Artificial intelligence-based automatic welding robot and welding system thereof

Abstract

The application relates to the technical field of artificial intelligence welding, and discloses an automatic welding robot based on artificial intelligence and a welding system thereof. The system comprises four modules, namely a welding feature acquisition module, a process parameter generation module, a real-time regulation and control module and a motion planning module. The welding feature acquisition module acquires three-dimensional contour data of a workpiece to be welded, material composition information and weld geometric parameters, and generates a welding feature atlas through feature fusion; the process parameter generation module retrieves a matching template from a welding knowledge base according to the welding feature atlas, combines with an environment temperature and humidity compensation coefficient, and outputs a combination of reference welding process parameters; the real-time regulation and control module dynamically corrects the reference parameters to generate an optimization instruction according to the molten pool shape, heat radiation distribution and arc voiceprint features in the welding process; and the motion planning module calculates the motion trajectory, posture parameters and speed curve of a welding execution mechanism according to the optimization instruction, and forms a robot control instruction set. The system can improve the intelligent level of welding, guarantee the stability of welding quality, and is suitable for welding scenes in multiple manufacturing industries.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence welding technology, specifically to an automated welding robot and welding system based on artificial intelligence. Background Technology

[0002] In modern manufacturing, welding, as a key joining process, is widely used in various fields such as machinery manufacturing, automobile production, shipbuilding, and aerospace. However, with the increasing demands for product precision, quality stability, and production efficiency in industrial production, traditional welding methods have gradually revealed many shortcomings. Traditional welding operations heavily rely on the experience and skills of operators; differences in the technical levels of different operators directly lead to fluctuations in weld quality, especially when handling complex structural workpieces or workpieces made of special materials, making it difficult to guarantee weld consistency and reliability.

[0003] In the preparation stage before welding, traditional methods for obtaining the workpiece's three-dimensional contour, material composition, and weld geometry parameters often rely on manual measurement or single testing equipment. This results in low measurement efficiency and data accuracy is easily affected by environmental factors, making it difficult to form comprehensive and accurate workpiece welding characteristic information, which in turn affects the rationality of subsequent process parameter settings. In the process parameter determination stage, traditional methods usually rely on preset fixed parameter templates or adjust parameters through multiple trial welds. This not only consumes a lot of manpower and material costs but also cannot dynamically compensate for changes in real-time ambient temperature and humidity, leading to poor stability of weld quality under different environmental conditions.

[0004] During welding, traditional systems lack effective monitoring and analysis methods for key real-time characteristics such as molten pool morphology, heat radiation distribution, and arc sound patterns. This makes it difficult to detect anomalies and adjust process parameters in a timely manner, leading to weld defects such as porosity, cracks, and lack of fusion. Furthermore, the motion trajectory planning of welding actuators is often based on pre-set fixed paths, making it difficult to flexibly optimize motion trajectories, attitudes, and speeds according to dynamically adjusted process parameters, further limiting the improvement of welding accuracy and efficiency.

[0005] As the manufacturing industry transforms towards intelligence and automation, traditional welding technology can no longer meet the production demands for high precision, high efficiency, and high stability. There is a need for an automated welding system that can integrate multi-dimensional feature acquisition, intelligent process parameter generation, real-time dynamic control, and flexible motion planning to solve problems such as low efficiency, large quality fluctuations, and strong dependence on manual labor in traditional welding processes. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence-based automated welding robot and its welding system to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides an automated welding system based on artificial intelligence, the system comprising:

[0008] The welding feature acquisition module is used to acquire the three-dimensional contour data, material composition information and weld geometric parameters of the workpiece to be welded, and generate the workpiece welding feature map through feature fusion technology;

[0009] The process parameter generation module, based on the workpiece welding feature map, retrieves matching process parameter templates from the welding knowledge base, and outputs a combination of benchmark welding process parameters by combining the ambient temperature and humidity compensation coefficients.

[0010] The real-time control module dynamically corrects the combination of reference welding process parameters based on the molten pool morphology, heat radiation distribution, and arc acoustic characteristics during the welding process, and generates optimized welding process parameter instructions.

[0011] The motion planning module calculates the motion trajectory, attitude adjustment parameters, and speed control curve of the welding actuator based on the optimized welding process parameter instructions, forming a robot control instruction set.

[0012] Preferably, the process of generating the workpiece welding feature map includes:

[0013] The three-dimensional contour data is processed by surface segmentation, and the curvature features and normal vectors of each surface segment are extracted to establish a geometric feature matrix;

[0014] Analyze the element content ratio in the material composition information to calculate the material's welding performance index;

[0015] Measure the width variation rate and angle deviation values ​​of the weld geometry parameters to construct the weld feature vector;

[0016] The geometric feature matrix, material welding performance index, and weld feature vector are fused to generate a workpiece welding feature map.

[0017] Preferably, the process of determining the combination of reference welding process parameters includes:

[0018] Based on the material welding performance index in the workpiece welding feature map, process parameter templates in the welding knowledge base that meet the preset threshold of material matching degree are selected.

[0019] Based on the width change rate of the weld feature vector, adjust the current fluctuation range in the process parameter template;

[0020] By combining the ambient temperature and humidity compensation coefficients, the voltage reference value in the process parameter template is corrected to form a reference welding process parameter combination.

