adjustment amount of laser power, arc current, welding speed and defocus amount; The comprehensive parameter adjustment instruction set is matched with the built-in welding process knowledge base, which is stored in the form of a relational database. Each record contains material combination, bevel type, plate thickness, historical quality rating, and corresponding set of successful process parameters. The matching process first performs a preliminary screening of the welding process knowledge base records based on the current workpiece material, joint type, and plate thickness. Then, the weighted Euclidean distance between the comprehensive parameter adjustment instruction set and each candidate parameter set after screening is calculated in the adjustment space. Finally, an algorithm based on case reasoning is called to evaluate the applicability of the comprehensive parameter adjustment instruction set by combining the final quality scores of similar cases. The quality status identification result and the real-time defect confidence level output in step S104 are logically compared with the expert rules stored in the welding process knowledge base. The expert rules include prioritizing fine-tuning the defocusing amount and the protective gas flow rate when there is a tendency for porosity, and prioritizing increasing the laser power and reducing the welding speed when there is a tendency for incomplete penetration. The matching and verification results are input into an optimization decision engine, which generates a final executable fine-tuning strategy based on fuzzy reasoning and constraint satisfaction algorithms. The executable fine-tuning strategy can take the form of a parameter gradient curve designed to adapt to the gradual change in plate thickness or gap, in which the four key parameters of laser power, arc current, welding speed and defocusing amount change continuously at a specific slope in the next welding length, or a coordinated compensation pulse parameter package of laser power and arc current for a period of time triggered to suppress specific defects in real time. The executable fine-tuning strategy is encoded as a sequence of execution instructions that can be directly issued to the welding power source and motion controller.

[0012] As a preferred technical solution, step S107 includes: The main controller receives and parses the executable fine-tuning strategy generated in step S106; For the parameter gradient curve strategy in the executable fine-tuning strategy, the main controller discretizes the parameter gradient curve into a series of time-parameter value pairs with serial numbers according to the spatial position based on the welding speed. Through the high-speed fieldbus based on EtherCAT, the sub-instructions and the corresponding target position information are synchronously and in real time sent to the laser power supply, the arc welding power supply and the motion servo driver, driving the laser power, the arc current, the welding speed and the defocusing amount to change smoothly along a predetermined trajectory within the preset welding length. For the instantaneous compensation pulse parameter strategy in the executable fine-tuning strategy, within a millisecond delay after receiving the defect trigger signal, the main controller sends a pulse trigger command and parameter packet with nanosecond-level precision timestamps to the fast digital interface of the laser and arc power supply to ensure that the energy output of the laser and arc power supply is coordinated and adjusted according to the preset waveform in a short time. During the execution process, the main controller collects and monitors the feedback status words of the welding power supply and servo driver in real time, including the actual output power, current, speed and position, and compares the actual values ​​with the expected values ​​in the command sequence in a loop to generate the execution status deviation. When the deviation of the execution status exceeds the preset tolerance, an online safety protection strategy including parameter rollback and process interruption is immediately triggered to ensure the accuracy of parameter adjustment and the safety of the process.

[0013] As a preferred technical solution, step S108 includes: After a welding unit is completed, the main controller packages the original multimodal signals, extracted fusion feature vectors, quality assessment results, executed fine-tuning strategy instruction set, and final macroscopic quality sampling results of the welding unit collected during steps S102 to S107 into a complete process-result data packet, which is then encrypted and stored and marked with a unique spatiotemporal identifier. The main controller calls the built-in self-evaluation module to compare and analyze the actual process data and quality results of the welding unit with the theoretical process and results predicted by the welding process digital twin model based on the initial parameters, and generates a self-evaluation report on sensor effectiveness, feature extraction robustness and parameter adjustment strategy accuracy. Based on the self-assessment report, during the welding interval, the incremental learning of the model and the dynamic update process of the knowledge base are initiated. The incremental learning process uses the newly added process-result data package to incrementally train and fine-tune the meta-learner of the Stacking ensemble learning quality assessment model and the second-layer PSO-BP neural network optimization model. The dynamic update process of the knowledge base abstracts the verified successful fine-tuning strategy cases and corresponding working condition features into new expert rules, which are then integrated and updated into the welding process knowledge base. In step S105, the incrementally optimized PSO-BP neural network optimization model performs online optimization based on the updated weight parameters. In step S106, the optimization decision engine calls the dynamically expanded welding process knowledge base for matching and verification.

[0014] This application also provides a laser-arc hybrid welding system, the system comprising: The pose deviation compensation module is used to acquire measurement data of the workpiece to be welded, generate real-time pose deviation signals between the welding torch and the weld centerline, and drive the welding head to perform initial positioning. The online quality assessment module is used to collect multimodal sensor signals in real time during the welding process and preprocess them to extract and fuse deep features. The resulting high-dimensional fused feature vector is then input into a pre-trained quality assessment model built on the Stacking ensemble learning framework to output the identification results and confidence levels of welding defects in real time. The optimization decision module is used to trigger welding parameter optimization decisions when the confidence level of the identification result is lower than the threshold or a specific defect tendency is detected. First, the relative distance between the laser and the electric arc is controlled in a closed loop, and then a comprehensive parameter adjustment instruction set is generated online based on a multi-objective optimization algorithm. The strategy verification module is used to match and verify the comprehensive parameter adjustment instruction set with the welding process knowledge base to generate the final executable fine-tuning strategy. The parameter execution module is used to receive and execute the executable fine-tuning strategy to dynamically adjust the output parameters of the laser-arc hybrid welding head. The welding control module is used to control the online quality assessment module, optimization decision module, strategy verification module and parameter execution module to run cyclically during the welding process until the welding is completed.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: This application achieves high-precision initial positioning by fusing 3D point cloud and contact measurement, and simultaneously acquires multimodal information such as molten pool images, spectra, electrical signals, and temperature fields. Through deep feature extraction and standardized fusion, a high-dimensional feature vector that comprehensively characterizes the welding state is constructed. Furthermore, a Stacking ensemble model, using random forests and support vector machines as base learners and deep forests with attention mechanisms as meta-learners, is employed as the core of quality assessment. This significantly improves the accuracy of identifying various defects such as weld formation, porosity, and incomplete penetration, as well as its early warning capability. When a quality risk is detected, the system not only stabilizes the laser-arc coupling distance but also optimizes multiple properties such as weld penetration and strength online in real time, generating comprehensive parameter adjustment instructions. After verification by the process knowledge base, executable fine-tuning strategies (such as parameter gradient curves and compensation pulses) are formed, ultimately achieving dynamic closed-loop optimization of welding parameters. The entire method forms a complete technical chain of "precise perception - intelligent assessment - real-time decision-making - closed-loop execution", which significantly enhances the adaptability and robustness of the welding process, thereby effectively ensuring the uniformity of weld formation and the stability and reliability of the joint's overall performance. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in the specific embodiments of the present application will be clearly and completely described below in conjunction with the embodiments of the present application.

[0017] This application provides a laser-arc hybrid welding method, which includes the following steps: S101: Acquire the three-dimensional point cloud data and contact measurement data of the workpiece to be welded, perform fusion processing on the three-dimensional point cloud data and contact measurement data to generate a real-time pose deviation signal between the welding torch and the center line of the weld, and drive the laser-arc composite welding head to perform initial positioning based on the real-time pose deviation signal so that the pose deviation is not greater than the preset pose deviation threshold.

[0018] S102, the laser-arc hybrid welding head is started for welding. During the welding process, the visual image of the molten pool, the composite plasma spectrum, the welding current and voltage waveforms, and the temperature field distribution image of the weld area are acquired in real time. The acquired raw signals are preprocessed by noise reduction and synchronous alignment to obtain the preprocessed multimodal signal.

