Weld seam recognition and welding parameter self-adaptive regulation method and system

Abstract

The application discloses a kind of weld seam identification and welding parameter self-adaptive control method and system, method steps: acquisition workpiece three-dimensional point cloud, welding real-time data, workpiece and working condition data;Using deep learning weld seam identification model to extract point cloud multiscale feature, output weld seam position posture and groove geometric parameter;Optimal welding parameter is generated by relying on welding parameter analysis model matching process parameter library;Welding quality is determined on-line by quality prediction model when welding, and closed-loop parameter adjustment is carried out when abnormal;Unqualified data after welding is realized process self-healing by deep reinforcement learning, and model is regularly iterated, and production line efficiency is analyzed to optimize scheduling.System includes data acquisition layer, intelligent analysis layer, decision output layer, execution control layer.The application realizes high-precision weld seam perception by relying on three-dimensional point cloud, constructs multilayer quality closed loop and process autonomous iteration mechanism, adapts to complex industrial working conditions, and improves welding quality and production line operation efficiency.

Description

Technical Field

[0001] This invention relates to the fields of intelligent welding manufacturing and machine vision technology, specifically to a method and system for weld seam recognition and adaptive control of welding parameters. Background Technology

[0002] In current welding production scenarios, traditional manual welding relies heavily on welders' practical experience, resulting in problems such as low weld positioning accuracy, poor parameter setting consistency, large quality fluctuations under complex working conditions, and limited production efficiency. Existing automated welding equipment mostly adopts preset trajectories and fixed process parameter schemes, which are insufficiently adaptable to workpiece assembly errors, differences in bevel morphology, and changes in on-site working conditions.

[0003] Some welding systems that incorporate visual recognition can only achieve basic weld seam positioning and struggle to finely deconstruct the geometric features of the bevel. Furthermore, welding parameter matching largely relies on preset process tables, failing to dynamically optimize based on real-time operating conditions. They lack online quality closed-loop control capabilities and an autonomous iteration mechanism for process knowledge. In addition, for complex scenarios common in industrial settings such as rust, oil stains, arc light reflections, and dust interference, the robustness and accuracy of existing weld seam recognition models are insufficient to meet the demands of high-precision welding, hindering the intelligent upgrading of welding production lines. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for weld identification and adaptive control of welding parameters, which solves the technical problems of low weld identification accuracy, poor adaptability to complex working conditions, insufficient parameter matching accuracy, and lack of quality closed-loop control and process autonomous iteration capability in existing welding systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for weld identification and adaptive control of welding parameters includes the following steps:

[0007] S1. Collect 3D point cloud data of the workpiece to be welded, real-time running data of the welding process, and workpiece attribute and on-site working condition data;

[0008] The real-time operating data of the welding process includes welding current, welding voltage, and molten pool temperature data;

[0009] S2. Multi-scale feature extraction is performed on the 3D point cloud data of the workpiece using a deep learning weld recognition model to obtain the position coordinates, spatial orientation parameters and weld type information of the weld, and output the bevel geometric feature parameters.

[0010] S3. Based on weld information, groove geometric feature parameters, workpiece material properties, workpiece thickness parameters and on-site working condition data, the welding parameter analysis model calls the pre-built welding process parameter library to perform multi-dimensional matching calculations, generate the optimal welding parameters that match the current welding conditions and output them.

[0011] S4. Perform welding operations according to the output optimal welding parameters, and simultaneously perform online quality analysis and anomaly determination based on the real-time collected operation data during the welding process through a quality prediction model; when an anomaly is determined, the welding process parameters are adjusted in a closed loop.

[0012] Preferably, the three-dimensional point cloud data is acquired using a line laser three-dimensional scanning unit; the line laser three-dimensional scanning unit is mounted on a welding robotic arm or fixed to the front end of the welding station, and the scanning speed is not less than 2 million points / second, and there is no need to spray developer on the surface of the workpiece to be welded during scanning;

[0013] The deep learning weld recognition model is trained by a multi-scene welding point cloud dataset, which includes point cloud samples of workpieces to be welded under complex lighting, workshop dust environment, and complex weld morphology. It covers scenarios such as rust, oil stains, oxide scale, spot welding residue, component deformation, bevel error, uneven gap, misalignment, burrs, and assembly deviation of large steel components.

[0014] The deep learning weld seam recognition model has a single-frame point cloud recognition processing time of 10ms to 50ms and a weld seam position recognition output accuracy of 0.015 to 0.040mm.

[0015] Preferably, the bevel geometric feature parameters include bevel type, bevel angle, blunt edge dimension, root gap, and bevel depth;

[0016] The welding process parameter library contains benchmark process parameters for various bevel types under different material grades, workpiece thickness ranges, and welding position conditions.

[0017] The generated optimal welding parameter combination includes welding technique type, weld layer and pass number planning, robotic arm offset compensation value, robotic arm pitch angle attitude, welding machine wire feed speed, welding current and voltage matching value, welding speed, oscillation frequency and oscillation amplitude;

[0018] The welding parameter analysis model outputs the optimal welding parameters and simultaneously provides explanations of the derivation basis for the corresponding parameters. The relative deviation between the output core welding parameters and the benchmark optimal parameters verified by standardized process tests under the same working conditions is no more than 10%.