[0021] Preferably, the process of generating the optimized welding process parameter instructions includes:

[0022] The contour changes of the molten pool morphology are captured by a high-speed camera system, and the stability index of the molten pool is calculated.

[0023] Temperature gradient data of thermal radiation distribution are obtained using an infrared thermometry system to establish a model of the heat-affected zone.

[0024] Collect the energy distribution spectrum of electric arc acoustic signature features to identify abnormal electric arc states;

[0025] Based on the molten pool stability index, heat-affected zone model, and abnormal arc state, adjust the key parameters in the baseline welding process parameter combination.

[0026] Preferably, the calculation process of the molten pool stability index further includes:

[0027] Analyze the width fluctuation rate and length consistency in the characteristics of molten pool profile variation;

[0028] Extract the distribution pattern and amplitude characteristics of surface ripples in the molten pool;

[0029] The stability index of the molten pool is calculated by combining the width fluctuation rate, length consistency and surface ripple characteristics.

[0030] Preferably, the process of forming the robot control instruction set includes:

[0031] The welding speed requirement in the optimized welding process parameter command is analyzed and decomposed into axial movement component and radial compensation amount;

[0032] Based on the three-dimensional contour data in the workpiece welding feature map, plan the approach path of the welding actuator;

[0033] The attitude control parameters are dynamically adjusted based on the real-time changes in weld geometry parameters.

[0034] The motion component, approach path, and attitude parameters are integrated to form a robot control instruction set.

[0035] Preferably, the dynamic adjustment process of the attitude control parameters further includes:

[0036] Monitor the real-time width changes of weld geometry parameters;

[0037] Calculate the angular deviation between the welding torch axis and the weld centerline;

[0038] The compensation amount for attitude adjustment is determined based on the width change and the included angle deviation;

[0039] Update the pitch and deflection angle parameters of the welding actuator.

[0040] Preferably, the system further includes:

[0041] The quality prediction module establishes a quality risk prediction model based on the correlation between defect characteristics and process parameters in historical welding data.

[0042] The system analyzes the molten pool stability index and heat-affected zone model in real time. When the risk coefficient exceeds the set value, it triggers an emergency intervention command for process parameters.

[0043] Preferably, the process of establishing the quality risk prediction model includes:

[0044] Extract porosity distribution characteristics and non-fusion region characteristics from historical welding data;

[0045] Analyze the correspondence between defect characteristics and molten pool stability indices;

[0046] Establish a quality risk prediction model based on a multidimensional feature space.

[0047] Preferably, the present invention also includes an artificial intelligence-based automatic welding machine, the machine comprising all the modules and functions of the artificial intelligence-based automatic welding system described above.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] The welding feature acquisition module enables comprehensive acquisition of multi-dimensional information about the workpiece. Unlike traditional single detection or manual measurement methods, it uses feature fusion technology to integrate three-dimensional contour data, material composition information, and weld geometric parameters to generate a workpiece welding feature map. This can more comprehensively and accurately reflect the welding-related characteristics of the workpiece, providing more reliable basic information for subsequent process parameter setting. It effectively avoids deviations in process parameter setting caused by incomplete or inaccurate feature information, thereby improving the rationality of welding process design from the source.

[0050] In the process parameter generation stage, the system does not rely on fixed templates or repeated trial welding. Instead, it retrieves matching process parameter templates from the welding knowledge base based on the generated workpiece welding feature map, and outputs a baseline welding process parameter combination by combining environmental temperature and humidity compensation coefficients. This process makes full use of the large amount of welding experience data accumulated in the knowledge base, while taking into account the impact of environmental factors on the welding process. This makes the output baseline process parameters more in line with the actual welding scenario requirements, reduces welding quality fluctuations caused by environmental changes, eliminates the need for parameter adjustments through multiple trial welding, significantly reduces manpower and material consumption, and shortens process preparation time.

[0051] The real-time control module enables the system to continuously monitor the molten pool morphology, heat radiation distribution, and arc acoustic signature during welding. By analyzing this real-time data, subtle changes or potential anomalies in the welding process can be detected promptly. This allows for dynamic correction of the baseline welding process parameter combination, generating optimized welding process parameter instructions. Compared to traditional welding processes that lack real-time monitoring and adjustment, this module effectively avoids weld defects caused by various uncertainties during welding, ensuring a stable and ideal welding process and improving the consistency and reliability of weld quality.

[0052] Based on optimized welding process parameters, the motion planning module accurately calculates the motion trajectory, posture adjustment parameters, and speed control curve of the welding actuator, forming a robot control instruction set. This motion planning method based on dynamically optimized process parameters breaks through the limitations of traditional fixed path planning, enabling the motion of the welding actuator to match real-time process parameters. Whether welding complex workpieces or in scenarios requiring adjustments to welding speed and posture to adapt to changes in process parameters, it achieves more precise welding operations, further improving welding accuracy. Simultaneously, it avoids reduced welding efficiency or quality issues caused by mismatches between motion trajectory and process parameters. Overall, it promotes welding operations towards higher efficiency, higher quality, and lower cost, reducing reliance on operator experience and minimizing human interference in the welding process, thus better meeting the intelligent and automated development needs of modern manufacturing. Attached Figure Description

[0053] Figure 1 This is a timing diagram of the artificial intelligence-based automatic welding system described in this invention;

[0054] Figure 2 A flowchart for generating welding feature maps of a workpiece;

[0055] Figure 3 A flowchart for optimizing the generation of welding process parameter instructions. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 The present invention provides an automatic welding system based on artificial intelligence, the system comprising: a welding feature acquisition module, a process parameter generation module, a real-time control module, and a motion planning module.