[0019] S103, perform deep feature extraction on the preprocessed multimodal signal obtained in step S102, and perform standardization and feature-level fusion to form a high-dimensional fused feature vector for evaluating the stability and defect tendency of the welding process.

[0020] In step S104, the high-dimensional fused feature vector constructed in step S103 is input into the pre-trained quality assessment model. This model is built upon the Stacking ensemble learning framework. The base learners of the Stacking ensemble learning framework include random forest, support vector machine, and K-nearest neighbor learner. The meta-learner of the Stacking ensemble learning framework is a deep forest model based on an attention mechanism, used to weight and integrate the outputs of the base learners. The quality assessment model outputs in real-time the identification results and confidence scores for defect states, including weld formation, porosity, incomplete penetration, undercut, and hump.

[0021] S105, when the confidence level of the identification result output in step S104 is lower than the preset confidence threshold or a specific defect tendency is detected, a welding parameter optimization decision is triggered: First, based on the visual image of the molten pool, the relative distance between the laser beam and the arc is monitored in real time and controlled in a closed loop to stabilize the relative distance within the preset optimal coupling range. Then, using the current quality status, workpiece material, bevel size, and the real-time defect confidence level output by the quality assessment model as inputs, and with weld penetration depth, tensile strength, and resistance to intergranular corrosion as multiple objectives, online multi-objective optimization is performed under process constraints, generating a comprehensive parameter adjustment instruction set in real time that includes laser power, arc current, welding speed, and defocusing amount.

[0022] S106, the comprehensive parameter adjustment instruction set generated in step S105 is matched and verified with the built-in welding process knowledge base. Based on the matching and verification results, the final executable fine-tuning strategy is generated. The executable fine-tuning strategy includes at least the parameter gradient curve within the length of the next welding unit or the instant compensation pulse parameter for a specific defect.

[0023] S107: Receive and execute the executable fine-tuning strategy generated in step S106 to dynamically adjust the output parameters of the laser-arc hybrid welding head.

[0024] S108, Repeat steps S102 to S107 until the entire weld seam is completed.

[0025] Specifically, a line laser scanner mounted on the side of the welding head rapidly scans the workpiece's welding area to obtain three-dimensional point cloud data characterizing its surface contour and bevel shape. Simultaneously, a contact probe at the end of the welding torch is controlled to precisely measure the three-dimensional coordinates of key feature points such as the bevel root and sidewalls in a point-to-point manner, serving as contact measurement data. Subsequently, an iterative nearest-point algorithm and coordinate system registration technology are applied, using the high-precision contact measurement points as a reference, to calibrate and fuse the broader but potentially noisy point cloud data, generating a precise and unified three-dimensional weld model.

[0026] Based on this fusion model, the system calculates the centerline trajectory of the weld and its normal vector at the welding point in real time using a feature extraction algorithm. Simultaneously, the actual position and orientation of the welding torch tool center point (TCP) are acquired in real time using sensors built into the welding head. By comparing the position and direction vectors of the TCP with those of the corresponding points on the weld centerline, the system calculates the pose deviation in real time, including the position deviation in three-dimensional space. and angular deviations around each axis These components together constitute the pose deviation signal. This signal is sent to the robot's motion controller in real time. The controller generates adjustment commands for the servo motors of each joint based on the deviation amount, driving the laser-arc hybrid welding head to adjust its pose until all deviation components are less than the preset threshold, thus completing the initial positioning.

[0027] This application achieves high-precision and robust reconstruction of weld trajectories by integrating data from non-contact surface scanning and contact point measurement, overcoming the limitations of single sensors under complex conditions such as reflection and beveling errors. This step proactively eliminates initial pose deviations caused by workpiece assembly errors, thermal deformation, or inaccurate offline programming before welding begins, laying a precise benchmark for the subsequent high-quality welding process. It avoids defects such as weld deviation and incomplete penetration directly caused by improper starting positions, improving the adaptability and process consistency of the entire welding system.

[0028] Furthermore, step S102 includes: Visual images of the molten pool and keyhole area were acquired using a combination of a high-speed CMOS camera and narrow-band filters. A six-channel high-speed beam splitter coupled with a fiber optic spectrometer was used to simultaneously acquire the emission spectra of the laser-arc composite plasma in the ultraviolet to near-infrared bands, serving as the composite plasma spectrum. Instantaneous waveforms of welding current and voltage were acquired using a Hannover welding quality analyzer, serving as the welding current and voltage waveforms, and the short-time Fourier transform spectra of these instantaneous waveforms were calculated. Temperature field distribution images of the weld and its heat-affected zone were acquired using a mid-wave infrared thermal imager, and real-time temperature calibration was performed based on colorimetric thermometry.

[0029] For the acquired visual images of the molten pool, a hybrid algorithm combining adaptive weighted median filtering and wavelet thresholding was employed to remove spatter and plasma scintillation noise. The composite plasma spectrum was smoothed using a moving window Savitzky-Golay filter, and baseline correction was performed using iterative polynomial fitting to eliminate continuous background radiation. The welding current and voltage waveforms were first processed using a power frequency interference cancellation technique based on linear prediction, and then processed by a finite-length unit impulse response low-pass filter designed with a Kaiser window to separate and smooth high-frequency noise. For the temperature field distribution image, a two-point non-uniformity correction method based on a reference blackbody and a bad pixel compensation algorithm using neighborhood weighted interpolation were employed.

[0030] Synchronous alignment preprocessing forces all acquisition devices to start synchronously at the start of welding by using a global hardware trigger signal based on the output of a precision programmable delay pulse generator. During continuous acquisition, each data frame is marked with a high-precision, low-drift absolute timestamp provided by a rubidium atomic clock module. Based on the absolute timestamp, a dynamic time warping algorithm is used to align and interpolate the cross-modal data over time, generating time-series multimodal signal data blocks that are strictly synchronized in the time domain and aligned in the feature dimensions, serving as the preprocessed multimodal signal.

[0031] Specifically, this step constructs a multi-source heterogeneous sensing system and a high-fidelity preprocessing pipeline. At the data acquisition end, optical, electrical, and thermal multi-physics sensors operate simultaneously: a high-speed CMOS camera equipped with a narrow-band filter of a specific wavelength (e.g., 808 nm ± 5 nm) effectively suppresses arc interference and captures clear images of the molten pool and keyhole; a fiber optic spectrometer, in conjunction with a high-speed spectrometer module, achieves high-speed parallel acquisition of composite plasma emission spectra from the ultraviolet to near-infrared bands; a Hannover analyzer captures instantaneous details of current and voltage at a sampling rate up to 100 kHz and converts them to the frequency domain using a short-time Fourier transform (STFT), the mathematical expression of which is: ; in For window functions, The frame shift is used to reveal the spectral characteristics of the electrical signal; the mid-wave infrared thermal imager, based on the principle of colorimetric thermometry, inverts the true temperature by comparing the radiation intensity of two bands and generates a temperature field image.

[0032] The acquired raw signals are then fed into a targeted noise reduction channel. For molten pool images filled with splash and plasma scintillation noise, adaptive weighted median filtering (with weights dynamically adjusted based on local gray-level variance) and wavelet soft thresholding are used for noise reduction (threshold set to a specific value). ,in For noise standard deviation estimation, The processing involves a hybrid algorithm (for signal length). For spectral signals, a moving-window Savitzky-Golay filter is first used for smoothing to preserve the peak shape, followed by baseline drift subtraction through iterative polynomial fitting. For electrical signals, a power frequency (50 / 60 Hz) sinusoidal template is first constructed for linear prediction and cancellation, followed by high-frequency noise removal using an FIR low-pass filter designed with a Kaiser window (cutoff frequency determined by the process). For thermal images, a two-point correction is performed using a reference blackbody, and neighbor-weighted interpolation is used to compensate for bad pixels.

