An online identification and control method and quality control system for welding defects

Abstract

The present application relates to the technical field of welding quality control, and proposes an online identification and control method for welding defects. The method comprises the following steps: S1, constructing a pre-welding process benchmark containing workpiece state information; S2, during the welding process, repairing the multi-source process data when it is abnormal based on the pre-welding process benchmark to generate a continuous time series data set; S3, analyzing the time series data set based on an embedded intelligent model, identifying welding defects and determining the deviation root causing the defects; S4, generating an adaptive process adjustment strategy according to the deviation root; S5, optimizing and deciding the process adjustment strategy based on a multi-objective optimization algorithm, and outputting the final process parameters to drive the welding execution, while updating the embedded intelligent model. The present application realizes full-process adaptive closed-loop control of welding quality, improves the accuracy of process adjustment through root determination, and endows the manufacturing system with the ability of continuous self-optimization.

Description

Technical Field

[0001] This invention relates to the field of welding quality control technology, and in particular to an online identification and control method and quality control system for welding defects. Background Technology

[0002] Welding, as a crucial material joining process, is widely used in key fields such as aerospace, automotive manufacturing, and energy equipment. During the welding process, various factors, including material properties, process parameters, and environmental disturbances, can easily lead to defects such as porosity, cracks, lack of fusion, and undercut, severely impacting the mechanical properties, sealing performance, and service safety of the welded structure. Therefore, achieving accurate online identification and real-time adaptive control of welding defects is a core challenge for improving welding quality and intelligence.

[0003] Existing technologies include numerous solutions for online monitoring and quality control of the welding process. For example, patent CN121061318A discloses an ultrasonic welding quality control system and method, which collects data through multiple sensors, extracts features, establishes correlations, and dynamically adjusts parameters. Patent CN120791072A proposes a method for monitoring the quality of surfacing welding, utilizing edge computing nodes to run a predictive model and adjusting welding parameters in real time based on quality assessment results. Patent CN121715731A relates to an intelligent control method and system for online monitoring of the quality of laser welding of busbar profiles, which constructs a welding feature matrix, analyzes the correlation strength of influencing factors, and generates control strategies. Patent CN121564528A provides an online detection method for welding defects based on molten pool image recognition, which can accurately identify defects and establish a data link between defect identification information and the production management system.

[0004] However, the aforementioned existing technical solutions still have the following limitations: First, most lack accurate modeling and benchmark construction of the individualized state of the workpiece before welding (such as geometric morphology, surface foreign matter, and material properties), resulting in insufficient benchmark adaptability for monitoring and control. Second, in the real-time acquisition of multi-source heterogeneous sensor data, the ability to repair and reconstruct missing or abnormal data is lacking, affecting the continuity and reliability of subsequent analysis. Third, defect identification mostly stops at type and location judgment, failing to deeply quantify the multiple root causes of defects and their causal contributions, resulting in a lack of accurate decision-making basis for subsequent process adjustments. In addition, process adjustment strategies are often relatively simple or static, failing to perform non-uniform, adaptive parameter field optimization in the three-dimensional space of the weld based on the root causes of defects, and lacking collaborative optimization under multiple objectives (such as quality, efficiency, and energy consumption). Finally, the systems generally lack the ability to continuously, safely, and collaboratively evolve diagnostic and decision-making models using real-time data, making it difficult to adapt to complex and changing working conditions. Summary of the Invention

[0005] In view of this, the present invention proposes an online identification and control method and quality control system that can realize a leap in quality management from passive detection to proactive prevention and adaptive control, and has good defect detection and compensation capabilities, so as to solve the related technical problems mentioned above.

[0006] The technical solution of this invention is implemented as follows: On the one hand, the present invention provides an online identification and control method for welding defects, comprising the following steps: S1. Construct a pre-welding process baseline that includes workpiece status information; S2. During the welding process, multi-source process data are collected synchronously, and the multi-source process data is repaired based on the pre-welding process benchmark when it is abnormal, generating a continuous time-series dataset. S3. Based on the embedded intelligent model, analyze the time series dataset to identify welding defects and determine the root causes of the deviations that lead to the defects; S4. Generate adaptive process adjustment strategies based on the root causes of deviations; S5. Based on a multi-objective optimization algorithm, optimize the process adjustment strategy and output the final process parameters to drive welding execution, while updating the embedded intelligent model.

[0007] Based on the above technical solutions, preferably, step S1 includes: scanning the weld area with the multi-source sensing unit at the end of the welding robot, reconstructing the three-dimensional geometric shape of the weld and identifying surface foreign objects based on the fusion data of three-dimensional point cloud and multispectral image, and constructing a personalized pre-welding digital twin benchmark containing geometric, surface state, initial physical field and material weldability information by combining the material property database and historical welding data. Simultaneously, based on the process and physical field information contained in the digital twin benchmark, a welding defect causal knowledge graph is initialized in parallel. The nodes of the welding defect causal knowledge graph include process parameters, physical field variables, and potential defect types. The edge weights of the welding defect causal knowledge graph are set based on domain expert experience or historical statistical relationships to construct a priori causal relationship framework related to welding quality.