[0058] The welding feature acquisition module is responsible for acquiring the 3D contour data, material composition information, and weld geometry parameters of the workpiece to be welded, and generating a workpiece welding feature atlas using feature fusion technology. This atlas comprehensively reflects the geometric characteristics, material properties, and weld morphology of the workpiece. The process parameter generation module, based on the workpiece welding feature atlas, retrieves matching process parameter templates from a pre-built welding knowledge base, and, combined with real-time ambient temperature and humidity compensation coefficients, outputs a baseline welding process parameter combination. This combination provides initial parameter settings for the welding process. The real-time control module dynamically monitors the molten pool morphology, heat radiation distribution, and arc acoustic signature during welding. By analyzing this real-time data, it corrects the baseline welding process parameter combination and generates optimized welding process parameter instructions, ensuring the stability of welding quality. The motion planning module, based on the optimized welding process parameter instructions, calculates the motion trajectory, attitude adjustment parameters, and speed control curve of the welding actuator, ultimately forming a robot control instruction set to drive the welding robot to complete precise operations. All modules are interconnected via a data bus to achieve real-time information sharing and collaborative control. The system adopts a hierarchical processing architecture: the bottom layer uses sensors to collect raw data, the middle layer uses artificial intelligence algorithms for feature extraction and decision-making, and the upper layer uses control units to execute actions and output results.

[0059] Example 1: See Figure 2 The generation process of the workpiece welding feature map begins with the precise capture of the workpiece's three-dimensional spatial information. A high-precision laser 3D scanner is used to perform a 360-degree scan of the workpiece to be welded, which is stationary on the welding fixture. The laser beam emitted by the scanner forms a light band on the workpiece surface. The image of the deformed light band is captured by a binocular vision sensor, and a point cloud reconstruction algorithm is used to generate dense 3D point cloud data. This data completely records the spatial coordinate information of the workpiece surface. In the point cloud data preprocessing stage, a statistical filtering algorithm is first used to remove outlier noise points. Then, voxel mesh downsampling is used to reduce the amount of data while maintaining the model accuracy, creating favorable conditions for subsequent surface segmentation processing. The surface segmentation processing adopts a region growing algorithm based on curvature features. This algorithm selects seed points from the point cloud and expands the region based on the normal vector and curvature continuity between adjacent points. Each generated surface segment is assigned an independent identifier. Its curvature features are obtained by calculating the Gaussian curvature and average curvature value of all points in the segment, and the normal vector is determined by fitting a local tangent plane through principal component analysis. The geometric feature matrix is ​​constructed using surface segments as basic units. The rows of the matrix correspond to different segments, while the columns contain geometric attributes such as the average curvature, rate of change of curvature, azimuth and tilt angle of the normal vector of the segment, forming a multi-dimensional numerical matrix for quantitatively describing the geometric morphological features of the workpiece surface.

[0060] Material composition information is obtained using a mobile laser-induced breakdown spectroscopy (LASPS) analyzer. Driven by a robotic arm, the analyzer approaches the workpiece's welding area and emits laser pulses to generate plasma on the material surface. The content ratios of elements such as carbon, silicon, manganese, chromium, and nickel in the material are determined by analyzing the wavelength and intensity of the plasma emission spectrum. The calculation of the material's weldability index comprehensively considers the influence of each element on weld hot cracking sensitivity, hardening tendency, and toughness. A multi-factor weighted model is used, with carbon equivalent calculation as the basis, and corrections made based on the content of harmful elements such as phosphorus and sulfur. The final result is a standardized index ranging from zero to one; a higher index value indicates better weldability. The measurement of weld geometry parameters is accomplished using an industrial digital camera equipped with a high-resolution lens. The camera acquires two-dimensional images of the weld area from multiple angles, and the three-dimensional morphology of the weld is reconstructed using a stereo vision algorithm. The measured parameters include the width, depth, and angle of the weld bevel, as well as the variation of these parameters along the weld length. The width variation rate is calculated using the sliding window method. Width values ​​are measured at fixed intervals along the weld length, and the standard deviation of the width variation percentage between adjacent measurement points is then calculated. The angle deviation is determined by comparing the absolute difference between the actual measured weld bevel angle and the design standard angle. The construction of the weld feature vector includes statistical characteristics of the width sequence, such as mean, variance, and skewness, as well as the distribution characteristics of the angle deviation, forming a numerical sequence describing the continuous change in weld geometry.

[0061] The feature fusion process employs a multilayer perceptron neural network structure, which includes an input layer, multiple hidden layers, and an output layer. The geometric feature matrix, after flattening, is fed into the network input layer. The material welding performance index serves as a scalar input, while the weld feature vector is input synchronously after standardization. In the hidden layers, the network performs nonlinear transformations and weighted combinations on features from different sources. Nonlinear factors are introduced through activation functions, enabling the system to learn the complex relationships between different features. The trained network maps heterogeneous input features to a unified low-dimensional feature space, generating a workpiece welding feature map. This workpiece welding feature map is transmitted and stored within the system as a feature tensor. Each dimension of the tensor corresponds to different feature types and spatial location information. The map not only contains static features but also records the distribution of features at different weld locations, providing comprehensive and detailed input for subsequent process parameter decisions. The entire generation process adopts a pipelined architecture, with data buffers between each processing stage to ensure data processing continuity and real-time performance even when processing time fluctuates at certain stages. The feature fusion stage also incorporates an attention mechanism, enabling the network to adaptively focus on feature dimensions that have a greater impact on welding quality, thereby enhancing the representational power of the feature map. The system periodically uses newly acquired workpiece data to incrementally train the feature fusion network, continuously optimizing its feature extraction capabilities over time and gradually adapting to more diverse workpiece types and welding conditions.