[0033] To achieve strict synchronization of cross-modal data, at the moment welding begins, a precision programmable delay pulse generator sends a global hardware trigger signal to all acquisition devices, forcing synchronization to start. During continuous acquisition, each data frame is marked with an absolute timestamp (with microsecond-level accuracy) provided by a highly stable rubidium atomic clock. Subsequently, based on these timestamps, a Dynamic Time Warping (DTW) algorithm is used to perform nonlinear alignment and interpolation on the modal data sequences at different sampling rates. The core of this algorithm is to find an optimal curved path. This results in the cost of normalization: Minimum, of which For sequence points and The distance metric between them ultimately generates multimodal time-series data blocks that are strictly aligned in the time domain and correspond to the feature dimensions.

[0034] In this application, a sensor array with complementary physical principles is configured to achieve holographic perception of the "optical-electrical-thermal" multi-physics fields during the welding process, providing an unprecedented information dimension for subsequent analysis. Advanced noise reduction algorithms are employed to address the differences in noise characteristics of various modal signals, significantly improving the signal-to-noise ratio and fidelity. Furthermore, a three-level synchronization mechanism—"hardware global triggering + high-precision time stamping + software dynamic normalization"—completely resolves the time mismatch problem of multi-source asynchronous data, ensuring that the data used for subsequent feature extraction and fusion have a true causal correspondence in time and space. This lays a solid and consistent data foundation for constructing a highly reliable digital twin of the welding process and achieving accurate real-time quality assessment and closed-loop control.

[0035] Furthermore, step S103 includes: A dual-branch convolutional neural network based on an attention mechanism is used to process the molten pool visual image in the preprocessed multimodal signal obtained in step S102. One branch extracts the static morphological features of the molten pool, including area, aspect ratio, contour roundness, and trailing angle. The other branch calculates the dynamic motion vector field of the molten pool region through optical flow method, extracts the main frequency, amplitude, and flow velocity distribution features of the keyhole oscillation and the tail of the molten pool, and then weights and fuses the features of the two branches in the channel dimension to form the molten pool depth visual feature vector.

[0036] For the welding current and voltage waveforms in the preprocessed multimodal signal obtained in step S102, statistical features including probability density distribution skewness and kurtosis, short-circuit transition frequency, coefficient of variation, and waveform factors are extracted in the time domain. In the frequency domain, wavelet packet transform decomposition is performed to calculate the Shannon entropy of the energy proportion of each sub-band. Waveform nonlinear dynamic features based on recursive quantitative analysis are constructed. The waveform nonlinear dynamic features include recursion rate and determinism, which together constitute a multidimensional feature set of the electrical signal.

[0037] For the composite plasma spectrum in the preprocessed multimodal signal obtained in step S102, the characteristic spectral lines of argon, iron, silicon, manganese and chromium at specific wavelengths are identified, the net intensity is calculated, and the electron temperature and electron density of the plasma are inverted in real time. The intensity ratio, intensity ratio change rate, electron temperature and electron density of each element are used as key spectral physical features.

[0038] The temperature field distribution image in the preprocessed multimodal signal obtained in step S102 is used to extract the isotherm of the molten pool profile and the width gradient of the heat-affected zone. The thermal cycling curve of a specific location is calculated based on the time series image, and then thermodynamic features including residence time above the phase transition point, t8 / 5 cooling time and peak temperature are extracted.

[0039] The visual feature vector of molten pool depth, the multi-dimensional feature set of electrical signals, the key spectral physical features and thermodynamic features are respectively normalized by Z-score based on sliding window to eliminate the influence of dimensions and short-term fluctuations.

[0040] The maximum correlation minimum redundancy algorithm based on mutual information is used to filter all standardized features, remove redundant features, and then the filtered features are concatenated according to a preset time window. A learnable feature weight vector is introduced to assign appropriate weights to features of different modalities, and finally a high-dimensional fusion feature vector is formed.

[0041] Specifically, this step constructs a multi-layered, multi-modal deep feature extraction and fusion framework. For the molten pool visual image, a dual-branch convolutional neural network based on an attention mechanism is designed: the static topography branch extracts high-level feature maps through an encoder and regresses the molten pool area, aspect ratio, and contour roundness (defined as...) through a fully connected layer. The dynamic motion branch first calculates the dense motion vector field of the molten pool region using an improved Farneback optical flow method. Its core equation involves least-squares solving of polynomial expansion coefficients. Then, the keyhole oscillation frequency and amplitude are extracted from the vector field time series using a fast Fourier transform, and the velocity distribution at the tail of the molten pool is statistically analyzed. The output features of the two branches are weighted and fused through a learnable channel attention module, with weights... ,in and For activation function, The features after global pooling are ultimately formed into a visual feature vector of the melt pool depth.

[0042] For electrical signals, the probability density distribution skewness of the waveform data can be directly calculated in the time domain. Kurtosis, short-circuit transition frequency, etc.; in the frequency domain, the signal is decomposed into N-level wavelet packet transforms. Each sub-band, calculate the energy of each sub-band. and its proportion of total energy Then calculate the Shannon entropy. Simultaneously, based on the recursion graph reconstructed from the current-voltage phase space, the recursion rate is calculated. and certainty These nonlinear dynamic characteristics together constitute a multi-dimensional feature set of electrical signals.

[0043] For composite plasma spectroscopy, the characteristic spectral lines of elements such as ArI, FeI, SiI, MnI, and CrI are first separated from the original spectrum using Lorentz fitting, and their net intensities are calculated. Electron temperature. Inversion was performed using the Boltzmann diagram method based on the relative intensity ratio of the two Fe I spectral lines: Electron density The full width at half maximum (FWHM) of the Hα or Ar I lines broadened by Stark is then used for calculation. These physical quantities and their time derivatives are taken as key spectral physical features.

[0044] For the temperature field image, isotherms of the molten pool contour are extracted through image processing, and the width gradient of the heat-affected zone perpendicular to the weld direction is calculated. Multiple specific points are selected along the welding direction, and their temperature-time curves (thermal cycling curves) are plotted to accurately extract the residence time above the phase transition point Ac3. Time to cool from 800°C to 500°C and peak temperature Isothermodynamic characteristics.

[0045] Subsequently, all features are standardized by modality using Z-score based on a sliding window: After standardization, the maximum relevance and minimum redundancy algorithm is used for feature selection, with the objective function being: ,in This represents the average mutual information (correlation) between features and target categories. This represents the average mutual information (redundancy) among features. The filtered features are concatenated according to a time-series window and passed through a learnable weight vector. (In model training, it is optimized together with subsequent classifiers) weighted and finally combined into a unified high-dimensional fusion feature vector.

[0046] In this application, this step transforms the original multimodal signals into highly abstract, physically meaningful, and complementary fused feature vectors. By combining deep learning (CNN, attention mechanism) with classical signal processing methods (wavelet packet, recursive quantitative analysis, spectral physical inversion), not only are the deep abstract patterns of the molten pool image captured, but quantitative physical features strongly correlated with the welding physical metallurgical process (such as electron temperature and cooling time) in the electrical, optical, and thermal signals are also extracted. The mRMR is used for feature selection, effectively reducing the dimensionality and redundancy of the feature space and improving the efficiency and generalization ability of subsequent models. Finally, by introducing learnable feature weight vectors for weighted fusion, the model can adaptively assign appropriate contributions to features of different modalities and importance, thereby constructing a high-dimensional fused feature representation that is most discriminative of welding process stability and defect tendency, providing optimal information input for subsequent high-precision real-time quality assessment models.