[0008] Based on the above technical solutions, preferably, step S2 includes: during the welding process, the sensor network of the welding area is regarded as a dynamic graph node, and electrical parameters, visual data of the molten pool, temperature field and force data are collected synchronously to form a spatiotemporal graph sequence; When missing or abnormal data is detected, the spatiotemporal graph sequence is used as input. The pre-welding digital twin benchmark established in step S1 is used to run a physical information spatiotemporal graph neural network. The physical information spatiotemporal graph neural network embeds the welding heat conduction equation and the simplified model of the molten pool fluid as physical constraints into the loss function. It performs physically reliable interpolation and high-resolution reconstruction of the three-dimensional temperature field and stress field during the missing or abnormal period to generate a continuous time series dataset.

[0009] Based on the above technical solutions, preferably, step S3 includes: a lightweight causal inference network running in the real-time environment of the controller, receiving continuous time-series datasets, outputting defect type and location, and further outputting a contribution vector. The contribution vector quantifies the causal influence of multiple potential deviation sources on the currently identified defect. The deviation sources include at least different deviation types caused by foreign matter on the surface before welding, workpiece clamping deformation, sudden changes in local heat dissipation conditions, and mismatch of dynamic process parameters.

[0010] Based on the above technical solutions, the preferred lightweight causal inference network adopts a dual machine learning framework based on attention mechanism. In the first stage, the network learns the representation from confounding factors to each potential source of bias. In the second stage, the network evaluates the intervention effect of each source of bias on the defect to calculate the contribution vector.

[0011] Based on the above technical solutions, preferably, step S4 includes: according to the root cause of the deviation and the contribution vector, for the process parameters that need to be adjusted, designing their three-dimensional non-uniform, anisotropic spatial distribution functions in the weld length direction, width direction and plate thickness direction. This step includes a basic control layer and a meta-optimization layer; The basic control layer is used to generate the three-dimensional spatial distribution function of the process parameters at the current moment; The meta-optimization layer utilizes a lightweight temporal prediction network to predict the evolution trend of the physical field in the next few welding steps based on current and historical physical field data, and dynamically adjusts the key parameters of the three-dimensional spatial distribution function of the process parameters generated by the basic control layer accordingly.

[0012] Based on the above technical solutions, preferably, the welding current is designed as a distribution function that decays exponentially with the heat accumulation effect in the weld length direction, as a Gaussian distribution function with the weld centerline as the axis of symmetry in the weld cross-section direction, and as a distribution function that decreases linearly with the depth in the plate thickness direction.

[0013] Based on the above technical solutions, the preferred lightweight time-series prediction network used in the meta-optimization layer is a gated cyclic unit network. The input of the gated cyclic unit network is the feature vector of the molten pool image, the temperature field gradient vector, and the welding current value of the past N time steps, and the output is the adjustment amount of the key control points of the process parameter field function for the next M time steps.

[0014] Based on the above technical solutions, preferably, step S5 includes: making optimization decisions on the process adjustment strategy based on a multi-objective optimization algorithm, and outputting the final process parameters to drive welding execution; The multi-objective optimization algorithm adopts a non-dominated sorting genetic algorithm with an elitist strategy. The optimization objectives include at least welding defect rate, welding efficiency, energy consumption, process robustness index, and model prediction uncertainty. Meanwhile, the data from the entire welding process was used to perform federated causal learning-based incremental updates on the embedded intelligent model. Specifically, this included: uploading the statistical summary of the model gradient and contribution vector to the edge server after homomorphic encryption; the edge server aggregating encrypted information from multiple welding devices to iteratively update and verify the shared causal knowledge graph of welding defects; and distributing the updated causal knowledge graph to each local controller to constrain and guide the incremental learning of the local diagnostic model.

[0015] On the other hand, the present invention provides a welding quality control system for performing the above-mentioned online identification and control method for welding defects, comprising: The personalized benchmark and knowledge graph construction module is used to drive sensors and fuse data to build a personalized pre-weld digital twin benchmark and initialize a causal knowledge graph of welding defects. The physical information spatiotemporal graph neural network reconstruction module is used to realize the synchronous acquisition of multi-source data and run the physical information spatiotemporal graph neural network to reconstruct the three-dimensional field under physical constraints of abnormal data. An embedded causal diagnostic module is used to load and execute an embedded intelligent model to achieve welding defect identification and root cause contribution quantification. The 3D process field generation and two-layer optimization module is used to generate a 3D spatial distribution function of process parameters based on the contribution quantification results, and to perform basic control and meta-optimization. The Federated Causal Learning and Evolution module is used to perform multi-objective optimization decisions and manage federated learning processes based on causal knowledge graph co-evolution.

[0016] The online identification and control method and quality control system for welding defects of the present invention have the following advantages over the prior art: (1) Through five core steps—constructing a baseline, repairing data, determining the root cause, generating a strategy, optimizing decisions, and updating the model—a complete technical closed loop of perception, diagnosis, decision-making, execution, and evolution is formed during the welding process. This systematic solution fundamentally solves the core problem of existing technologies where various links are isolated from each other and cannot be dynamically adjusted and self-optimized according to real-time working conditions. This solution can realize online identification of welding defects, determination of root causes, and adaptive control closed loop, promoting the leap in welding quality management from passive detection to proactive prevention and adaptive control, and has good defect detection and compensation capabilities.

[0017] (2) By introducing personalized digital twin benchmarks and causal knowledge graphs, the system is provided with accurate initial state and prior knowledge framework; by adopting physical information spatiotemporal graph neural networks, reliable fusion and physical credibility reconstruction of multi-source heterogeneous data are ensured; by outputting contribution vectors through lightweight causal inference networks, defect diagnosis is advanced from simple type identification to quantitative analysis of the causal contribution of multiple root causes; the technical problems of insufficient benchmark adaptability, unreliable data quality, and fuzzy root cause analysis of deviations in existing technologies are effectively solved, thus effectively improving the comprehensiveness of system perception, making diagnosis more accurate, and making decision-making basis more solid.