[0062] Example 2: The process of determining the combination of baseline welding process parameters is a decision-making process based on knowledge retrieval and dynamic correction. The welding knowledge base, as the core data resource of the system, stores a large amount of welding case data that has been verified in practice. Each case contains a complete process parameter template, which is categorized by multi-level indexing according to characteristics such as material type, thickness range, and weld type, forming a structured parameter database. The knowledge base adopts a distributed architecture for storage, supporting fast parallel retrieval. Upon receiving the workpiece welding feature map from the feature acquisition module, the system initiates a multi-condition matching retrieval process.

[0063] The retrieval process first focuses on calculating the material matching degree. The system extracts the material welding performance index from the workpiece welding feature map and compares it with the material feature data recorded in each template in the knowledge base. The similarity algorithm adopts a feature-weighted distance metric, comprehensively considering the similarity of multiple indicators such as material chemical composition and mechanical properties. The system will filter out all candidate templates whose material matching degree exceeds a preset threshold. This threshold is usually dynamically adjusted according to welding quality requirements, and a higher matching threshold is set for high-requirement welding applications. The candidate templates are sorted in descending order of matching degree score to form a preliminary set of candidate parameters. Next, the system enters the parameter adjustment stage, where the width change rate in the weld feature vector becomes the key adjustment basis. The process parameter template originally contained a standard current fluctuation range, which now needs to be adjusted according to the actual geometric characteristics of the weld. When a large weld width change rate is detected, it indicates that there is a significant fluctuation in the weld gap along the welding direction. The system will correspondingly expand the allowable fluctuation range of the current. This adjustment is achieved by looking up a preset parameter adjustment rule table. The rule table records the current adjustment coefficients corresponding to different width change rate ranges. These coefficients are empirical values ​​derived from a large amount of process test data. The adjustment process keeps the current reference value unchanged and mainly modifies its upper and lower fluctuation ranges.

[0064] The environmental compensation process involves real-time data collection from temperature and humidity sensors installed within the welding workstation. Temperature readings are accurate to 0.1 degrees Celsius, and humidity measurements are accurate to ±2%RH. These environmental parameters are converted into temperature and humidity compensation coefficients. Correction of the voltage reference value is the core of environmental compensation. The process parameter templates in the knowledge base contain voltage reference values ​​under standard environmental conditions, which now need to be corrected based on actual environmental conditions. The temperature compensation coefficient primarily affects arc stability; as the ambient temperature increases, the voltage required to maintain a stable arc decreases accordingly. Humidity compensation focuses on the shielding gas effect; as humidity increases, the voltage needs to be appropriately increased to compensate for changes in shielding gas density. The two compensation coefficients are combined through a multiplicative model to affect the voltage reference value. The final stage of forming the reference welding process parameter combination requires consistency verification of all parameters. The system checks the logical relationships between parameters to ensure their rationality, such as the matching of welding speed and heat input, and the correspondence between wire feed speed and current value. After successful verification, the system generates a complete reference welding process parameter combination document. This document records all necessary parameters, including current range, voltage value, welding speed, wire feed speed, and shielding gas flow rate, in a standardized format. The parameter combination also includes a confidence score, which is derived from the combined confidence of the matching degree calculation and environmental compensation, providing a reference for the parameter adjustment range of the subsequent real-time control module.

[0065] The knowledge base's update and maintenance mechanism is crucial for continuous system optimization. After each welding task is completed, the system stores the actual process parameters used and the welding quality assessment results as new cases in the knowledge base. New cases undergo quality review before being added; only those with satisfactory welding quality are adopted. The knowledge base also has a self-learning function, analyzing parameter patterns in successful cases and automatically optimizing coefficient settings in the parameter adjustment rule table, thus continuously improving the system's decision-making capabilities over time. For welding tasks involving special materials or novel joint types, the system supports manual import of verified process parameter templates to enrich the knowledge base's coverage. The entire process of determining the baseline welding process parameter combination employs a fault-tolerant design. When the number of retrieved candidate templates is insufficient, the system gradually relaxes the matching threshold, ensuring a feasible parameter solution is provided under any circumstances. Multiple safety checks are implemented during parameter determination to prevent obviously unreasonable parameter combinations. All parameter adjustments are performed within preset safety ranges to avoid damage to welding equipment from extreme parameters. The system records detailed logs for each parameter decision, including complete information such as search conditions, matching results, and adjustment processes. These logs are used for subsequent process analysis and system optimization. The output of the baseline welding process parameter combination adopts a hierarchical structure, with core parameters as the main output and derived and auxiliary parameters as supplementary information output simultaneously. All parameters are labeled with their data source and adjustment basis, facilitating subsequent modules' understanding of the logical foundation for parameter decisions. The parameter combination also includes version control information; when the knowledge base is updated, the system marks the version of the knowledge base on which the parameter combination is based, maintaining the traceability of process parameters. The entire determination process is completed within seconds, meeting the high efficiency requirements of modern automated welding for process preparation.