[0047] Furthermore, the operational mechanism of the quality assessment model built based on the Stacking ensemble learning framework in step S104 includes: The high-dimensional fused feature vector constructed in step S103 is input in parallel into three base learners: Random Forest, Support Vector Machine, and K-Nearest Neighbors. The Random Forest learner, constructed using an extreme random tree algorithm, is used to evaluate feature importance and output the first preliminary prediction probability for each type of defect. The Support Vector Machine learner uses a Gaussian radial basis kernel function to construct a classification hyperplane in high-dimensional space and output the second preliminary prediction probability. The K-Nearest Neighbors learner employs a hybrid metric combining Euclidean distance and dynamic time curvature distance to perform similarity matching in the historical sample database and output the third preliminary prediction probability.

[0048] The vector composed of the first, second, and third preliminary prediction probabilities output by the three base learners is concatenated with the preset key physical features in the original fused feature vector formed in step S103 to form a meta-feature vector, which is then input into the LightGBM meta-learner whose hyperparameters have been optimized by the particle swarm optimization algorithm.

[0049] The LightGBM meta-learner integrates an attention weight allocation module. This module dynamically calculates the decision reliability weights based on the entropy values ​​of the output probability vectors of each base learner, and performs weighted fusion of the first, second, and third preliminary prediction probabilities. The LightGBM model then makes the final decision based on the weighted meta-features, outputting defect category labels and confidence scores.

[0050] Specifically, the quality assessment model operates through a two-stage ensemble learning process. In the first stage, the high-dimensional fused feature vector generated in step S103 is fed in parallel into three heterogeneous base learners. The random forest learner is constructed using the Extra-Trees algorithm, which randomly selects features and thresholds when splitting nodes, outputting the first preliminary prediction probability vector for each type of defect. The support vector machine (SVM) learner employs a Gaussian radial basis function (RBF) kernel function, and simultaneously evaluates feature importance. The feature is mapped to a high-dimensional space to find the optimal classification hyperplane, and a second preliminary prediction probability vector is output based on the distance to the hyperplane. The K-Nearest Neighbors (KNN) learner uses a hybrid distance metric. ,in The dynamic time-bending distance is used to find the K most similar samples in the historical sample database and output a third preliminary prediction probability vector based on their class distribution. .

[0051] In the second stage, the predicted probability vectors of the three base learners are... , , The meta-feature vector is concatenated with key physical features (such as electron temperature and t8 / 5 cooling time) pre-defined in the original fused feature vector. This meta-feature vector is then input into the meta-learner—a LightGBM model optimized for hyperparameters such as the number of trees and learning rate using Particle Swarm Optimization (PSO). This LightGBM model integrates an attention weight allocation module, which calculates the entropy of the output probability vector of each base learner. To measure the uncertainty of its predictions, a lower entropy value indicates a more certain prediction. Subsequently, decision reliability weights are dynamically assigned based on the entropy value. The weights are then used to weight and fuse the probability outputs of the base learners to obtain weighted probabilistic meta-features. Finally, the LightGBM model makes a final decision based on all the weighted meta-features, outputting the defect category label and the corresponding confidence score.

[0052] This application creatively integrates the advantages of multiple machine learning paradigms by constructing a quality assessment model based on the Stacking ensemble learning framework. Random forests effectively evaluate feature importance and handle nonlinear relationships, SVMs have strong generalization capabilities in high-dimensional spaces, and KNNs combined with DTW distance can keenly capture correlations with historically similar working conditions. Compared to a single model, this heterogeneous ensemble strategy significantly improves the robustness and accuracy of defect identification in complex and variable welding processes. Furthermore, instead of simply averaging the results of the base learners, this application introduces an attention weight allocation mechanism, dynamically adjusting the contribution of each base learner based on its current prediction "confidence" (entropy value). This allows the meta-learner's decision to focus more on the more reliable initial judgment, thereby improving the reliability of the final decision. Finally, the meta-learner's hyperparameters are optimized using a particle swarm optimization algorithm to ensure optimal performance of the entire ensemble model. This mechanism enables the quality assessment results to possess both high accuracy and interpretable confidence, providing a reliable basis for subsequent decision optimization.

[0053] Furthermore, the welding parameter optimization decision in step S105 includes: A first-layer CNN regression model was constructed and trained to monitor the relative distance between the laser beam and the electric arc. The construction and training process involved calibration experiments, setting a series of known laser-arc relative distances and simultaneously acquiring corresponding visual images of the molten pool, constructing an image-distance pairing dataset, and using the image-distance pairing dataset to train a deep convolutional neural network with a residual structure. The deep convolutional neural network takes a single frame of molten pool image as input and outputs an estimated value of the relative distance.

[0054] The trained CNN regression model is deployed on an online system to process the visual image of the molten pool in real time and output a distance estimate. The distance estimate is compared with the preset optimal coupling interval to generate a distance deviation signal. The distance deviation signal is input to a fuzzy adaptive PID controller, which outputs an adjustment command for the pose of the laser head or arc welding gun to achieve closed-loop stable control of the relative distance.

[0055] A second-layer PSO-BP neural network optimization model was constructed and invoked to generate a comprehensive parameter adjustment instruction set. The construction process involved collecting a historical welding process database, which contained multiple quality target results of process parameter combinations, working conditions, and corresponding weld penetration, tensile strength, and intergranular corrosion resistance. A multilayer perceptron neural network was trained using the backpropagation algorithm to establish an initial nonlinear mapping model from input parameters to output quality indicators.

[0056] An improved particle swarm optimization algorithm is used to globally optimize the connection weights and biases of the initial nonlinear mapping model. The improvements include the introduction of a linearly decreasing inertial weight strategy and a Pareto optimal solution selection mechanism based on crowding distance, thereby obtaining an optimized PSO-BP neural network model.

[0057] When called online, the real-time defect confidence and category output by the quality assessment model, the current workpiece material and bevel size are used as partial inputs. Combined with the current welding parameters, they are input into the optimized PSO-BP neural network model. The PSO-BP neural network model aims to maximize the comprehensive quality score. It performs forward calculation and fast optimization within the process constraint boundary and outputs a comprehensive parameter adjustment instruction set in real time, including laser power, arc current, welding speed and defocusing amount.

[0058] Specifically, the welding parameter optimization decision in this step adopts a two-level hierarchical optimization control structure. The first level is a closed-loop stable control of the laser-arc relative distance based on CNN regression. First, different known relative distances are obtained through calibration experiments. The melt pool images acquired synchronously are used to construct a training dataset. A deep residual network is trained using this dataset, with its last layer being a regression layer, to achieve processing from single-frame images. Distance estimate The mapping, i.e. When deployed online, real-time images are input into the CNN model, resulting in... and the preset optimal distance range Comparison, generating bias (in (The value is the median of the interval). The deviation and rate of change are input into the fuzzy adaptive PID controller, whose output control quantity drives the actuator to adjust the laser head or welding torch posture, forming a fast inner loop control.

[0059] The second stage involves multi-objective parameter optimization based on a PSO-BP neural network. First, a multilayer perceptron is trained using a historical process database to establish a model from the input vector. (Including process parameters and operating conditions) to the quality target vector Initial nonlinear model (for melting depth, strength, etc.) Subsequently, an improved particle swarm optimization algorithm was used to globally retrain the weights and biases of the neural network, in which the inertial weights decreased linearly. Furthermore, a Pareto solution selection mechanism based on congestion distance is introduced to maintain the diversity of optimization directions, thus obtaining the optimized PSO-BP model. During online invocation, the current quality status, defect confidence, operating conditions, and current parameters are used as inputs. Under process constraints, the goal is to maximize the overall quality score. (in To normalize quality indicators, Using weights as the objective, this model is used for rapid forward computation and optimization, and outputs a comprehensive set of adjustment instructions for laser power, current, velocity, and defocus in real time. .