[0018] (3) By adopting a three-dimensional process parameter field function and a two-layer optimization architecture, the process adjustment can be based on the quantitative diagnosis results to perform non-uniform and anisotropic fine matching in the length, width and thickness of the weld, and has short-term forward optimization capability, which solves the problem of rough process adjustment; at the same time, the federated causal learning mechanism aims to update the shared causal knowledge graph, which ensures that the system can continuously accumulate and verify the causal knowledge of welding process under the premise of protecting data privacy, realize system-level intelligent evolution across equipment and batches, and effectively solve the problems of model solidification and difficulty in adapting to new working conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the steps of the online identification and control method for welding defects and the quality control system of the present invention. Figure 2 This is a flowchart illustrating the online identification and control method for welding defects and the quality control system of the present invention. Figure 3This is a block diagram of the online identification and control method for welding defects and the quality control system of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] In the description of the embodiments of the present invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in the embodiments of the present invention based on the specific circumstances.

[0023] In the description of the embodiments of the present invention, it should be noted that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the embodiments of the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of the present invention.

[0024] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0025] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0026] The following disclosure provides numerous different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples. Such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. Additionally, examples of various specific processes and materials are provided in this invention; however, those skilled in the art will recognize the applicability of other processes and / or the use of other materials.

[0027] like Figures 1-3 As shown, the online identification and control method for welding defects of the present invention includes the following steps: S1, constructing a pre-welding process benchmark containing workpiece state information; S2, during the welding process, synchronously collecting multi-source process data, and repairing the multi-source process data when it is abnormal based on the pre-welding process benchmark, generating a continuous time-series dataset; S3, analyzing the time-series dataset based on an embedded intelligent model, identifying welding defects and determining the root cause type of the defects; S4, generating an adaptive process adjustment strategy based on the root cause of the deviation; S5, optimizing the process adjustment strategy based on a multi-objective optimization algorithm, and outputting the final process parameters to drive welding execution, while updating the embedded intelligent model; Specifically, in step S1, the multi-source sensing unit at the end of the welding robot scans the weld area. Based on the fusion data of three-dimensional point cloud and multispectral image, the three-dimensional geometry of the weld is reconstructed and surface foreign objects are identified. Combining the material property database and historical welding data, a personalized pre-welding digital twin benchmark containing geometric, surface state, initial physical field and material weldability information is constructed. At the same time, based on the process and physical field information contained in the digital twin benchmark, a welding defect causal knowledge graph is initialized in parallel. Its nodes include process parameters, physical field variables and potential defect types. The edge weights of the welding defect causal knowledge graph are set based on domain expert experience or historical statistical relationships to construct a priori causal relationship framework related to welding quality. As described above, a multi-source sensing unit mounted on the end effector of a welding robot scans the workpiece to be welded. This unit includes at least one 3D vision sensor (such as a structured light or laser scanner) and one multispectral camera. The 3D vision sensor acquires precise 3D point cloud data of the weld and its surrounding area. Each data point contains 3D spatial coordinates (X, Y, Z). The 3D point cloud is processed using filtering, denoising, and surface reconstruction algorithms to generate a precise 3D CAD model of the weld bevel. Key geometric dimensions such as bevel angle, root gap, and misalignment can be quantified and extracted for reconstructing the 3D geometric morphology of the weld bevel, gap, etc.

[0028] Multispectral cameras analyze the spectral reflectance characteristics of different bands to identify the distribution of foreign matter on the workpiece surface, such as oil stains, rust, or coatings. Based on the differences in reflectance characteristics of multispectral images in different bands, a preset classification algorithm (such as Support Vector Machine, SVM) is used to classify image pixels, identify and locate the areas of surface foreign matter such as oil stains, rust, and coating residues, and their approximate composition categories. The identification results are then mapped onto the aforementioned 3D model surface to form a 3D entity with "foreign matter annotations." The extracted geometric and surface condition information is integrated with the background database to construct a dynamic and personalized benchmark model: a material property database is accessed to obtain key physical property parameters such as density, specific heat capacity, thermal conductivity, and phase transition temperature based on the workpiece grade. Simultaneously, historical welding data for the workpiece batch or similar materials are queried to obtain empirical parameters such as typical welding thermal cycle curves and crack sensitivity indices.

[0029] Using the above information as input, a finite element analysis (FEA) model or a simplified heat conduction calculation model is initialized in an embedded simulation environment. This model uses the actual scanned geometry as its entity, assigns realistic material properties, and calculates the initial stress / strain field and temperature field (usually a uniform ambient temperature) distribution before welding based on preset initial welding parameters, serving as initial physical field information. Finally, a complete data package integrating "actual geometry," "surface condition," "material properties," "initial physical field," and "historical weldability data" constitutes a personalized pre-welding digital twin baseline for this specific workpiece. This baseline provides accurate initial conditions and comparison references for subsequent online monitoring and control.

[0030] While constructing the geometric baseline, a causal knowledge graph of welding defects is initialized in parallel within the system's knowledge base. This graph contains three core node types: process parameter nodes (e.g., welding current, voltage, speed, wire feed speed, shielding gas flow rate); physical field variable nodes (e.g., molten pool temperature, cooling rate, thermal stress, molten pool flow velocity); and potential defect type nodes (e.g., porosity, cracks, lack of fusion, undercut).