[0066] Example 3: See Figure 3The process of generating optimized welding process parameters is based on multi-sensor information fusion and dynamic analysis. A high-speed camera system continuously acquires image sequences of the weld pool area at a fixed frame rate. The image acquisition process is complemented by filters of specific wavelengths to suppress interference from the strong arc light. After preprocessing, each frame enters the contour extraction stage, where an edge detection algorithm identifies the boundary between the weld pool and the base material. The analysis of weld pool contour variation characteristics focuses on the temporal evolution of geometric morphology. The instantaneous fluctuation of the weld pool width is calculated by comparing the positional changes of contour points between consecutive frames. The evaluation of length consistency is achieved by measuring the change in distance from the tail of the weld pool to the center of the arc. These dynamic parameters collectively reflect the stability of the weld pool. The quantification of weld pool stability indicators requires the integration of multiple feature parameters. The system uses a sliding time window statistical method to calculate the ratio of the standard deviation to the average value of the weld pool width within a set time segment. The evaluation of length consistency is achieved by calculating the coefficient of variation of the weld pool length, which reflects the dispersion of the length value. The feature extraction of surface ripples in the molten pool is achieved by using image texture analysis technology. After frequency domain transformation of the molten pool area, the distribution pattern of the ripples is analyzed. The amplitude features are obtained by measuring the height difference between the peak and valley values ​​of the ripples. These feature parameters are combined by weighting to form a comprehensive molten pool stability index, where the weight coefficients of each feature are trained based on historical welding quality data.

[0067] The infrared temperature measurement system was carefully deployed, with multiple infrared sensors installed around the welding area at specific angles, forming a three-dimensional monitoring network for the thermal field. Temperature data collected by the sensors underwent spatial interpolation to generate a thermal radiation distribution map of the welding area. Temperature gradient data extraction was based on isotherm distribution in the thermal image, quantifying areas of concentrated heat by calculating the change in spacing between adjacent isotherms. The heat-affected zone (HAZ) model was established using basic heat transfer principles, combined with the material's thermal conductivity to predict the thermal cycling curve. This model can simulate the heat conduction process in the workpiece and predict the width and microstructure changes of the HAZ under different welding parameters. Arc acoustic signature features were acquired using a directional microphone array, the array arrangement taking into account the sound wave propagation path and environmental noise shielding. The acquired audio signal was first bandpass filtered to remove low-frequency mechanical noise and high-frequency interference, and then processed in frames. Energy distribution spectrum analysis employed a short-time Fourier transform method to convert the time-domain signal into a frequency-domain representation, and a peak detection algorithm was used to identify characteristic frequency components. The identification of abnormal arc states was based on pattern matching of the acoustic signature spectrum, comparing the real-time acoustic signature with typical abnormal acoustic signatures in the knowledge base. The threshold for judging abnormal conditions is adaptively adjusted based on the welding material and the type of shielding gas.

[0068] The calculation process for the molten pool stability index incorporates a comprehensive evaluation model that considers the interaction relationships of multiple characteristic parameters:

[0069] ,

[0070] in: The final calculation result represents the stability index of the molten pool; the higher the value, the better the stability. The volatility of the melt pool width is derived from the standard deviation of the width values ​​within a statistical time window. This represents the consistency coefficient of the molten pool length, reflecting the degree of variation in the length value; This represents the characteristic parameter of surface ripple amplitude, which is the average amplitude value of the ripple obtained through image analysis; , , The weighting coefficients for the three characteristics were determined through regression analysis of extensive process test data. This model comprehensively reflects the geometric stability characteristics of the molten pool, providing a quantitative basis for adjusting process parameters.

[0071] The adjustment strategy for the baseline welding process parameter combination employs a multi-level decision-making mechanism. The system first determines the urgency level of the adjustment based on the deviation of the molten pool stability index. When the index indicates the molten pool is in a critically unstable state, the system prioritizes adjusting the current parameter, as current has the most direct impact on molten pool stability. The output of the heat-affected zone (HAZ) model guides fine-tuning of the heat input. When the model predicts the HAZ width exceeds the allowable range, the system appropriately reduces the welding speed or adjusts the voltage. The detection results of abnormal arc states are primarily used for protective adjustments. When arc drift or short-circuit tendency is identified, the system immediately fine-tunes the combination of voltage and wire feed speed. Coordinated control during parameter adjustment is crucial. The system maintains a parameter influence matrix that describes the coupling relationships between various process parameters. When adjusting a parameter, the system simultaneously adjusts other strongly correlated parameters to avoid introducing new instability factors due to single-parameter adjustments. All parameter adjustments are performed within preset safety boundaries, with the adjustment range dynamically weighted based on the reliability of real-time monitoring data. High-confidence sensor data corresponds to larger adjustment ranges, while data with higher noise levels corresponds to more conservative adjustments.