[0019] Preferably, during the online quality analysis process, the quality prediction model performs time-series calculations on the real-time collected welding current, voltage, and molten pool temperature data; the quality prediction model, after offline training and verification, has a welding quality grade prediction accuracy of no less than 85%.

[0020] The welding arc tracking system operates in real time, continuously feeding back the torch posture deviation signal and arc length fluctuation characteristics to the quality prediction model, and autonomously correcting the torch height and oscillation amplitude in a closed loop. When a quality abnormality is detected, an early warning signal is triggered and the process parameter adjustment process is initiated. The process parameter adjustment process combines the molten pool temperature and electrical parameter fluctuation characteristics to dynamically correct thermal input parameters such as welding speed, welding current, and wire feed speed.

[0021] Preferably, when the quality is confirmed to be unqualified after welding, a welding parameter closed-loop self-healing process based on deep reinforcement learning is initiated based on the process closed-loop adjustment.

[0022] The deep reinforcement learning-based closed-loop self-healing process for welding parameters specifically includes: using current working condition characteristics, real-time welding parameters, and quality non-compliance results as state inputs; constructing a reward function using weld formation pass rate and thermal input stability as quantitative indicators; optimizing parameter strategies offline through a deep reinforcement learning agent based on historical non-compliance samples under the same working conditions; outputting a corrected parameter scheme adapted to the current working conditions; synchronously updating the verified corrected parameters and corresponding working condition characteristics to the welding process parameter library; and simultaneously optimizing and updating the welding parameter analysis model and quality prediction model across scenarios through online incremental learning.

[0023] Preferably, it also includes a model periodic iterative update step: based on the labeled welding point cloud data, process parameters and quality result data accumulated during the production process, each analysis model is periodically retrained offline and the parameters are iteratively updated, which complements the online incremental update triggered by quality failure;

[0024] The welding process parameter library uses the collected workpiece surface point cloud features and the corresponding actual welding effect as driving data. The welding parameter analysis model, which has been optimized by incremental learning, infers and analyzes the optimal parameters under new working conditions. After verification, the parameters are automatically added to the library and updated.

[0025] Preferably, it also includes a production line efficiency analysis step: performing multi-dimensional efficiency statistics on the full operation data of the welding production line, comparing and analyzing it with historical benchmarks and design indicators, generating corresponding production line efficiency optimization plans, and outputting them to the production line scheduling system to execute production scheduling adjustments;

[0026] The multi-dimensional performance statistical analysis covers multiple statistical dimensions, including welding cycle time, product qualification rate, equipment utilization rate, and consumable loss rate.

[0027] A weld seam identification and adaptive welding parameter control system includes a data acquisition layer, an intelligent analysis layer, a decision output layer, and an execution control layer that are sequentially connected in communication.

[0028] The data acquisition layer is used to acquire three-dimensional point cloud data of the workpiece to be welded, real-time parameters of the welding process, workpiece attributes and on-site working condition data.

[0029] The intelligent analysis layer has a built-in deep learning weld recognition model, welding parameter analysis model and quality prediction model, and is equipped with a weld recognition deconstruction module, a welding parameter matching module, an online quality discrimination module and a pre-built welding process parameter library to analyze and calculate the collected data.

[0030] The decision output layer is used to output weld location information, optimal welding parameters, quality anomaly warning signals and production line efficiency optimization schemes, and supports integration with external production line scheduling systems.

[0031] The execution control layer is used to control the operation of welding equipment and adjust welding parameters according to the decision output results, and to correct the welding torch position in real time through the arc tracking execution unit in a closed loop.

[0032] Preferably, the data acquisition layer includes a three-dimensional vision acquisition unit, a welding process sensing unit, and a working condition data acquisition unit;

[0033] The three-dimensional vision acquisition unit is used to acquire three-dimensional point cloud data of the surface of the workpiece to be welded, and output raw data containing information on bevel contour, surface morphology and assembly gap.

[0034] The welding process sensing unit is used to collect welding current, voltage, molten pool temperature and welding torch posture deviation signals fed back by the arc tracking system in real time.

[0035] The working condition data acquisition unit is used to collect data on workpiece attributes, equipment operating status, production cycle time, and consumable consumption.

[0036] Preferably, the intelligent analysis layer is further configured with a process knowledge iteration unit and a production line efficiency analysis unit;

[0037] The weld identification and deconstruction module has a built-in deep learning weld identification model, which performs multi-scale feature extraction on three-dimensional point cloud data and outputs weld position coordinates, spatial attitude parameters, as well as groove type, groove angle, blunt edge size, root gap and groove depth parameters.

[0038] The welding parameter matching module has a built-in welding parameter analysis model and welding process parameter library. Based on the bevel characteristic parameters, workpiece attribute data and on-site working condition data, it outputs a complete combination of process parameters that is adapted to the current working condition through multi-dimensional matching calculation.

[0039] The online quality judgment module has a built-in quality prediction model that integrates real-time welding process data and arc tracking feedback signals for time-series analysis and anomaly determination.

[0040] The process knowledge iteration unit has a built-in deep reinforcement learning strategy optimization module and incremental learning module. Based on the actual welding effect feedback data, it continuously updates and optimizes the welding process parameter library, welding parameter analysis model and quality prediction model, and regularly iterates and optimizes the deep learning weld recognition model based on the newly added labeled point cloud data.