[0031] Directed edges between nodes represent potential causal relationships (e.g., "low welding current" -> "insufficient weld pool temperature" -> "lack of fusion"). The weight of each edge is set based on one or more of the following methods: Domain expert experience: assigned by process experts based on theory and experience. Historical statistical relationships: correlation analysis and causal discovery calculations are performed on accumulated historical welding big data, using statistical measures (such as conditional probability, causal effect strength) as initial weights. This knowledge graph is not fixed after initialization and will be continuously updated and optimized in subsequent online learning and federated evolution.

[0032] Thus, through high-precision 3D scanning and multispectral analysis, a unique digital twin benchmark is created for each specific workpiece. This provides a reference system that perfectly matches the actual situation of the workpiece for all subsequent online monitoring, defect diagnosis, and process adjustments, greatly improving the adaptability and control accuracy of the control system to incoming material fluctuations and complex working conditions, thereby improving quality consistency from the source. The initialized causal knowledge graph integrates fragmented process knowledge, physical principles, and defect phenomena into a structured, machine-understandable, and reasonable semantic network. It not only provides guidance containing physical laws for the training of embedded intelligent models, enabling them to converge faster and be more consistent with mechanisms, but more importantly, it provides an explanatory path and verification basis for the "root causes of deviations" diagnosed by the model online that conform to the thinking logic of human experts, significantly enhancing the credibility and acceptability in engineering applications.

[0033] Specifically, in step S2, during the welding process, the sensor network of the welding area is regarded as a dynamic graph node, and electrical parameters, molten pool vision, temperature field and force data are collected synchronously to form a spatiotemporal graph sequence. When data missing or abnormality is detected, the contribution vector uses the pre-welding digital twin benchmark established in step S1 and runs a physical information spatiotemporal graph neural network. The physical information spatiotemporal graph neural network embeds the welding heat conduction equation and the simplified model of molten pool fluid as physical constraints into the loss function, and performs physically reliable interpolation and high-resolution reconstruction of the three-dimensional temperature field and stress field during the missing or abnormal period to generate a continuous time series dataset. The system abstracts various sensors (current / voltage Hall sensors, high-speed vision cameras, infrared thermal imagers, and six-dimensional force sensors) deployed near the welding torch and on the workpiece into a dynamic sensor network. Each sensor is defined as a graph node, and node types include: current / voltage Hall sensor nodes, high-speed vision camera nodes, infrared thermal imager nodes, and six-dimensional force / torque sensor nodes. Each node is assigned a feature vector at time t. For example, the features of a current sensor are its instantaneous current and voltage values; the features of a vision camera node are morphological feature vectors extracted from the molten pool image. Specifically, the node features include real-time acquired electrical parameters, visual features of the molten pool, temperature field matrix, and force / torque vector. The edges between nodes are defined by their physical location associations and signal propagation relationships, thus forming a spatiotemporal graph sequence that evolves with time steps.

[0034] When the system detects an interruption in the data stream at a node or a significant deviation from physical laws (such as a sudden temperature jump), it triggers a repair process. The core of this process is running a physical information spatiotemporal graph neural network. The network takes a spatiotemporal graph sequence as input, and its training is optimized using a combined loss function. , in, For the total loss, For data reconstruction losses, As a weighting factor, Loss due to physical constraints; Of which, total loss It is the final objective function that is to be minimized when training the network model.

[0035] Data reconstruction loss The mean squared error (MSE) is used to measure the difference between the physical field data (such as temperature and stress) predicted by the neural network and the data collected by actual sensors. Its purpose is to force the network predictions to approximate the actual observations. The tradeoff coefficient is used to adjust... and The relative weights of the two factors in the total loss are used to balance the prediction's fidelity to the data and its adherence to physical laws. , in, The total number of valid data points considered when calculating this loss. During the training phase, this is typically the sum of the data volume of all normal sensor nodes across all time steps within a single batch; during the online repair phase, it is the total number of node data points in the current spatiotemporal graph that are not marked as anomalous or missing. For the neural network to the first A vector of predicted values ​​for each data point, which contains the values ​​of each sensor node (such as current, temperature of a pixel, stress at a point, etc.) inferred by the network. To and The corresponding number The true measurement vector of each data point is derived from the following stages: the model training stage uses offline, well-labeled historical welding datasets, while the online repair stage uses real-time sensor readings that are not marked as "missing" or "abnormal" by the system during the current welding process.

[0036] Physical constraint loss This is the core of our method. It embeds the prior physical knowledge of the welding process (specifically, the mathematical equations of the welding heat conduction equation and the simplified model of the molten pool fluid dynamics) as constraints into the network. This penalty penalizes the degree to which the network's prediction results violate the aforementioned physical laws, thereby ensuring that even when data is missing, the network's interpolation and reconstruction results conform to basic physical principles, achieving "physical reliability."

[0037]

[0038] in, The density of the material of the workpiece to be welded; The specific heat capacity of the material of the workpiece to be welded; The thermal conductivity of the material of the workpiece to be welded; The volumetric heat flux density input into the workpiece from a heat source such as a welding arc or laser. The kinematic viscosity of the molten metal during welding; It is the acceleration due to gravity; The coefficient of volumetric thermal expansion of the welding fluid; For reference temperature; To predict the temperature field; To predict the velocity field; The sign of the partial derivative; This is the Nabla operator.