[0060] In this application, this step achieves efficient and precise online optimization of welding parameters by constructing a two-level architecture combining "fast-response inner-loop control" and "multi-objective optimization decision-making." The first level, CNN regression combined with fuzzy PID control, achieves millisecond-level direct visual perception and closed-loop stabilization of the key coupling relationship between the laser and the electric arc, ensuring process stability from the energy source. The second level, the PSO-BP model, combines the predictive power of historical data-driven neural networks with the powerful global optimization capability of the improved particle swarm optimization algorithm. It can solve for the optimal parameter adjustment scheme for the current specific defect tendency and working condition in real time under complex multi-objective and multi-constraint conditions. This two-level optimization mechanism enables the system to not only quickly correct immediate physical deviations but also intelligently plan a comprehensive process parameter adjustment strategy oriented towards the final performance goal based on a deep understanding of welding quality, significantly improving the adaptability and intelligence level of the welding process.

[0061] Furthermore, step S106 includes: The system receives the integrated parameter adjustment instruction set generated in step S105. The integrated parameter adjustment instruction set includes the adjustment amount of laser power, arc current, welding speed, and defocusing amount.

[0062] The comprehensive parameter adjustment instruction set is matched with a built-in welding process knowledge base, which is stored in a relational database. Each record contains material combination, bevel type, plate thickness, historical quality rating, and corresponding set of successful process parameters. The matching process first performs a preliminary screening of the welding process knowledge base records based on the current workpiece material, joint type, and plate thickness. Then, it calculates the weighted Euclidean distance between the comprehensive parameter adjustment instruction set and each candidate parameter set after screening in the adjustment space. Finally, it calls a case-based reasoning algorithm and combines the final quality scores of similar cases to evaluate the applicability of the comprehensive parameter adjustment instruction set.

[0063] The quality status identification result and real-time defect confidence output in step S104 are logically compared with the expert rules stored in the welding process knowledge base. The expert rules include prioritizing fine-tuning the defocusing amount and shielding gas flow rate when there is a tendency for porosity, and prioritizing increasing the laser power and reducing the welding speed when there is a tendency for incomplete penetration.

[0064] The matching and verification results are input into an optimization decision engine, which generates a final executable fine-tuning strategy based on fuzzy reasoning and constraint satisfaction algorithms. The executable fine-tuning strategy can take the form of a parameter gradient curve designed to adapt to gradual changes in plate thickness or gap, where the four key parameters—laser power, arc current, welding speed, and defocusing amount—change continuously at a specific slope within the next welding length; or a coordinated compensation pulse parameter package of laser power and arc current for a sustained period of time, triggered to instantly suppress specific defects.

[0065] The executable fine-tuning strategy is encoded into a sequence of execution instructions that can be directly issued to the welding power source and motion controller.

[0066] Specifically, this step is the key security verification and policy conversion module that connects optimization decision-making and physical execution. It receives a comprehensive set of parameter adjustment instructions from the PSO-BP neural network output in step S105. First, the instruction set is matched and verified against a relational welding process knowledge base: a preliminary screening is performed based on the current workpiece material, bevel type, and plate thickness, and all similar historical case records are retrieved from the knowledge base. Subsequently, the weighted Euclidean distance between the instruction set to be executed and the parameter set of each candidate case in the adjustment space is calculated. Furthermore, by combining a case-based reasoning algorithm with the final quality scores of the K nearest cases, the applicability and risks of the current instruction set are assessed.

[0067] Simultaneously, the real-time quality status output in step S104 (such as "porosity tendency, confidence level 85%) is logically compared with the expert rules stored in the knowledge base (e.g., IF "porosity tendency exists" THEN "prioritize adjusting defocus amount and protective gas flow rate") to generate rule verification results.

[0068] The matching and validation results (case similarity, rule compliance, etc.) are input into an optimization decision engine. This engine, based on fuzzy reasoning (converting the input into fuzzy quantities, such as "high matching degree") and constraint satisfaction algorithms (ensuring new parameters do not exceed the physical limits of the device), optimizes the original... The strategy is then refined and improved to generate a final, executable fine-tuning strategy. This strategy has two main forms: one is a parameter gradient curve to adapt to varying working conditions, such as laser power varying with welding distance x. Linear variation; secondly, a collaborative compensation pulse parameter package for immediate suppression of defects, such as within a time window. A laser power pulse of a specific waveform is applied internally. Ultimately, this strategy is encoded into a sequence of instructions that can be directly parsed and executed by the welding power supply and motion controller.

[0069] In this application, this step adds a layer of "safety verification" and "practice correction" based on historical experience and domain knowledge to the optimization instructions generated by the purely data-driven model by introducing a welding process knowledge base and an expert rule base. This effectively prevents the model from outputting unrealistic or even dangerous parameter instructions due to getting stuck in local optima or data bias, greatly improving the reliability and safety of the system. At the same time, the optimization decision engine transforms abstract adjustment quantities into two highly structured executable strategies (gradual curves and compensation pulses), which enables control instructions to accurately match different types of process disturbances (slow condition changes or sudden defects), achieving a refined and reliable conversion from optimization objectives to execution actions, ensuring that intelligent decisions can be applied safely, effectively, and smoothly to the physical welding process.

[0070] Furthermore, step S107 includes: The main controller receives and parses the executable fine-tuning strategy generated in step S106.

[0071] For the parameter gradient curve strategy in the executable fine-tuning strategy, the main controller discretizes the parameter gradient curve into a series of time-parameter value pairs with serial numbers according to the spatial position based on the welding speed. Through the high-speed fieldbus based on EtherCAT, the sub-instructions and the corresponding target position information are synchronously and in real time sent to the laser power supply, arc welding power supply and motion servo driver, driving the laser power, arc current, welding speed and defocusing amount to change smoothly along a predetermined trajectory within the preset welding length.

[0072] For the real-time compensation pulse parameter strategy in the executable fine-tuning strategy, the main controller sends a pulse trigger command and parameter packet with nanosecond-level precision timestamps to the fast digital interface of the laser and arc power supply within a millisecond-level delay after receiving the defect trigger signal. This ensures that the energy output of the laser and arc power supply is coordinated and adjusted according to the preset waveform within a short time. During execution, the main controller collects and monitors the feedback status words of the welding power supply and servo driver in real time, including actual output power, current, speed, and position. It then compares the actual values ​​with the expected values ​​in the command sequence in a loop to generate execution status deviations.

[0073] When the deviation in the execution status exceeds the preset tolerance, an online safety protection strategy that includes parameter rollback and process interruption is immediately triggered to ensure the accuracy of parameter adjustment and the safety of the process.

[0074] Specifically, after receiving the executable fine-tuning strategy issued in step S106, the main controller first parses it to identify the strategy type (gradual curve or compensation pulse) and the contained parameter instructions. For the parameter gradual curve strategy, the main controller, based on the current welding speed v, generates a continuous curve describing the change of parameters with the weld position x. Through formula Discretize into a series of sub-instruction points that strictly correspond to time or position. These numbered command points are sent to each actuator in precise timing synchronization via a high-speed fieldbus based on EtherCAT, ensuring that laser power, arc current, welding speed, and defocusing amount can change smoothly and in a coordinated manner according to the predetermined trajectory.

[0075] For the instantaneous compensation pulse parameter strategy, after receiving the defect trigger signal from the quality assessment module, the main controller, within a millisecond delay, sends trigger commands with high-precision nanosecond-level timestamps and preset parameter packages (such as pulse waveform, amplitude, and width) to the laser and arc welding power source via a dedicated high-speed digital I / O interface (such as PXIe). This ensures that the two energy sources can coordinate their actions according to the predetermined waveform in an extremely short time, achieving precise energy compensation. Throughout the strategy execution, the main controller reads the feedback status words (actual power, current, position, etc.) of each actuator in real time via the EtherCAT bus and compares them cyclically with the expected values ​​in the command sequence to calculate the execution status deviation. Once any parameter deviation exceeds the preset safety tolerance... The system will immediately trigger safety protection strategies, such as reverting parameters to safe values ​​or interrupting the welding process.