[0041] The production line efficiency analysis unit is used to perform multi-dimensional statistical analysis on the full operation data of the production line, generate efficiency optimization schemes, and output them to the production line scheduling system through the decision output layer.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention uses a three-dimensional point cloud perception scheme to identify weld seams. It has strong resistance to interference from the environment and workpiece surface conditions, and can accurately acquire the spatial orientation of the weld seam and the three-dimensional geometric parameters of the bevel. The recognition accuracy and reliability are significantly better than two-dimensional vision schemes.

[0044] This invention constructs a multi-dimensional process parameter matching mechanism, which dynamically generates optimal welding parameters by combining bevel characteristics, workpiece attributes and on-site working conditions. The parameter matching accuracy is high and the range of working conditions is wide, which can effectively reduce the workload of manual process debugging.

[0045] This invention constructs an online quality prediction and closed-loop adjustment system, which integrates arc tracking signals to achieve real-time control of the welding process. It can promptly correct quality deviations caused by fluctuations in operating conditions and improve the stability of welding quality.

[0046] This invention possesses the ability to self-heal through a closed-loop process knowledge loop and to continuously iterate the model. It can continuously optimize the parameter library and analysis model based on production data to achieve autonomous improvement of process capabilities.

[0047] This invention integrates production line efficiency analysis functions, which can evaluate the production line operation status from multiple dimensions, support production line scheduling optimization, and improve the overall operation efficiency of the production line. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method of the present invention.

[0049] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0051] Example 1: Weld Identification and Adaptive Adjustment Method for Welding Parameters

[0052] like Figure 1 As shown in the figure, this embodiment provides a method for weld identification and adaptive control of welding parameters, and the specific implementation method is as follows:

[0053] S1. Multi-source welding data acquisition:

[0054] The system collects three core types of data: three-dimensional point cloud data of the workpiece to be welded, real-time operation data of the welding process, and workpiece attribute and on-site working condition data, providing complete data support for subsequent analysis and decision-making.

[0055] The 3D point cloud data is acquired using a 3D scanning unit. This unit, based on the principle of laser triangulation, projects a structured laser onto the workpiece surface, acquires the reflected laser contour image, and obtains the 3D coordinates of each point on the workpiece surface through geometric calculation, forming a point cloud dataset. The 3D scanning unit can be mounted on a welding robotic arm for motion acquisition or fixedly deployed at the front end of the welding station. The scanning process does not require the application of developing agent to the workpiece surface and can directly acquire data on workpieces with various surface conditions, adapting to the complex workpiece conditions in industrial settings.

[0056] The real-time operating data of the welding process includes welding current, welding voltage and molten pool temperature data, which are collected in real time through corresponding sensing units. The sampling frequency meets the requirements of real-time closed-loop control, which can accurately capture the dynamic changes of parameters during the welding process and fully reflect the heat input and molten pool state of the welding process.

[0057] The workpiece attributes and on-site operating condition data include inherent information such as workpiece material attributes and workpiece thickness, as well as operating condition information such as on-site environment, equipment status, and consumable status. This data can be obtained by connecting to the workshop production management system, equipment controller, and distributed sensor network, providing comprehensive operating condition support for process decision-making.

[0058] S2. Weld seam recognition and bevel feature extraction based on deep learning:

[0059] Using the acquired 3D point cloud data as input, multi-scale feature extraction is performed through a deep learning weld recognition model, and the output includes the weld's position coordinates, spatial attitude parameters, weld type information, and groove geometric feature parameters.

[0060] Before inputting the model, the original 3D point cloud data is preprocessed, including background filtering, noise removal, and downsampling, to improve the quality of the point cloud data and the efficiency of subsequent calculations.

[0061] Background filtering uses a spatial range filtering method. It sets a spatial coordinate threshold based on the preset position of the workpiece, removes background point clouds outside the threshold range, and shrinks the calculation area to the periphery of the weld to reduce the amount of invalid calculation.

[0062] Noise removal employs a statistical filtering algorithm. By calculating the average distance between each point and its nearest neighbors, outlier noise points are identified and removed based on the statistical distribution characteristics of these distances, thus eliminating abnormal points caused by environmental interference during the scanning process. For each point cloud data point, the average distance to its k nearest neighbors is calculated. The average distances of all points follow a Gaussian distribution, with the mean and standard deviation as follows:

[0063]

[0064]

[0065] In the formula, This represents the total number of points in the point cloud. Let be the average nearest neighbor distance of the i-th point. A distance threshold is set; points exceeding the threshold are identified as outliers and removed. The threshold coefficient can be adjusted based on the point cloud quality. Downsampling employs voxel downsampling, dividing the 3D space into meshes using cubic voxels of a set size. Within each voxel, only the centroid representative point of all points is retained. This reduces the point cloud density while preserving the overall geometric features of the point cloud, thus improving computational efficiency.

[0066] After preprocessing, the point cloud data is input into a deep learning weld seam recognition model for computation. This deep learning weld seam recognition model is trained using a multi-scenario welding point cloud dataset. This dataset includes point cloud samples of workpieces to be welded under complex lighting, dusty environments, and complex weld seam morphologies. It covers various typical industrial scenarios such as rust, oil, oxide scale, spot weld residue, component deformation, bevel errors, uneven gaps, misalignments, burrs, and assembly deviations, ensuring the model's ability to generalize in the field.