[0039] This constraint forces the network's predicted output (temperature field) Flow velocity field The method must adhere to the physical laws described by the embedded welding heat conduction equations and the simplified model of molten pool fluid dynamics, fundamentally ensuring the physical authenticity and reliability of the reconstructed temperature field, stress field, and other key data. This overcomes the drawbacks of purely data-driven methods that may produce "physical distortion" predictions. Furthermore, by modeling the discretely distributed heterogeneous sensors (electrical, visual, thermal, and mechanical) as a spatiotemporal graph, and leveraging the inherent strong spatiotemporal correlation of graph neural networks, deep information fusion across modalities and spatiotemporal scales is effectively achieved, generating a more comprehensive digital representation of the welding state. Compared to traditional solutions that rely on a single or few sensors, this solution can automatically repair itself through physical laws and contextual information even if some sensors malfunction or are interfered with, generating reliable continuous data. This ensures the continuous and stable operation of the intelligent diagnosis and control closed loop, significantly enhancing the system's robustness.

[0040] In this scheme, step S3 includes: a lightweight causal inference network running in the real-time environment of the controller, receiving continuous time-series datasets, outputting defect type and location, and further outputting a contribution vector. The contribution vector quantifies the causal influence of multiple potential deviation sources on the currently identified defect. The deviation sources include at least different deviation types caused by foreign matter on the surface before welding, workpiece clamping deformation, sudden changes in local heat dissipation conditions, and mismatch of dynamic process parameters. The lightweight causal inference network adopts a dual machine learning framework based on the attention mechanism. In the first stage, the network learns the representation from the confounding factors to each potential source of bias. In the second stage, the network evaluates the intervention effect of each source of bias on the defect to calculate the contribution vector.

[0041] Specifically, a pruned and quantized lightweight causal inference network model is deployed in a welding real-time controller (such as an industrial PC or embedded module that supports AI inference). The network takes the continuous time-series dataset generated in the aforementioned steps as input and completes forward inference within millisecond latency.

[0042] The core of the network is to use a dual machine learning framework to calculate the conditionally averaged treatment effect (CATE) of each potential root cause of bias on the defect, and use this as its contribution. .

[0043] Phase 1 (Bias Removal): Fitting the results using an auxiliary network. (Defect indicators) and processing variables (No. The intensity characterization of each potential source of bias for confounding factors The residuals are obtained by considering the dependencies of (other processes and state variables): , , in, For defect indicators; For the first Characterization of the intensity of each potential source of deviation; This is a confounding factor vector, containing other possible influences from the four possible calls. and Process and state variables; and Used for fitting respectively and Auxiliary functions for dependencies; and To remove the biased residual; , in, The total number of potential source types of deviations to be considered; For the first The average conditional treatment effect of each root cause of deviation on the defect, i.e., its causal contribution. , This is a contribution vector; This is the random error term; Thus, this solution outputs a contribution vector through a causal inference framework, advancing defect analysis from traditional "presence" or "type judgment" to precise quantification of the impact of multiple deviation root causes, making the diagnostic results more instructive. It provides a direct and reliable basis for subsequent precise process adjustments: the quantified contribution clarifies the primary and secondary relationships of different deviation root causes. This enables the process adjustment strategy (S4) to perform differentiated and precise parameter compensation based on the specific contribution weight of each deviation root cause, avoiding the blindness of adjustments based on fuzzy qualitative judgments. Real-time causal analysis is achieved in resource-constrained environments: through model lightweighting technology, the computationally complex causal inference algorithm is deployed on an embedded real-time controller, achieving online, low-latency causal contribution analysis without cloud reliance, meeting the stringent requirements of real-time control in industrial settings.

[0044] In this scheme, step S4 includes: based on the root cause of the deviation and the contribution vector, designing a three-dimensional non-uniform, anisotropic spatial distribution function of the process parameters to be adjusted in the weld length direction, width direction, and plate thickness direction; this step includes a basic control layer and a meta-optimization layer; the basic control layer is used to generate the three-dimensional spatial distribution function of the process parameters at the current moment; the meta-optimization layer uses a lightweight temporal prediction network to predict the evolution trend of the physical field in the next few welding steps based on the current and historical physical field data, and dynamically adjusts the key parameters of the three-dimensional spatial distribution function of the process parameters generated by the basic control layer accordingly; Specifically, for the welding current, it is designed as a distribution function that decays exponentially with the heat accumulation effect in the weld length direction, a Gaussian distribution function with the weld centerline as the axis of symmetry in the weld cross-section direction, and a distribution function that decreases linearly with the depth in the plate thickness direction. Specifically, the lightweight time-series prediction network used in the meta-optimization layer is a gated recurrent unit network. The input of the gated recurrent unit network is the feature vector of the molten pool image, the temperature field gradient vector, and the welding current value of the past N time steps, and the output is the adjustment amount of the key control points of the process parameter field function for the next M time steps.

[0045] In practical applications, based on the contribution vector output from previous steps, the system designs a three-dimensional spatial distribution function for the process parameters that need adjustment. This is exemplified by the welding current. For example, this function is synthesized after being designed independently in three dimensions of the weld: In the length direction (x-axis), to counteract the effect of welding heat accumulation, the current is designed as an exponential decay function: ; in, The starting position of the weld ( The preset current value (=0); This is the attenuation coefficient, which is related to the material's thermal properties and welding speed, and is a positive value. The coordinates are along the length of the weld.