[0072] The frequency of welding process parameter generation commands is matched to the welding speed, with a higher command update rate used during high-speed welding to ensure the system can respond promptly to changes in process conditions. Each optimization command includes a rationale for parameter adjustment, recording which sensor(s) data anomaly triggered the adjustment. Before outputting commands, a rationality check is performed to verify whether the adjusted parameters are within the equipment's allowable operating range and whether the parameter combination conforms to the fundamental principles of welding metallurgy. The entire optimization process forms a complete closed-loop control, with real-time monitoring, analysis, decision-making, and execution tightly integrated, maintaining the optimal state of the welding process through continuous parameter fine-tuning. The system's learning mechanism records the effect of each parameter adjustment, evaluating the effectiveness of the adjustment strategy by comparing changes in molten pool stability indicators before and after the adjustment. This feedback data is used to periodically optimize the rule base, allowing the system's decision-making capabilities to continuously evolve with accumulated experience. For welding processes involving special materials, the system supports importing expert experience rules, which work in conjunction with the automatic optimization algorithm to ensure the generation of reasonable optimization commands under various operating conditions.

[0073] Example 4: The formation process of the robot control instruction set can be clearly demonstrated through a specific welding case. Assume the system is processing a longitudinal seam welding task for a large pressure vessel cylinder. The workpiece is made of low-alloy high-strength steel with a wall thickness of 25 mm and a weld length of approximately 3 meters. The welding actuator is a six-axis articulated robot with a water-cooled welding torch at its end. The optimized welding process parameters require the welding speed to be dynamically adjusted between 350 mm / min and 420 mm / min while maintaining a constant heat input. The system first analyzes these speed requirements, decomposing them into an axial movement component along the weld direction and a radial compensation amount perpendicular to the workpiece surface. The axial movement component is calculated based on the welding speed reference value; the system converts the 350 mm / min speed into the angular velocity of each joint of the robot. The determination of the radial compensation amount needs to consider the weld height changes caused by workpiece assembly errors and thermal deformation. The compensation amount is dynamically adjusted by measuring the distance between the welding torch and the workpiece surface in real time using a laser displacement sensor. The approach path planning adopts a three-dimensional spatial interpolation algorithm. The robot controller generates a smooth approach trajectory based on the three-dimensional contour data in the workpiece welding feature map. This trajectory ensures that the welding torch approaches the weld start point from the starting point in the optimal posture, avoiding interference with the tooling fixture.

[0074] The dynamic adjustment of attitude control parameters is achieved through a real-time monitoring system. A vision sensor mounted on the side of the welding torch acquires weld images at a rate of 50 frames per second, and an image processing algorithm calculates the changes in weld width in real time. When a 10% increase in weld width is detected, the system adjusts the welding torch's oscillation amplitude and dwell time accordingly. The angular deviation between the welding torch axis and the weld centerline is measured using a binocular vision system. Two cameras capture images of the weld area from different angles, and the relative positional relationship in three-dimensional space is calculated using triangulation. Based on the measurement results of width changes and angular deviations, the system determines the compensation amount for attitude adjustment. The welding actuator's attitude parameters are updated using a progressive adjustment strategy. The pitch angle adjustment follows a cosine function curve to ensure a smooth transition in attitude changes; the deflection angle adjustment is synchronized with the welding speed, using a smaller adjustment range in high-speed welding sections and allowing a larger adjustment range in low-speed sections. Throughout the adjustment process, the system continuously monitors the robot's joint torque and motor current, automatically reducing the adjustment range when abnormal loads are detected to prevent overload. Table 1 shows the attitude parameter adjustment records for a specific time segment during the welding process.

[0075] Table 1: Record of Dynamic Adjustment of Welding Posture Parameters

[0076]

[0077] The motion trajectory is generated using spline curve interpolation. The system sets a series of key points on the weld path and connects these points with cubic spline curves to ensure the smoothness and continuity of the robot's motion. The attitude parameters of each key point are optimized to keep the welding torch at the optimal working angle. The speed control curve is planned considering the robot's dynamic characteristics, using uniform motion in straight sections and appropriately reducing the speed in curved sections to avoid vibration caused by centrifugal force.

[0078] The final integration process of the robot control instruction set involves coordinate transformation and kinematic calculation. The system transforms the global coordinate system of the welding path to the robot's base coordinate system and converts the pose of the end effector into the angle values ​​of the six joints through inverse kinematics calculation. The motion trajectory of each joint is processed through velocity planning and acceleration limiting to generate a smooth joint space trajectory. The instruction set adopts a hierarchical structure: high-level instructions describe the welding task and path planning, mid-level instructions define the motion trajectory and posture sequence, and low-level instructions directly control the pulse output of the servo drivers. The collision avoidance algorithm runs continuously during instruction generation. The system calculates the distance field between the robotic arm and the surrounding environment and adjusts the trajectory in real time to avoid interference. When a potential collision is predicted, the system automatically replans the path to ensure the safe progress of the welding process. The instruction set also includes a fault-tolerant processing mechanism. When a sensor data is abnormal, the system can switch to backup control mode and continue to complete the welding task using historical data or model predictions. Real-time corrections during the welding process are achieved through closed-loop control. The laser tracking system continuously monitors the deviation between the actual position of the welding torch and the planned trajectory. When the deviation exceeds a set threshold, the system immediately generates a trajectory correction command. These corrections are smoothly integrated into the subsequent motion trajectory through the interpolation algorithm of the robot controller, avoiding sudden position jumps. The response time of the entire control system is controlled within ten milliseconds, ensuring timely compensation for various external disturbances.