[0067] The model adopts a point cloud deep learning network architecture with multi-scale feature extraction capabilities, and sets up multiple feature extraction modules with different receptive fields to extract multi-dimensional features of the overall weld trajectory, local contour and micro geometry, so as to realize semantic segmentation and feature parameter calculation of the weld area, while taking into account the continuity of the weld trajectory and the accuracy of the bevel details.

[0068] The model outputs weld position coordinates as a continuous three-dimensional coordinate sequence of the weld centerline. Spatial attitude parameters include spatial angle information such as weld direction and inclination angle, which are used to guide the motion trajectory planning of the welding robot arm. Weld types include common welding joint types such as butt joint, corner joint, and lap joint, providing a joint type basis for process parameter matching.

[0069] The model synchronously outputs bevel geometric feature parameters, specifically including bevel type, bevel angle, blunt edge dimension, root clearance, and bevel depth. These parameters are calculated analytically using a geometric method: first, point clouds of each region of the bevel are separated based on semantic segmentation results; then, the least squares method is used to fit the plane equations of each region. The general expression for the plane is:

[0070]

[0071] In the formula, Let be the unit normal vector of the plane. Let be the distance from the plane to the origin. For all point cloud data points in the slope region, construct a least-squares optimization objective:

[0072]

[0073] The above optimization problem is solved by matrix decomposition to obtain the planar parameters of the slopes on both sides.

[0074] Based on the normal vectors of the two bevel faces, the bevel angle is calculated:

[0075]

[0076] In the formula, , These are the normal vectors of the inclined planes on both sides of the bevel.

[0077] The blunt edge dimension is calculated using the vertical distance between the blunt edge plane and the lowest point of the bevel root. The root gap is calculated using the three-dimensional Euclidean distance between the two bevel root endpoints. The bevel depth is calculated using the vertical distance between the top of the bevel and the root. All calculations are based on the fitted plane equations and feature point coordinates, ensuring the accuracy of parameter calculation. The deep learning weld recognition model's single-frame point cloud recognition processing time meets the real-time requirements of online production, and the weld position recognition output accuracy can reach a high level, adapting to the process requirements of precision welding.

[0078] S3. Generation of optimal welding parameters for multi-dimensional matching:

[0079] Based on weld information, groove geometric feature parameters, workpiece material properties, workpiece thickness parameters, and on-site working condition data, the welding parameter analysis model calls a pre-built welding process parameter library to perform multi-dimensional matching calculations, generate the optimal welding parameters that match the current welding conditions, and output them.

[0080] The welding process parameter library is a pre-built structured process database containing process parameters for various bevel types under different material grades, workpiece thickness ranges, and welding positions. All benchmark parameters have been verified through process experiments and possess reliable process applicability. The parameter library covers commonly used welding materials, thickness ranges, and welding positions, and can be adapted to most industrial welding scenarios.

[0081] Welding heat input is a core control indicator for process parameter matching, directly determining the weld penetration state and forming quality. Its calculation formula is as follows:

[0082]

[0083] In the formula, For heat input per unit length of weld, This is the welding thermal efficiency coefficient. For welding voltage, For welding current, The welding speed is set at the target range. All parameter matching and adjustment are constrained by the target heat input range to ensure that the welding heat input is within a reasonable range and to avoid quality defects such as incomplete fusion and burn-through.

[0084] The matching operation of the welding parameter analysis model is divided into two stages: basic matching and working condition correction.

[0085] The first stage is basic parameter matching. Based on core features such as bevel type, material grade, workpiece thickness, and welding position, a similarity matching algorithm is used to retrieve the benchmark process parameter set with the highest matching degree from the welding process parameter database, which serves as the initial parameter benchmark. Similarity calculation is based on a weighted distance of multi-dimensional features, with the weight of each dimension set according to the degree of process influence.

[0086] The second stage is dynamic correction under operating conditions. The baseline parameters are corrected based on the field operating condition data. The final optimal parameters are determined by the baseline parameters and the correction coefficients for each operating condition.

[0087]

[0088] In the formula, For optimal output parameters, As the baseline process parameters, These are correction factors corresponding to various operating conditions, including correction factors for ambient temperature, equipment condition, and shielding gas. Each factor is calibrated based on process experiments and empirical data. By correcting for operating conditions, the impact of factors such as fluctuations in the on-site environment and changes in equipment condition on welding quality can be offset, thereby improving the adaptability of parameters to operating conditions.

[0089] The final optimal welding parameter combination includes welding technique type, weld layer and pass planning, robotic arm offset compensation value, robotic arm pitch angle attitude, welding machine wire feed speed, welding current and voltage matching value, welding speed, oscillation frequency and oscillation amplitude, covering all core process parameters required for welding execution. The welding parameter analysis model outputs the optimal welding parameters along with explanations of their derivation, including parameter matching sources and operating condition correction logic, facilitating traceability and verification by process engineers. The relative deviation between the output core welding parameters and the optimal parameters verified under the same operating conditions is controlled within a reasonable range, ensuring the accuracy of parameter matching.

[0090] S4. Welding execution and online quality closed-loop control:

[0091] Welding operations are performed according to the output optimal welding parameters. At the same time, online quality analysis and abnormal status judgment are carried out through the quality prediction model based on the real-time operation data collected during the welding process. If the quality is determined to be unqualified, the welding parameters are adjusted in a closed loop.