[0076] In this application, this step represents the "last mile" in translating intelligent decision-making into physical action. Its core value lies in achieving high-precision synchronous execution and closed-loop safety monitoring. By discretizing the spatial curve into time-series instructions according to welding speed and utilizing the high deterministic communication of the EtherCAT bus, the smoothness and accuracy of multi-parameter coordinated changes are ensured. By adding nanosecond-level timestamps to the compensation pulse instructions, strict synchronization of laser and arc energy adjustments is ensured, which is crucial for suppressing transient defects. More importantly, this step constructs a safety closed loop for real-time instruction tracking and comparison, capable of instantly detecting execution deviations and triggering protection. This extends the reliability of the intelligent optimization system from the "decision level" to the "execution level," effectively preventing process runaway caused by actuator errors or communication delays, and ensuring the safety and stability of the entire adaptive welding process.

[0077] Furthermore, step S108 includes: After a welding unit is completed, the main controller packages the original multimodal signals, extracted fusion feature vectors, quality assessment results, executed fine-tuning strategy instruction set, and final macroscopic quality sampling results of the welding unit collected during steps S102 to S107 into a complete process-result data packet, which is then encrypted and stored and marked with a unique spatiotemporal identifier.

[0078] The main controller calls the built-in self-evaluation module to compare and analyze the actual process data and quality results of the welding unit with the theoretical process and results predicted by the welding process digital twin model based on the initial parameters, and generates a self-evaluation report on sensor effectiveness, feature extraction robustness, and accuracy of parameter adjustment strategy.

[0079] Based on the self-assessment report, the incremental learning of the model and the dynamic update of the knowledge base are initiated during the welding interval. The incremental learning of the model uses the newly added process-result data package to incrementally train and fine-tune the meta-learner of the Stacking ensemble learning quality assessment model and the second-layer PSO-BP neural network optimization model. The dynamic update of the knowledge base abstracts the validated successful fine-tuning strategy cases and corresponding working condition features into new expert rules, which are then integrated and updated into the welding process knowledge base.

[0080] In step S105, the incrementally optimized PSO-BP neural network optimization model performs online optimization based on the updated weight parameters. In step S106, the optimization decision engine calls the dynamically expanded welding process knowledge base for matching and verification.

[0081] Specifically, after a welding unit is completed, the system automatically executes a data archiving and knowledge evolution process. The main controller packages all data generated throughout the unit, including the raw multimodal signals acquired in step S102, the fused feature vectors extracted in step S103, the quality assessment results and confidence levels in step S104, the executable fine-tuning strategy instruction set generated in step S106, and the final macroscopic quality sampling results (such as a non-destructive testing report), into a structured "process-result" data package. This data package is encrypted and stored, and marked with a unique spatiotemporal identifier (such as "workpiece ID, weld number, timestamp"). Subsequently, the system calls the self-evaluation module to compare the actual welding process data with the theoretical process curves and results predicted by the welding process digital twin model based on the initial set parameters. By calculating the root mean square error and correlation coefficient between the predicted and actual values, the system quantitatively analyzes the effectiveness of the sensor data, the robustness of feature extraction, and the accuracy of the parameter adjustment strategy, generating a structured self-evaluation report.

[0082] Based on this report, during welding breaks, the system initiates incremental learning and knowledge updates. For incremental model learning, the system uses new "process-result" data packages as training samples to incrementally train the LightGBM meta-learner of the Stacking ensemble learning quality assessment model and the second-layer PSO-BP neural network optimization model. This is achieved by minimizing the incremental loss function. To fine-tune the model parameters and adapt them to new working conditions, the system dynamically updates the knowledge base. Successfully validated fine-tuning strategy cases, along with their triggering working condition characteristics (such as defect type, material, and bevel), are abstracted into expert rules in "IF-THEN" form. After deduplication and consistency checks, these rules are integrated into the relational database of the welding process knowledge base. Subsequently, in step S105, the incrementally optimized PSO-BP neural network optimization model uses the updated weight parameters for online optimization. In step S106, the optimization decision engine calls the dynamically expanded welding process knowledge base, which contains more cases and rules, for matching and verification.

[0083] In this application, this step constructs a complete data closed loop of "perception-decision-execution-learning" and a system self-evolution mechanism. By systematically archiving the entire "process-result" data chain, valuable digital assets are provided for subsequent analysis and optimization. The self-evaluation module realizes the monitoring and quantitative evaluation of the system's own performance, ensuring the continuous reliability of each submodule. Most importantly, the model update based on incremental learning and the dynamic expansion of the knowledge base based on case reasoning enable the entire intelligent welding system to self-optimize using the experience of each welding practice. This breaks the limitations of traditional welding systems with fixed parameters and the inability to accumulate experience, enabling the system to continuously learn, adapt to new working conditions, and evolve process knowledge in actual use, thereby continuously improving the consistency, reliability, and adaptability of welding quality and process window.

[0084] This application also provides a laser-arc hybrid welding system, the system comprising: The position deviation compensation module is used to acquire measurement data of the workpiece to be welded, generate a real-time position deviation signal between the welding torch and the center line of the weld, and drive the welding head to perform initial positioning.

[0085] The online quality assessment module is used to collect multimodal sensor signals in real time during the welding process and perform preprocessing, then extract and fuse deep features, and input the obtained high-dimensional fused feature vector into a pre-trained quality assessment model built on the Stacking ensemble learning framework to output the identification results and confidence levels of welding defects in real time.

[0086] The optimization decision module is used to trigger welding parameter optimization decisions when the confidence level of the identification result is lower than the threshold or a specific defect tendency is detected. First, the relative distance between the laser and the electric arc is controlled in a closed loop, and then a comprehensive parameter adjustment instruction set is generated online based on a multi-objective optimization algorithm.

[0087] The strategy verification module is used to match and verify the comprehensive parameter adjustment instruction set with the welding process knowledge base to generate the final executable fine-tuning strategy.

[0088] The parameter execution module is used to receive and execute executable fine-tuning strategies to dynamically adjust the output parameters of the laser-arc hybrid welding head.

[0089] The welding control module is used to control the online quality assessment module, optimization decision-making module, strategy verification module, and parameter execution module to run cyclically during the welding process until the welding is completed.