[0079] The instruction set optimization also considers energy consumption. The system selects the lowest energy-consuming motion trajectory to reduce the robot's energy consumption while ensuring welding quality. The trajectory planning algorithm avoids unnecessary acceleration and deceleration processes, maintaining smooth motion. After each welding task is completed, the system analyzes the difference between the actual motion data and the planned trajectory, using this data to optimize the trajectory planning parameters for subsequent tasks and continuously improve control accuracy. The robot control instruction set is verified in a virtual simulation environment. The system uses a robot dynamics model to simulate the actual motion process, detecting any singularities or out-of-limit positions. Only after the simulation passes will the instruction set be downloaded to the actual robot controller for execution. The entire instruction generation process maintains strict timing synchronization, with precise coordination of all motion axes to ensure stable and reliable welding quality. The system fully records the execution data of each welding task, including the actual trajectory, attitude parameters, and sensor readings. This data is used for subsequent process analysis and system optimization.

[0080] Example 5: The implementation of the quality prediction module uses the circumferential welding of a storage tank as a specific application scenario. The tank is made of austenitic stainless steel, with a diameter of four meters and a plate thickness of twelve millimeters. The welding process adopts a double-wire submerged arc welding process, and the welding speed is controlled at about 400 millimeters per minute. The historical welding database stores welding records of similar structures completed in the past three years, totaling about two hundred complete cases. Each case includes process parameter records, non-destructive testing results, and mechanical property test data. After the module is started, it first retrieves these historical data from the database for preprocessing and feature extraction. The extraction of pore distribution features is based on the digital analysis of X-ray inspection films. The system uses image processing algorithms to identify pore images on the films and measures the diameter and position coordinates of each pore. Feature parameters include the number of pores per unit area, the distribution pattern of pore diameter, and the pore aggregation index. The acquisition of features of the unfused area relies on ultrasonic phased array detection data. The system analyzes the signal features in the B-scan images to identify the size and orientation of the unfused defects. Feature parameters include the proportion of unfused area, defect length, depth location, and the angle between the defect orientation and the weld centerline. These defect features are correlated with corresponding process parameters to form a training sample set.

[0081] The correlation analysis between defect features and molten pool stability indices employs a time-series matching method. The system spatially correlates the time curves of molten pool stability indices recorded during welding with the locations of subsequently detected defects. By analyzing the variation patterns of molten pool stability indices near the defect occurrence time point, correlation patterns are identified. Analysis reveals that when molten pool stability indices fluctuate drastically within a specific time period, the probability of porosity in that area increases significantly; and non-fusion defects are often associated with abnormal temperature gradients in the thermal radiation distribution. The quality risk prediction model based on a multi-dimensional feature space utilizes a deep learning architecture. The model input layer contains fifteen feature parameters, including real-time molten pool stability indices, the maximum temperature gradient in the heat-affected zone, and an abnormal arc sound pattern index. The three hidden layers contain thirty-two, sixteen, and eight neurons respectively, and the output layer generates a quality risk coefficient between zero and one. Model training uses 150 cases from historical data, with the remaining fifty cases used to validate the model's prediction accuracy. The training process employs an adaptive moment estimation algorithm, and the loss function comprehensively considers classification accuracy and generalization ability. During real-time risk monitoring, the system collects current welding status data ten times per second, including molten pool stability indicators and heat-affected zone temperature distribution. Each data point, after standardization, is input into a trained quality risk prediction model, which outputs the risk coefficient of the current welding point in real time. The risk coefficient calculation considers a weighted combination of multiple feature parameters, with the molten pool stability indicator having the highest weight, followed by the heat-affected zone temperature gradient, and arc acoustic signature features serving as an auxiliary judgment criterion. The system sets a dynamic risk threshold, which is adjusted according to the importance level of the weld, with stricter control standards applied to critical load-bearing welds.

[0082] When the risk coefficient exceeds the set value, the system triggers an emergency intervention command for process parameters. The intervention command generation employs a tiered response mechanism, implementing intervention measures of varying intensities based on the magnitude and duration of the risk coefficient exceeding the limit. Level 1 intervention addresses minor exceedances, automatically adjusting welding parameters for compensation, such as fine-tuning welding current or voltage. Level 2 intervention addresses persistent exceedances, reducing welding speed and simultaneously adjusting heat input parameters. Level 3 intervention addresses severe exceedances, immediately halting the welding process and prompting operator intervention. The execution of emergency intervention commands follows a gradual strategy to avoid secondary interference to the welding process from sudden parameter changes. After the command is issued, the system continuously monitors the risk coefficient's changing trend; if the indicators do not improve, higher-level intervention measures are initiated. All intervention actions are recorded in the process log, including intervention time, intervention type, parameter adjustment amount, and post-intervention effect evaluation; this data is used for subsequent analysis of the effectiveness of the intervention strategy.