[0092] After the welding operation is started, the optimal welding parameters are sent to the welding execution equipment. The robotic arm plans the motion trajectory according to the weld position coordinates and attitude parameters. The welding power supply outputs welding energy according to the set electrical parameters and wire feeding parameters, and completes the welding operation according to the preset process.

[0093] The arc tracking system operates in real time throughout the welding process, continuously monitoring the relative positional deviation between the welding torch and the weld, and autonomously completing closed-loop correction of the torch posture to ensure that the torch is always aligned with the weld center. The arc tracking system calculates the posture deviation by detecting fluctuations in the electrical parameters of the welding arc; the left and right deviations satisfy a mapping relationship with the current difference.

[0094]

[0095] In the formula, This represents the left-right deviation between the welding torch and the weld. , These are the average currents when the welding torch swings to its extreme positions on both sides. This is the deviation conversion factor, obtained through calibration. The arc tracking system can correct the welding torch height and oscillation amplitude in real time, offsetting weld offsets caused by workpiece assembly deviations and thermal deformation, and is the basic positional guarantee for the welding process.

[0096] The welding torch posture deviation signal fed back by the arc tracking system and the arc length fluctuation characteristics are synchronously input into the quality prediction model as auxiliary features for quality analysis.

[0097] Online quality analysis is performed by a quality prediction model. The model processes real-time collected welding current, voltage, and molten pool temperature data using time-series computation to extract the fluctuation, trend, and statistical characteristics of the parameters. This quality prediction model, trained and validated offline, employs a temporal deep learning network architecture, effectively capturing the mapping relationship between welding process parameters and welding quality. The prediction results show a high degree of agreement with actual welding quality, meeting industrial application requirements. The quality prediction model integrates multi-dimensional features from electrical parameters, molten pool temperature, and arc tracking to output a comprehensive quality assessment result and classify quality levels. The comprehensive quality score is calculated by weighting multiple dimensions of indicators.

[0098]

[0099] In the formula, For the overall quality score, Scores are awarded for each individual indicator, including electrical parameter stability score, molten pool temperature stability score, and positional deviation score. The weighting coefficients for the corresponding indicators are set according to the degree of influence of each indicator on welding quality.

[0100] Different quality levels can be divided based on the comprehensive score. When the score is lower than the qualified threshold, it is judged as a quality abnormality, triggering an early warning signal and initiating the parameter adjustment process.

[0101] The parameter adjustment process combines the types of quality anomalies to specifically adjust welding process parameters. If the molten pool temperature is abnormal or the heat input deviates from the reasonable range, the welding current, welding speed, wire feed speed, and other heat input-related parameters are adjusted accordingly. If the positional deviation is too large, leading to poor fusion, the torch offset and oscillation parameters are adjusted accordingly. Among these, the positional deviation feedback from the arc tracking serves as the core basis for positional adjustment. Heat input parameters are dynamically corrected based on the fluctuation characteristics of the molten pool temperature and electrical parameters. Through real-time closed-loop adjustment, quality deviations caused by fluctuations in operating conditions can be corrected in a timely manner, ensuring the stability of the welding process.

[0102] Closed-loop self-healing process for post-weld quality defects:

[0103] When post-weld inspection confirms substandard quality, a self-healing process for welding parameters based on deep reinforcement learning is initiated, building upon the existing closed-loop process adjustments, to achieve iterative optimization of process knowledge. This process and the real-time parameter adjustments during welding belong to different levels of closed loops; the former targets iterative optimization at the process knowledge level, while the latter targets real-time process control of a single weld pass.

[0104] The closed-loop self-healing process for welding parameters based on deep reinforcement learning uses current working condition characteristics, real-time welding parameters, and quality non-compliance results as state inputs, constructs a reward function with welding quality compliance and process stability as quantitative indicators, and iteratively searches for a correction parameter scheme that adapts to the current working condition through a deep reinforcement learning agent.

[0105] The reward function of deep reinforcement learning consists of a quality reward term and a stability constraint term:

[0106]

[0107] Among them, quality awards Based on the quality prediction results corresponding to the parameters, a positive reward is given if the quality meets the standards, and a negative reward is given if the quality fails to meet the standards, depending on the degree of defect; stability constraint item. This constraint is used to limit the range of parameter adjustments, preventing drastic fluctuations from affecting process stability. Larger parameter adjustments result in higher penalties. Considering that online trial-and-error iterations are not feasible in industrial production scenarios, this process employs offline iterative optimization. A reinforcement learning agent is trained based on historically accumulated sample data from similar operating conditions, and the optimal parameter correction scheme is output through offline iteration, avoiding disruption to normal production.

[0108] After the correction parameter scheme output by the intelligent agent is verified to be effective through process verification, the correction parameters and corresponding working condition characteristics are synchronously updated to the welding process parameter library, replacing the original deviation parameters. At the same time, online incremental learning is used to optimize the parameters and update the welding parameter analysis model and quality prediction model for cross-scenario adaptation, so that the model learns the optimal parameter mapping relationship under the working condition and improves the processing accuracy of subsequent similar working conditions.

[0109] The incremental learning process employs a knowledge distillation mechanism to constrain model updates, and the loss function consists of both the new data loss and the distillation loss.

[0110]

[0111] In the formula, For the prediction loss of new operating condition data, The knowledge distillation loss is used to measure the difference between the updated model and the original model's output on historical samples. This represents the distillation weighting coefficient. This mechanism allows the model to learn new operating conditions while retaining its ability to handle existing conditions, thus avoiding catastrophic forgetting problems.