[0090] It should be noted that the terms "first," "second," and similar terms used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "a" or "one," and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. "A plurality" or "several" indicates at least two. Unless otherwise stated, terms such as "front," "back," "left," "right," "lower," and / or "upper" are for illustrative purposes only and are not limited to a location or spatial orientation. Terms such as "comprising" or "including" indicate that the elements or objects preceding "comprising" encompass the elements or objects listed following "comprising" or "including" and their equivalents, and do not exclude other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0091] The singular forms “a,” “the,” and “the” used in this application specification and appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0092] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A laser-arc hybrid welding method, characterized in that, Includes the following steps: S101, acquire the three-dimensional point cloud data and contact measurement data of the workpiece to be welded, perform fusion processing on the three-dimensional point cloud data and the contact measurement data to generate a real-time positional deviation signal between the welding gun and the center line of the weld, and drive the laser-arc composite welding head to perform initial positioning according to the real-time positional deviation signal so that the positional deviation is not greater than the preset positional deviation threshold. S102, the laser-arc composite welding head is started to perform welding. During the welding process, the visual image of the molten pool, the composite plasma spectrum, the welding current and voltage waveforms, and the temperature field distribution image of the weld area are acquired in real time and synchronously. The acquired raw signals are preprocessed by noise reduction and synchronous alignment to obtain the preprocessed multimodal signal. S103, Deep feature extraction is performed on the preprocessed multimodal signal obtained in step S102, and standardization and feature-level fusion are performed to form a high-dimensional fusion feature vector for evaluating the stability and defect tendency of the welding process. S104, The high-dimensional fusion feature vector constructed in step S103 is input into the pre-trained quality assessment model. The quality assessment model is built based on the Stacking ensemble learning framework. The base learners of the Stacking ensemble learning framework include random forest, support vector machine and K-nearest neighbor learner. The meta-learner of the Stacking ensemble learning framework is a deep forest model based on attention mechanism, which is used to weight and integrate the outputs of the base learners. The quality assessment model outputs the identification results and confidence scores of defect states including weld formation, porosity, incomplete penetration, undercut and hump in real time. S105, when the confidence level of the identification result output in step S104 is lower than the preset confidence threshold or a specific defect tendency is detected, a welding parameter optimization decision is triggered: First, based on the visual image of the molten pool, the relative distance between the laser beam and the arc is monitored in real time and controlled in a closed loop to stabilize the relative distance within the preset optimal coupling range; then, with the current quality status, workpiece material, bevel size, and the real-time defect confidence level output by the quality assessment model as inputs, and with weld penetration depth, tensile strength, and intergranular corrosion resistance as multiple objectives, online multi-objective optimization is performed under process constraints to generate a comprehensive parameter adjustment instruction set including laser power, arc current, welding speed, and defocusing amount in real time; S106, the comprehensive parameter adjustment instruction set generated in step S105 is matched and verified with the built-in welding process knowledge base, and the final executable fine-tuning strategy is generated based on the matching and verification results. The executable fine-tuning strategy includes at least a parameter gradient curve within the length of the next welding unit or an instant compensation pulse parameter for a specific defect. S107, Receive and execute the executable fine-tuning strategy generated in step S106, and dynamically adjust the output parameters of the laser-arc composite welding head; S108, Repeat steps S102 to S107 until the entire weld seam is completed.

2. The laser-arc hybrid welding method according to claim 1, characterized in that, Step S102 includes: Visual images of the molten pool and keyhole area are acquired using a combination of a high-speed CMOS camera and a narrow-band filter. The emission spectra of the laser-arc composite plasma in the ultraviolet to near-infrared bands are simultaneously acquired using a fiber optic spectrometer coupled with a six-channel high-speed beam splitter, serving as the composite plasma spectrum. Instantaneous waveforms of welding current and voltage are acquired using a Hannover welding quality analyzer, serving as the welding current and voltage waveforms, and the short-time Fourier transform spectrum of these instantaneous waveforms is calculated. Temperature field distribution images of the weld and its heat-affected zone are acquired using a mid-wave infrared thermal imager, and real-time temperature calibration is performed based on colorimetric thermometry. The acquired molten pool visual image is processed using a hybrid algorithm combining adaptive weighted median filtering and wavelet thresholding to remove spatter and plasma scintillation noise. The composite plasma spectrum is smoothed using a moving window Savitzky-Golay filter, and baseline correction is performed using iterative polynomial fitting to eliminate continuous background radiation. The welding current and voltage waveforms are first processed using a power frequency interference cancellation technique based on linear prediction, and then processed by a finite-length unit impulse response low-pass filter designed with a Kaiser window to separate and smooth high-frequency noise. The temperature field distribution image is processed using a two-point non-uniformity correction method based on a reference blackbody and a bad pixel compensation algorithm based on neighborhood weighted interpolation. Synchronous alignment preprocessing forces all acquisition devices to start synchronously at the welding start moment by using a global hardware trigger signal based on the output of a precision programmable delay pulse generator. During continuous acquisition, each data frame is marked with a high-precision, low-drift absolute timestamp provided by a rubidium atomic clock module. Based on the absolute timestamp, a dynamic time warping algorithm is used to perform time series alignment and interpolation compensation on the cross-modal data, generating time-series multimodal signal data blocks that are strictly synchronized in the time domain and aligned in the feature dimension, which serve as the preprocessed multimodal signal.

3. The laser-arc hybrid welding method according to claim 2, characterized in that, Step S103 includes: A dual-branch convolutional neural network based on an attention mechanism is used to process the molten pool visual image in the preprocessed multimodal signal obtained in step S102. One branch extracts the static morphological features of the molten pool, including area, aspect ratio, contour roundness, and trailing angle. The other branch calculates the dynamic motion vector field of the molten pool region using optical flow method, extracts the main frequency, amplitude, and flow velocity distribution features of the keyhole oscillation, and the flow velocity distribution features at the tail of the molten pool. The features of the two branches are then weighted and fused in the channel dimension to form a molten pool depth visual feature vector. For the welding current and voltage waveforms in the preprocessed multimodal signal obtained in step S102, statistical features including probability density distribution skewness and kurtosis, short-circuit transition frequency, coefficient of variation, and waveform factors are extracted in the time domain. In the frequency domain, wavelet packet transform decomposition is performed to calculate the Shannon entropy of the energy proportion of each sub-band. Waveform nonlinear dynamic features based on recursive quantitative analysis are constructed. The waveform nonlinear dynamic features include recursion rate and determinism, which together constitute a multidimensional feature set of the electrical signal. For the composite plasma spectrum in the preprocessed multimodal signal obtained in step S102, the characteristic spectral lines of argon, iron, silicon, manganese and chromium at specific wavelengths are identified, the net intensity is calculated, and the electron temperature and electron density of the plasma are inverted in real time. The intensity ratio, intensity ratio change rate, electron temperature and electron density of each element are used as key spectral physical features. For the temperature field distribution image in the preprocessed multimodal signal obtained in step S102, extract the isotherm of the molten pool profile and the width gradient of the heat-affected zone, and calculate the thermal cycling curve at a specific location based on the time series image, and then extract thermodynamic features including the residence time above the phase change point, t8 / 5 cooling time and peak temperature. The visual feature vector of the molten pool depth, the multi-dimensional feature set of the electrical signal, the key spectral physical features and the thermodynamic features are respectively normalized by Z-score based on a sliding window to eliminate the influence of dimensions and short-term fluctuations. The maximum correlation minimum redundancy algorithm based on mutual information is used to filter all standardized features, remove redundant features, and then the filtered features are concatenated according to a preset time window. A learnable feature weight vector is introduced to assign appropriate weights to features of different modalities, and finally a high-dimensional fusion feature vector is formed.

4. The laser-arc hybrid welding method according to claim 3, characterized in that, The operating mechanism of the quality assessment model built based on the Stacking ensemble learning framework described in step S104 includes: The high-dimensional fused feature vector constructed in step S103 is input in parallel to the three base learners: the random forest, the support vector machine, and the K-nearest neighbors. The random forest learner is constructed using an extreme random tree algorithm to evaluate feature importance and output a first preliminary prediction probability for various defects. The support vector machine learner uses a Gaussian radial basis kernel function to construct a classification hyperplane in high-dimensional space and output a second preliminary prediction probability. The K-nearest neighbors learner uses a hybrid metric that combines Euclidean distance and dynamic time curvature distance to perform similarity matching in the historical sample database and output a third preliminary prediction probability. The vector composed of the first preliminary prediction probability, the second preliminary prediction probability, and the third preliminary prediction probability output by the three base learners is concatenated with the preset key physical features in the original fused feature vector formed in step S103 to form a meta-feature vector, which is then input into the LightGBM meta-learner whose hyperparameters have been optimized by the particle swarm algorithm. The LightGBM meta-learner integrates an attention weight allocation module. The attention weight allocation module dynamically calculates the decision reliability weight of each base learner based on the entropy value of the output probability vector of each base learner, and performs weighted fusion of the first preliminary prediction probability, the second preliminary prediction probability, and the third preliminary prediction probability. The LightGBM model then makes the final decision based on the weighted meta-features and outputs the defect category label and confidence score.