[0083] The quality prediction module's self-learning function is implemented through an online update mechanism. After each welding task is completed, the system compares the actual inspection results with the predicted results. When a large prediction deviation is found, the case is automatically marked. After accumulating a certain number of new cases, the system initiates a model parameter optimization process, using new data to fine-tune the prediction model. The model update process uses an incremental learning algorithm, incorporating new experience while maintaining existing knowledge, allowing the prediction capability to continuously evolve over time. The module's reliability is ensured through a multi-verification mechanism. The prediction results for important welds are cross-validated by two independent models. Intervention is only taken when the prediction conclusions of the two models are consistent. For uncertainties in the prediction results, the system adopts a conservative strategy, preferring to falsely report rather than miss serious defects. The visualization interface of the prediction results displays the risk coefficient change curve in real time and marks high-risk sections with colors, allowing operators to intuitively understand the welding quality status. Historical data management uses a version control method. When welding processes or material specifications change, the system creates new data branches to avoid data interference between different process conditions. Data cleaning is performed every six months to remove invalid data and outliers, maintaining database quality. The module also supports expert experience import, allowing senior engineers to adjust risk assessment rules based on practical experience. These manual adjustments are stored as special cases in the knowledge base, enriching the system's decision-making basis. Collaboration with other modules is achieved through a data bus. The quality prediction module receives molten pool monitoring data from the real-time control module in real time and feeds back the prediction results to the process parameter generation module. When a quality risk is predicted, the module sends a warning signal to the motion planning module, indicating that the welding path or speed may need to be adjusted. The entire quality prediction process forms a complete monitoring-prediction-intervention closed loop, providing additional assurance for the stability of welding quality. All data generated during module operation is encrypted and stored to ensure the security and traceability of process data.

[0084] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automated welding system based on artificial intelligence, characterized in that, include: The welding feature acquisition module is used to acquire the three-dimensional contour data, material composition information and weld geometric parameters of the workpiece to be welded, and generate the workpiece welding feature map through feature fusion technology; The process parameter generation module, based on the workpiece welding feature map, retrieves matching process parameter templates from the welding knowledge base, and outputs a combination of benchmark welding process parameters by combining the ambient temperature and humidity compensation coefficients. The real-time control module dynamically corrects the combination of reference welding process parameters based on the molten pool morphology, heat radiation distribution, and arc acoustic characteristics during the welding process, and generates optimized welding process parameter instructions. The motion planning module calculates the motion trajectory, posture adjustment parameters, and speed control curve of the welding actuator based on the optimized welding process parameter instructions, forming a robot control instruction set. The process of generating the optimized welding process parameter instructions includes: The contour changes of the molten pool morphology are captured by a high-speed camera system, and the stability index of the molten pool is calculated. Temperature gradient data of thermal radiation distribution are obtained using an infrared thermometry system to establish a model of the heat-affected zone. Collect the energy distribution spectrum of electric arc acoustic signature features to identify abnormal electric arc states; Based on the molten pool stability index, heat-affected zone model, and abnormal arc state, adjust the key parameters in the baseline welding process parameter combination; The calculation process for the molten pool stability index also includes: Analyze the width fluctuation rate and length consistency in the characteristics of molten pool profile variation; Extract the distribution pattern and amplitude characteristics of surface ripples in the molten pool; The stability index of the molten pool is calculated by combining the width fluctuation rate, length consistency and surface ripple characteristics; The process of forming the robot control instruction set includes: The welding speed requirement in the optimized welding process parameter command is analyzed and decomposed into axial movement component and radial compensation amount; Based on the three-dimensional contour data in the workpiece welding feature map, plan the approach path of the welding actuator; The attitude control parameters are dynamically adjusted based on the real-time changes in weld geometry parameters. Integrate motion components, proximity paths, and attitude parameters to form a robot control instruction set; The dynamic adjustment process of the attitude control parameters also includes: Monitor the real-time width changes of weld geometry parameters; Calculate the angular deviation between the welding torch axis and the weld centerline; The compensation amount for attitude adjustment is determined based on the width change and the included angle deviation; Update the pitch and deflection angle parameters of the welding actuator.

2. The artificial intelligence-based automatic welding system according to claim 1, characterized in that, The process of generating the workpiece welding feature map includes: The three-dimensional contour data is processed by surface segmentation, and the curvature features and normal vectors of each surface segment are extracted to establish a geometric feature matrix; Analyze the element content ratio in the material composition information to calculate the material's welding performance index; Measure the width variation rate and angle deviation values ​​of the weld geometry parameters to construct the weld feature vector; The geometric feature matrix, material welding performance index, and weld feature vector are fused to generate a workpiece welding feature map.

3. The artificial intelligence-based automatic welding system according to claim 2, characterized in that, The process of determining the combination of the reference welding process parameters includes: Based on the material welding performance index in the workpiece welding feature map, process parameter templates in the welding knowledge base that meet the preset threshold of material matching degree are selected. Based on the width change rate of the weld feature vector, adjust the current fluctuation range in the process parameter template; By combining the ambient temperature and humidity compensation coefficients, the voltage reference value in the process parameter template is corrected to form a reference welding process parameter combination.

4. The artificial intelligence-based automatic welding system according to claim 1, characterized in that, Also includes: The quality prediction module establishes a quality risk prediction model based on the correlation between defect characteristics and process parameters in historical welding data. The system analyzes the molten pool stability index and heat-affected zone model in real time. When the risk coefficient exceeds the set value, it triggers an emergency intervention command for process parameters.

5. The artificial intelligence-based automatic welding system according to claim 4, characterized in that, The process of establishing the quality risk prediction model includes: Extract porosity distribution characteristics and non-fusion region characteristics from historical welding data; Analyze the correspondence between defect characteristics and molten pool stability indices; Establish a quality risk prediction model based on a multidimensional feature space.

6. An automated welding machine based on artificial intelligence, characterized in that, It includes all modules and functions of the AI-based automated welding system as described in any one of claims 1 to 5.

Images