[0112] The model is updated periodically:

[0113] This method also includes a model iterative update step: based on the welding point cloud data, process parameters, and quality result data accumulated during the production process, each analysis model is periodically retrained and its parameters are iteratively updated. This periodic update mechanism complements the aforementioned quality-triggered online incremental update; the former ensures the overall generalization ability of the model, while the latter enables rapid adaptation to new working conditions.

[0114] Among them, the iterative training of the deep learning weld recognition model is based on the labeled weld point cloud data, with semantic segmentation accuracy as the core training indicator; the iterative training of the welding parameter analysis model and the quality prediction model is based on the full amount of process-quality paired data, with prediction accuracy as the core training indicator.

[0115] The welding process parameter library uses the collected workpiece surface point cloud features and the corresponding actual welding effect as driving data. Through an online incremental learning mechanism, the optimal parameters under new working conditions are analyzed and automatically updated into the library, continuously expanding the working condition coverage of the parameter library.

[0116] Production line efficiency analysis:

[0117] This method also includes a production line efficiency analysis step: performing multi-dimensional efficiency statistics and comparative analysis on the full operation data of the welding production line, generating corresponding production line efficiency optimization schemes, and outputting them to the production line scheduling system for production scheduling adjustments.

[0118] The multi-dimensional performance statistical analysis covers multiple statistical dimensions, including welding cycle time, product qualification rate, equipment utilization rate, and consumable consumption rate. By comparing with historical benchmarks and design indicators, bottlenecks in production line operation are identified. Targeted optimization solutions are generated for the identified bottlenecks to support production line scheduling and management decisions, thereby achieving continuous improvement in production line operational efficiency.

[0119] Example 2: Weld Identification and Adaptive Control System for Welding Parameters

[0120] like Figure 2 As shown, this embodiment provides a weld identification and adaptive welding parameter control system to implement the control method described in Embodiment 1. The system includes a data acquisition layer, an intelligent analysis layer, a decision output layer, and an execution control layer that are sequentially connected in communication. The specific composition and function of each layer are as follows:

[0121] Data Acquisition Layer:

[0122] The data acquisition layer serves as the system's data sensing input, responsible for acquiring 3D point cloud data of the workpiece to be welded, real-time welding process parameters, workpiece attributes, and on-site working condition data, providing raw data support for upper-level analysis. The data acquisition layer includes a 3D vision acquisition unit, a welding process sensing unit, and a working condition data acquisition unit.

[0123] The 3D vision acquisition unit is used to acquire 3D point cloud data of the surface of the workpiece to be welded, and outputs raw data containing information on bevel contours, surface morphology, and assembly gaps. This unit adopts a line laser 3D scanning scheme, which has high scanning speed and accuracy, and can adapt to various workpiece surface conditions without the need for developer spraying, making it suitable for the complex industrial environment of welding workshops.

[0124] The welding process sensing unit is used to collect welding current, voltage, molten pool temperature and welding torch posture deviation signals fed back by the arc tracking system in real time. Various sensors have appropriate sampling frequency and response speed, which can accurately capture the dynamic parameter changes in the welding process and ensure the real-time performance of closed-loop control.

[0125] The working condition data acquisition unit is used to collect data on workpiece attributes, equipment operating status, production cycle time, and consumable consumption. It achieves the aggregation of multi-source working condition data by connecting to the workshop production management system, equipment controller, and environmental sensor network.

[0126] Intelligent Analysis Layer:

[0127] The intelligent analysis layer is the core computing and decision-making layer of the system. It incorporates a deep learning weld recognition model, a welding parameter analysis model, and a quality prediction model. It is equipped with a weld recognition deconstruction module, a welding parameter matching module, an online quality judgment module, and a pre-built welding process parameter library. It also includes a process knowledge iteration unit and a production line efficiency analysis unit to perform comprehensive analysis and computation on the collected data. The weld recognition deconstruction module incorporates a deep learning weld recognition model, performs multi-scale feature extraction on 3D point cloud data, completes point cloud preprocessing, weld semantic segmentation, and geometric parameter calculation, and outputs weld position coordinates, spatial attitude parameters, and groove type, groove angle, blunt edge size, root gap, and groove depth parameters, corresponding to all the functions of S2 in the method embodiment.

[0128] The welding parameter matching module has a built-in welding parameter analysis model and welding process parameter library. Based on bevel characteristic parameters, workpiece attribute data and on-site working condition data, it outputs a complete combination of process parameters and derivation basis that are adapted to the current working condition through multi-dimensional matching calculations, which corresponds to all the functions of S3 in the method embodiment.

[0129] The online quality judgment module has a built-in quality prediction model that integrates real-time welding process data and arc tracking feedback signals for time-series analysis and anomaly determination, and outputs quality level and anomaly warning signals, corresponding to the quality analysis function in S4 of the method embodiment.

[0130] The process knowledge iteration unit incorporates a deep reinforcement learning strategy optimization module and an incremental learning module. Based on actual welding effect feedback data, it continuously updates and optimizes the welding process parameter library, welding parameter analysis model, and quality prediction model. Simultaneously, it periodically iterates and optimizes the deep learning weld recognition model based on newly added labeled point cloud data, corresponding to the process self-healing and model iteration functions in the method embodiment. The production line efficiency analysis unit performs multi-dimensional statistical analysis on all production line operation data, identifies operational bottlenecks by benchmarking, and generates efficiency optimization solutions, corresponding to the production line efficiency analysis function in the method embodiment.