5. The laser-arc hybrid welding method according to claim 4, characterized in that, The welding parameter optimization decision in step S105 includes: A first-layer CNN regression model was constructed and trained to monitor the relative distance between the laser beam and the electric arc. The construction and training process involved conducting calibration experiments, setting a series of known laser-arc relative distances and simultaneously acquiring corresponding visual images of the molten pool, constructing an image-distance pairing dataset, and using the image-distance pairing dataset to train a deep convolutional neural network containing a residual structure. The deep convolutional neural network takes a single frame of molten pool image as input and outputs an estimated value of the relative distance. The trained CNN regression model is deployed on an online system to process the visual image of the molten pool in real time and output a distance estimate. The distance estimate is compared with the preset optimal coupling interval to generate a distance deviation signal. The distance deviation signal is input to a fuzzy adaptive PID controller, which outputs an adjustment command for the pose of the laser head or arc welding gun to achieve closed-loop stable control of the relative distance. A second-layer PSO-BP neural network optimization model is constructed and invoked to generate the comprehensive parameter adjustment instruction set. The construction process involves collecting a historical welding process database, which contains multiple quality target results of process parameter combinations and working conditions with corresponding weld penetration, tensile strength, and intergranular corrosion resistance. A multilayer perceptron neural network is trained using the backpropagation algorithm to establish an initial nonlinear mapping model from input parameters to output quality indicators. An improved particle swarm optimization algorithm is used to globally optimize the connection weights and biases of the initial nonlinear mapping model. The improvements include the introduction of a linear decreasing strategy for inertial weights and a Pareto optimal solution selection mechanism based on crowding distance, thereby obtaining an optimized PSO-BP neural network model. When invoked online, the real-time defect confidence and category, current workpiece material and bevel size output by the quality assessment model are used as partial inputs. Combined with the current welding parameters, these are input to the optimized PSO-BP neural network model. The PSO-BP neural network model aims to maximize the comprehensive quality score. It performs forward calculation and rapid optimization within the process constraint boundary and outputs the comprehensive parameter adjustment instruction set, which includes laser power, arc current, welding speed and defocusing amount, in real time.

6. The laser-arc hybrid welding method according to claim 5, characterized in that, Step S106 includes: Receive the integrated parameter adjustment instruction set generated in step S105, the integrated parameter adjustment instruction set including the adjustment amount of laser power, arc current, welding speed and defocus amount; The comprehensive parameter adjustment instruction set is matched with the built-in welding process knowledge base, which is stored in the form of a relational database. Each record contains material combination, bevel type, plate thickness, historical quality rating, and corresponding set of successful process parameters. The matching process first performs a preliminary screening of the welding process knowledge base records based on the current workpiece material, joint type, and plate thickness. Then, the weighted Euclidean distance between the comprehensive parameter adjustment instruction set and each candidate parameter set after screening is calculated in the adjustment space. Finally, an algorithm based on case reasoning is called to evaluate the applicability of the comprehensive parameter adjustment instruction set by combining the final quality scores of similar cases. The quality status identification result and the real-time defect confidence level output in step S104 are logically compared with the expert rules stored in the welding process knowledge base. The expert rules include prioritizing fine-tuning the defocusing amount and the protective gas flow rate when there is a tendency for porosity, and prioritizing increasing the laser power and reducing the welding speed when there is a tendency for incomplete penetration. The matching and verification results are input into an optimization decision engine, which generates a final executable fine-tuning strategy based on fuzzy reasoning and constraint satisfaction algorithms. The executable fine-tuning strategy can take the form of a parameter gradient curve designed to adapt to the gradual change in plate thickness or gap, in which the four key parameters of laser power, arc current, welding speed and defocusing amount change continuously at a specific slope in the next welding length, or a coordinated compensation pulse parameter package of laser power and arc current for a period of time triggered to suppress specific defects in real time. The executable fine-tuning strategy is encoded as a sequence of execution instructions that can be directly issued to the welding power source and motion controller.

7. The laser-arc hybrid welding method according to claim 6, characterized in that, Step S107 includes: The main controller receives and parses the executable fine-tuning strategy generated in step S106; For the parameter gradient curve strategy in the executable fine-tuning strategy, the main controller discretizes the parameter gradient curve into a series of time-parameter value pairs with serial numbers according to the spatial position based on the welding speed. Through the high-speed fieldbus based on EtherCAT, the sub-instructions and the corresponding target position information are synchronously and in real time sent to the laser power supply, the arc welding power supply and the motion servo driver, driving the laser power, the arc current, the welding speed and the defocusing amount to change smoothly along a predetermined trajectory within the preset welding length. For the instantaneous compensation pulse parameter strategy in the executable fine-tuning strategy, within a millisecond delay after receiving the defect trigger signal, the main controller sends a pulse trigger command and parameter packet with nanosecond-level precision timestamps to the fast digital interface of the laser and arc power supply to ensure that the energy output of the laser and arc power supply is coordinated and adjusted according to the preset waveform in a short time. During the execution process, the main controller collects and monitors the feedback status words of the welding power supply and servo driver in real time, including the actual output power, current, speed and position, and compares the actual values ​​with the expected values ​​in the command sequence in a loop to generate the execution status deviation. When the deviation of the execution status exceeds the preset tolerance, an online safety protection strategy including parameter rollback and process interruption is immediately triggered to ensure the accuracy of parameter adjustment and the safety of the process.

8. The laser-arc hybrid welding method according to claim 7, characterized in that, Step S108 includes: After a welding unit is completed, the main controller packages the original multimodal signals, extracted fusion feature vectors, quality assessment results, executed fine-tuning strategy instruction set, and final macroscopic quality sampling results of the welding unit collected during steps S102 to S107 into a complete process-result data packet, which is then encrypted and stored and marked with a unique spatiotemporal identifier. The main controller calls the built-in self-evaluation module to compare and analyze the actual process data and quality results of the welding unit with the theoretical process and results predicted by the welding process digital twin model based on the initial parameters, and generates a self-evaluation report on sensor effectiveness, feature extraction robustness and parameter adjustment strategy accuracy. Based on the self-assessment report, during the welding interval, the incremental learning of the model and the dynamic update process of the knowledge base are initiated. The incremental learning process uses the newly added process-result data package to incrementally train and fine-tune the meta-learner of the Stacking ensemble learning quality assessment model and the second-layer PSO-BP neural network optimization model. The dynamic update process of the knowledge base abstracts the verified successful fine-tuning strategy cases and corresponding working condition features into new expert rules, which are then integrated and updated into the welding process knowledge base. In step S105, the incrementally optimized PSO-BP neural network optimization model performs online optimization based on the updated weight parameters. In step S106, the optimization decision engine calls the dynamically expanded welding process knowledge base for matching and verification.

9. A laser-arc hybrid welding system, characterized in that, The system includes: The pose deviation compensation module is used to acquire measurement data of the workpiece to be welded, generate real-time pose deviation signals between the welding torch and the weld centerline, and drive the welding head to perform initial positioning. The online quality assessment module is used to collect multimodal sensor signals in real time during the welding process and preprocess them to extract and fuse deep features. The resulting high-dimensional fused feature vector is then input into a pre-trained quality assessment model built on the Stacking ensemble learning framework to output the identification results and confidence levels of welding defects in real time. The optimization decision module is used to trigger welding parameter optimization decisions when the confidence level of the identification result is lower than the threshold or a specific defect tendency is detected. First, the relative distance between the laser and the electric arc is controlled in a closed loop, and then a comprehensive parameter adjustment instruction set is generated online based on a multi-objective optimization algorithm. The strategy verification module is used to match and verify the comprehensive parameter adjustment instruction set with the welding process knowledge base to generate the final executable fine-tuning strategy. The parameter execution module is used to receive and execute the executable fine-tuning strategy to dynamically adjust the output parameters of the laser-arc hybrid welding head. The welding control module is used to control the online quality assessment module, optimization decision module, strategy verification module and parameter execution module to run cyclically during the welding process until the welding is completed.