[0046] In the width direction (y-axis), to ensure symmetrical and concentrated heat input, a Gaussian function symmetrical about the weld centerline (y=0) is designed: ; in, For the weld centerline ( The current setting value at (=0); The standard deviation of the Gaussian distribution controls the distribution width of the heat input across the cross-section; The horizontal coordinate is perpendicular to the weld centerline; In the plate thickness direction (z-axis), to meet the requirements of penetration depth, it is designed to decrease linearly with depth: ; in, For the upper surface of the workpiece ( The current setting value is 0. It is a linear decreasing coefficient, related to plate thickness and weld depth requirements, and is a positive value; These are the depth coordinates along the thickness direction of the workpiece plate.

[0047] The final spatial distribution is a weighted synthesis of the above functions: ; in, In spatial location The final set current value at the location; The base welding current value; , , These are weighting coefficients, corresponding to the activation intensity adjusted in the length, width, and plate thickness directions, respectively, derived from the contribution vector; , , These are modulation functions for the length, width, and thickness directions, respectively.

[0048] After generating the above three-dimensional field function, the system enters a two-level optimization phase: Basic control layer: Performs the above calculations to generate the process parameter field for current control. ; Meta-optimization layer: A lightweight temporal prediction network (such as GRU) runs in parallel. This network predicts the evolution trend of the physical field (such as weld pool morphology and temperature gradient) for the next few welding steps based on current and historical molten pool images, temperature fields, and other physical field data. If the prediction indicates that an unfavorable state (such as overheating) is about to occur, the meta-optimization layer will dynamically adjust the key parameters in the basic control layer function (for example, increasing the attenuation coefficient α to reduce the heat input in advance), thus achieving feedforward optimization. The meta-optimization layer uses a gated recurrent unit network (GRU) as its temporal prediction core. Through its update and reset gate mechanisms, the GRU effectively captures the long-term and short-term dependencies of physical field changes during the welding process, and has a small number of model parameters, meeting real-time control requirements. This includes network selection and input feature construction, network output and parameter adjustment. The network input is a time-series sequence that incorporates multi-dimensional state information, including: Molten pool image feature vector: Abstract features extracted by running a lightweight convolutional neural network on the current molten pool region image (from a high-speed camera), representing the molten pool morphology and flow state; Temperature field gradient vector: Calculate the magnitude and direction of the temperature gradient at key locations (such as the molten pool front and the heat-affected zone) from the reconstructed 3D temperature field data; Welding current value: The process parameter at the current moment; The features from the past N consecutive time steps are concatenated to form the input sequence of the GRU network. ,in, Let be the eigenvector at time t; The GRU network uses the aforementioned N-step history as a condition to predict the evolution trend of the welding state over the next M time steps. Its output is not a direct state value, but rather the adjustment amount of key control points in the three-dimensional spatial distribution function of process parameters within the next M steps. ;in, For adjustment purposes; P represents the set of key control point parameters in the three-dimensional spatial distribution function of process parameters. Taking the three-dimensional function of welding current as an example, P includes... , , and possible function magnitude coefficients , , wait; This represents M consecutive future time steps, starting from the next time step. M is the prediction step size, which is also a preset hyperparameter. Therefore... This means that the network predicts that, in Constantly responding to key parameters The amount of adjustment made.

[0049] These parameters collectively define a closed-loop predictive control system: the GRU network learns a dynamic model of the welding process by analyzing a multidimensional state sequence that integrates morphology (image features), thermodynamics (temperature gradient), and energy (current) over the most recent N steps, and predicts the three-dimensional distribution function of the process parameters required to maintain ideal welding conditions (such as a stable weld pool and a reasonable temperature gradient) over the next M steps. What kind of subtle adjustments should be made in the sequence? This represents a leap from perceiving the present to predicting and optimizing the future, and is the core mathematical representation of the meta-optimization layer in achieving feedforward control.

[0050] Based on this, this scheme achieves precise and differentiated distribution of welding energy in three-dimensional space by designing independent and anisotropic spatial distribution functions for parameters such as welding current in three dimensions: length, width, and plate thickness. This enables control of thermal effect distribution and effectively compensates for local quality deviations caused by geometric and uneven heat dissipation, making it particularly suitable for welding complex structures such as thick plates and irregularly shaped welds. By introducing a two-layer optimization architecture, especially the meta-optimization layer's prediction and early intervention of future physical fields, the control system is upgraded from a traditional "detection-response" mode to a "prediction-prevention" mode. This significantly enhances the ability to suppress disturbances in the welding dynamic process and improves the overall robustness and stability of the process. The specific form of the three-dimensional distribution function (exponential, Gaussian, linear) and its composite weights are directly driven by the quantized contribution vector output from step S3. This establishes a clear and interpretable mathematical mapping relationship between the process adjustment strategy and the root cause of deviations, realizing a closed loop from problem diagnosis to the execution of adjustment amounts, greatly improving the accuracy and reliability of control.