[0131] Decision output layer:

[0132] The decision output layer serves as the system's external interaction interface layer, responsible for structured output of the computational results from the intelligent analysis layer. Specific outputs include weld location information, optimal welding parameters, quality anomaly warning signals, and production line efficiency optimization schemes. This layer provides a standardized industrial communication interface, supporting multiple mainstream industrial bus protocols. It connects to the internal execution control layer to issue control commands, and also supports integration with external systems such as production line scheduling systems and production management systems, enabling data and command exchange.

[0133] Execution control layer:

[0134] The execution control layer is the system's execution terminal, used to control the welding equipment operation and adjust welding parameters based on the decision output results. This layer includes a welding robot controller, a welding power controller, an arc tracking execution unit, and a parameter adjustment execution module.

[0135] The welding robot controller receives weld position and posture parameters and robotic arm motion parameters, and controls the welding robotic arm to complete weld trajectory tracking and posture adjustment; the welding power controller receives process parameters such as current, voltage, and wire feed speed, and controls the energy output of the welding power source; the arc tracking execution unit receives arc tracking deviation signals and corrects the welding torch height and swing amplitude in real time through a servo mechanism to achieve real-time closed-loop posture control; the parameter adjustment execution module receives quality anomaly adjustment commands and dynamically corrects welding process parameters and motion parameters to achieve closed-loop quality control of the welding process.

[0136] The overall workflow of this system corresponds completely to the method described in Example 1: After the workpiece is loaded, the data acquisition layer completes the acquisition of three-dimensional point cloud and various working condition data; the intelligent analysis layer completes weld identification, parameter matching and quality analysis calculation; the decision output layer issues and outputs control commands and analysis results; the execution control layer drives the welding equipment to complete welding operations and real-time closed-loop control; post-weld quality data is fed back to the intelligent analysis layer to realize continuous iterative optimization of the process parameter library and model; the intelligent analysis layer periodically outputs production line efficiency optimization schemes to support production line management and scheduling optimization.

[0137] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for weld identification and adaptive control of welding parameters, characterized in that, Includes the following steps: S1. Collect 3D point cloud data of the workpiece to be welded, real-time running data of the welding process, and workpiece attribute and on-site working condition data; The real-time operating data of the welding process includes welding current, welding voltage, and molten pool temperature data; S2. Multi-scale feature extraction is performed on the 3D point cloud data of the workpiece using a deep learning weld recognition model to obtain the position coordinates, spatial orientation parameters and weld type information of the weld, and output the bevel geometric feature parameters. S3. Based on weld information, groove geometric feature parameters, workpiece material properties, workpiece thickness parameters and on-site working condition data, the welding parameter analysis model calls the pre-built welding process parameter library to perform multi-dimensional matching calculations, generate the optimal welding parameters that match the current welding conditions and output them. S4. Perform welding operations according to the output optimal welding parameters, and simultaneously perform online quality analysis and anomaly determination based on the real-time collected operation data during the welding process through a quality prediction model; when an anomaly is determined, the welding process parameters are adjusted in a closed loop.

2. The weld identification and adaptive control method for welding parameters according to claim 1, characterized in that, The three-dimensional point cloud data is acquired using a line laser three-dimensional scanning unit; the line laser three-dimensional scanning unit is mounted on a welding robotic arm or fixed at the front end of the welding station, with a scanning speed of no less than 2 million points / second, and there is no need to spray developer on the surface of the workpiece to be welded during scanning. The deep learning weld recognition model is trained by a multi-scene welding point cloud dataset, which includes point cloud samples of workpieces to be welded under complex lighting, workshop dust environment, and complex weld morphology. It covers scenarios such as rust, oil stains, oxide scale, spot welding residue, component deformation, bevel error, uneven gap, misalignment, burrs, and assembly deviation of large steel components. The deep learning weld seam recognition model has a single-frame point cloud recognition processing time of 10ms to 50ms and a weld seam position recognition output accuracy of 0.015 to 0.040mm.

3. The weld identification and adaptive control method for welding parameters according to claim 1, characterized in that, The bevel geometric feature parameters include bevel type, bevel angle, blunt edge dimension, root gap, and bevel depth; The welding process parameter library contains benchmark process parameters for various bevel types under different material grades, workpiece thickness ranges, and welding position conditions. The generated optimal welding parameter combination includes welding technique type, weld layer and pass number planning, robotic arm offset compensation value, robotic arm pitch angle attitude, welding machine wire feed speed, welding current and voltage matching value, welding speed, oscillation frequency and oscillation amplitude; The welding parameter analysis model outputs the optimal welding parameters and simultaneously provides explanations of the derivation basis for the corresponding parameters. The relative deviation between the output core welding parameters and the benchmark optimal parameters verified by standardized process tests under the same working conditions is no more than 10%.