[0051] In this scheme, step S5 includes: making optimization decisions on process adjustment strategies based on a multi-objective optimization algorithm, and outputting the final process parameters to drive welding execution; the multi-objective optimization algorithm adopts a non-dominated sorting genetic algorithm with an elitist strategy, and the optimization objectives include at least welding defect rate, welding efficiency, energy consumption, process robustness indicators, and model prediction uncertainty; at the same time, the data of the entire welding process is used to perform federated causal learning-based incremental updates on the embedded intelligent model, specifically including: uploading the statistical summary of the model gradient and contribution vector to the edge server after homomorphic encryption; the edge server aggregates encrypted information from multiple welding devices to iteratively update and verify the shared welding defect causal knowledge graph; and the updated causal knowledge graph is distributed to each local controller to constrain and guide the incremental learning of the local diagnostic model. Specifically, the system receives the process adjustment strategy (i.e., the three-dimensional process parameter field function) from S4 and uses a non-dominated sorting genetic algorithm with an elitist strategy for optimization. This algorithm simultaneously optimizes at least five objectives: Welding defect rate ;in, This refers to the cumulative length of welds with defects identified by the online diagnostic module. This represents the total length of the completed welds. Process robustness indicators ,in, The nominal value of the set process parameters (such as welding current); This refers to the allowable range of process parameter fluctuations; An event indicating that the welding quality is acceptable; Represents probability; Welding efficiency ,in, To complete The total time spent on welding the length; Energy consumption ,in, In order to be in The instantaneous value of the welding voltage collected at all times; exist The instantaneous value of welding current collected at all times; This represents the time interval for data collection. Model prediction uncertainty ,in, Physical information spatiotemporal graph neural network for the first Predicted output for each data point (e.g., temperature, stress); For variance calculation; The number of samples used for evaluation; The algorithm outputs a Pareto optimal solution set, which is then filtered through multi-attribute decision-making to select the optimal final process parameters, and immediately drives welding execution. After welding, the entire process data is used to update the system. Each local controller homomorphically encrypts the statistical summary of the model gradient of the lightweight causal inference network and the contribution vector θ output by S3 in this welding process, and uploads it to the edge server. The edge server aggregates encrypted information from multiple welding devices, decrypts it, and uses this data (especially the statistical regularity of the contribution vector) to iteratively update and verify the shared welding defect causal knowledge graph, corrects the weights of the causal relationships between nodes (process parameters, physical fields, defects) in the graph, and securely distributes the updated causal knowledge graph to each local controller. When the local system performs incremental training on the diagnostic model in the future, the causal relationship rules in this graph are used as regularization constraints or sample weighting criteria, and incorporated into the loss function, thereby "guiding" the model to learn features that conform to global evolutionary knowledge, achieving continuous and co-evolution of diagnostic capabilities.

[0052] As mentioned above, by employing a multi-objective optimization algorithm, simultaneously seeking optimization across five dimensions—quality, efficiency, energy consumption, robustness, and reliability—the system outputs an optimal compromise solution that comprehensively considers all manufacturing elements, rather than focusing solely on local optimization of a single indicator. This achieves the best balance between economy, reliability, and high quality in the welding process. Through a federated causal learning mechanism, while protecting the privacy of on-site data, the system achieves the accumulation, verification, and co-evolution of process causal knowledge (rather than simply model parameters) across equipment and batches. This not only improves model performance but also accumulates and solidifies interpretable domain expert experience, significantly enhancing the adaptability and long-term intelligence level of the entire manufacturing system.

[0053] The welding quality control system of the present invention is used to execute the above-mentioned online identification and control method for welding defects, comprising: a personalized benchmark and knowledge graph construction module, used to drive sensors and fuse data, construct a personalized pre-welding digital twin benchmark, and initialize a causal knowledge graph of welding defects; a physical information spatiotemporal graph neural network reconstruction module, used to realize synchronous acquisition of multi-source data and run a physical information spatiotemporal graph neural network to reconstruct a three-dimensional field under physical constraints for abnormal data; an embedded causal diagnosis module, used to load and execute an embedded intelligent model to realize the identification of welding defects and the quantification of the contribution of deviation roots; a three-dimensional process field generation and two-layer optimization module, used to generate a three-dimensional spatial distribution function of process parameters based on the contribution quantification results, and execute basic control and meta-optimization; and a federated causal learning and evolution module, used to execute multi-objective optimization decisions and manage the federated learning process based on the causal knowledge graph co-evolution.

[0054] The system operates according to the logic of pre-welding preparation - online sensing and repair - intelligent diagnosis - precise control - decision execution and evolution; Personalized benchmark and knowledge graph construction module: Process start point: Before welding begins, this module drives the multi-source sensing unit (such as 3D vision, multispectral camera) integrated at the end of the robot to scan the workpiece.

[0055] Core work: Integrating 3D point cloud and multispectral image data to reconstruct the true 3D geometry of the weld and identify surface foreign objects. Combining material databases and historical data, a unique, personalized pre-weld digital twin benchmark is established for each workpiece. Simultaneously, a causal knowledge graph of welding defects is initialized, storing the potential causal relationships between processes, physical fields, and defects in a graphical structure, providing an interpretable reasoning framework for subsequent intelligent diagnosis.

[0056] Online multi-source sensing and reliable data generation: Process Implementation: After the welding process starts, this module is responsible for the real-time data link; Core functionality: Simultaneously acquire heterogeneous data from multiple sources, including electrical parameters, molten pool visual data, temperature field data, and force perception data, and organize them into a spatiotemporal graph sequence. Upon detecting missing or abnormal data, the embedded physical information spatiotemporal graph neural network is immediately invoked. This network uses the digital twin benchmark established in the previous module as initial conditions and forces its output to comply with welding physics laws such as heat conduction and fluid dynamics. This allows for physically reliable repair and reconstruction of missing key information such as the three-dimensional temperature field and stress field, outputting a continuous and reliable high-dimensional time-series dataset.