4. The weld identification and adaptive control method for welding parameters according to claim 1, characterized in that, During the online quality analysis process, the quality prediction model performs time-series calculations on the real-time collected welding current, voltage, and molten pool temperature data; the quality prediction model has been trained and verified offline, and its welding quality grade prediction accuracy is no less than 85%. The welding arc tracking system operates in real time, continuously feeding back the welding torch posture deviation signal and arc length fluctuation characteristics to the quality prediction model, and autonomously correcting the welding torch height and oscillation amplitude in a closed loop. When a quality anomaly is detected, an early warning signal is triggered and the process parameter adjustment procedure is initiated. The process parameter adjustment procedure combines the characteristics of molten pool temperature and electrical parameter fluctuations to dynamically correct thermal input parameters such as welding speed, welding current, and wire feeding speed.

5. The weld identification and adaptive control method for welding parameters according to claim 4, characterized in that, If the quality is found to be unqualified after welding, a welding parameter closed-loop self-healing process based on deep reinforcement learning is initiated based on the process closed-loop adjustment. The deep reinforcement learning-based closed-loop self-healing process for welding parameters specifically includes: using current working condition characteristics, real-time welding parameters, and quality non-compliance results as state inputs; constructing a reward function using weld formation pass rate and thermal input stability as quantitative indicators; optimizing parameter strategies offline through a deep reinforcement learning agent based on historical non-compliance samples under the same working conditions; outputting a corrected parameter scheme adapted to the current working conditions; synchronously updating the verified corrected parameters and corresponding working condition characteristics to the welding process parameter library; and simultaneously optimizing and updating the welding parameter analysis model and quality prediction model across scenarios through online incremental learning.

6. The weld identification and adaptive control method for welding parameters according to claim 1, characterized in that, It also includes a regular model iteration update step: based on the labeled welding point cloud data, process parameters and quality result data accumulated during the production process, each analysis model is regularly retrained offline and the parameters are iterated and updated, which complements the online incremental update triggered by quality failure. The welding process parameter library uses the collected workpiece surface point cloud features and the corresponding actual welding effect as driving data. The welding parameter analysis model, which has been optimized by incremental learning, infers and analyzes the optimal parameters under new working conditions. After verification, the parameters are automatically added to the library and updated.

7. The weld identification and adaptive control method for welding parameters according to claim 1, characterized in that, It also includes a production line efficiency analysis step: performing multi-dimensional efficiency statistics on the full operation data of the welding production line, comparing and analyzing it with historical benchmarks and design indicators, generating corresponding production line efficiency optimization plans, and outputting them to the production line scheduling system to execute production scheduling adjustments; The multi-dimensional performance statistical analysis covers multiple statistical dimensions, including welding cycle time, product qualification rate, equipment utilization rate, and consumable loss rate.

8. A weld seam identification and adaptive welding parameter control system, characterized in that, It includes a data acquisition layer, an intelligent analysis layer, a decision output layer, and an execution control layer that are connected in sequence: The data acquisition layer is used to acquire three-dimensional point cloud data of the workpiece to be welded, real-time parameters of the welding process, workpiece attributes and on-site working condition data. The intelligent analysis layer has a built-in deep learning weld recognition model, welding parameter analysis model and quality prediction model, and is equipped with a weld recognition deconstruction module, a welding parameter matching module, an online quality discrimination module and a pre-built welding process parameter library to analyze and calculate the collected data. The decision output layer is used to output weld location information, optimal welding parameters, quality anomaly warning signals and production line efficiency optimization schemes, and supports integration with external production line scheduling systems. The execution control layer is used to control the operation of welding equipment and adjust welding parameters according to the decision output results, and to correct the welding torch position in real time through the arc tracking execution unit in a closed loop.

9. The weld identification and adaptive control system for welding parameters according to claim 8, characterized in that, The data acquisition layer includes a 3D vision acquisition unit, a welding process sensing unit, and a working condition data acquisition unit. The three-dimensional vision acquisition unit is used to acquire three-dimensional point cloud data of the surface of the workpiece to be welded, and output raw data containing information on bevel contour, surface morphology and assembly gap. The welding process sensing unit is used to collect welding current, voltage, molten pool temperature and welding torch posture deviation signals fed back by the arc tracking system in real time. The working condition data acquisition unit is used to collect data on workpiece attributes, equipment operating status, production cycle time, and consumable consumption.

10. The weld identification and adaptive welding parameter control system according to claim 8, characterized in that, The intelligent analysis layer is also equipped with a process knowledge iteration unit and a production line efficiency analysis unit; The weld identification and deconstruction module has a built-in deep learning weld identification model, which performs multi-scale feature extraction on three-dimensional point cloud data and outputs weld position coordinates, spatial attitude parameters, as well as groove type, groove angle, blunt edge size, root gap and groove depth parameters. The welding parameter matching module has a built-in welding parameter analysis model and welding process parameter library. Based on the bevel characteristic parameters, workpiece attribute data and on-site working condition data, it outputs a complete combination of process parameters that is adapted to the current working condition through multi-dimensional matching calculation. The online quality judgment module has a built-in quality prediction model that integrates real-time welding process data and arc tracking feedback signals for time-series analysis and anomaly determination. The process knowledge iteration unit has a built-in deep reinforcement learning strategy optimization module and incremental learning module. Based on the actual welding effect feedback data, it continuously updates and optimizes the welding process parameter library, welding parameter analysis model and quality prediction model, and regularly iterates and optimizes the deep learning weld recognition model based on the newly added labeled point cloud data. The production line efficiency analysis unit is used to perform multi-dimensional statistical analysis on the full operation data of the production line, generate efficiency optimization schemes, and output them to the production line scheduling system through the decision output layer.

Images