[0057] Embedded real-time diagnostics and deviation root cause quantification: Process Analysis: This module receives clean, continuous time-series data provided by the previous module.

[0058] Core functionality: Running a lightweight causal inference network deployed within a real-time controller. This network not only quickly identifies the type and location of defects (such as porosity and cracks), but more importantly, performs causal analysis to determine the type of root cause of the deviation leading to the defect (such as surface foreign matter and assembly deformation), and outputs a contribution vector that quantifies the specific impact weight of each root cause on the current defect.

[0059] 3D spatial process optimization and forward-looking adjustments: Process Decision: Based on accurate quantitative diagnostic results, this module generates precise control instructions.

[0060] The core work involves the following steps: First, the basic control layer designs a non-uniform, anisotropic spatial distribution function across the weld length, width, and plate thickness dimensions for the parameters to be adjusted (such as current) based on the contribution vector. Then, the meta-optimization layer runs a lightweight prediction network (such as GRU) to predict future trends based on real-time physics data and proactively fine-tunes the parameters of the basic control layer, forming a two-layer optimization strategy of current optimization plus future pre-tuning.

[0061] Multi-objective decision-making and system co-evolution: Process execution and closed loop: This module completes the final decision and drives the system's self-improvement; Core work: First, a multi-objective optimization algorithm is used to globally optimize the control strategy, weighing multiple objectives such as quality, efficiency, energy consumption, and robustness, and outputting the comprehensive optimal process parameters to drive welding execution. Simultaneously, a federated causal learning process is managed: each local device uploads encrypted model update information to the edge server; the server aggregates global information, with the core task of iteratively updating the shared welding defect causal knowledge graph; the updated graph is then distributed to each device to constrain and improve the performance of its local diagnostic model, thereby achieving continuous, secure, and collaborative evolution of system-level knowledge.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for online identification and control of welding defects, characterized in that, Includes the following steps: S1. Construct a pre-welding process baseline that includes workpiece status information; S2. During the welding process, multi-source process data is collected synchronously, and when the multi-source process data is abnormal based on the pre-welding process benchmark, it is repaired to generate a continuous time-series dataset. S3. Based on the embedded intelligent model, analyze the time series dataset to identify welding defects and determine the root causes of the deviations that lead to the defects; S4. Generate an adaptive process adjustment strategy based on the root causes of the deviation; S5. Based on a multi-objective optimization algorithm, optimize the process adjustment strategy and output the final process parameters to drive welding execution, while updating the embedded intelligent model.

2. The online identification and control method for welding defects as described in claim 1, characterized in that, Step S1 includes: scanning the weld area with the multi-source sensing unit at the end of the welding robot, reconstructing the three-dimensional geometric shape of the weld and identifying surface foreign objects based on the fusion data of three-dimensional point cloud and multispectral image, and constructing a personalized pre-welding digital twin benchmark containing geometric, surface state, initial physical field and material weldability information by combining the material property database and historical welding data. Simultaneously, based on the process and physical field information contained in the digital twin benchmark, a welding defect causal knowledge graph is initialized in parallel. The nodes of the welding defect causal knowledge graph include process parameters, physical field variables, and potential defect types. The edge weights of the welding defect causal knowledge graph are set based on domain expert experience or historical statistical relationships to construct a priori causal relationship framework related to welding quality.

3. The online identification and control method for welding defects as described in claim 2, characterized in that, Step S2 includes: during the welding process, the sensor network of the welding area is regarded as a dynamic graph node, and electrical parameters, visual data of the molten pool, temperature field and force data are collected synchronously to form a spatiotemporal graph sequence; When data loss or anomaly is detected, the spatiotemporal graph sequence is used as input. The pre-welding digital twin benchmark established in step S1 is used to run a physical information spatiotemporal graph neural network. The physical information spatiotemporal graph neural network embeds the welding heat conduction equation and the simplified model of the molten pool fluid as physical constraints into the loss function. It performs physically reliable interpolation and high-resolution reconstruction of the three-dimensional temperature field and stress field during the missing or abnormal period to generate a continuous time series dataset.

4. The online identification and control method for welding defects as described in claim 1, characterized in that, Step S3 includes: a lightweight causal inference network running in the real-time environment of the controller, receiving the continuous time-series dataset, outputting the defect type and location, and further outputting a contribution vector. The contribution vector quantifies the causal influence of multiple potential deviation sources on the currently identified defect. The deviation sources include at least different deviation types caused by foreign matter on the surface before welding, workpiece clamping deformation, sudden changes in local heat dissipation conditions, and mismatch of dynamic process parameters.

5. The online identification and control method for welding defects as described in claim 4, characterized in that, The lightweight causal inference network employs a dual machine learning framework based on an attention mechanism. In the first stage, the network learns representations from confounding factors to each root cause of bias. In the second stage, the network evaluates the intervention effect of each root cause of bias on the defect to calculate the contribution vector.

6. The online identification and control method for welding defects as described in claim 4, characterized in that, Step S4 includes: Based on the aforementioned deviation root causes and contribution vectors, for the process parameters that need to be adjusted, a three-dimensional non-uniform, anisotropic spatial distribution function is designed in the weld length direction, width direction, and plate thickness direction; this step includes a basic control layer and a meta-optimization layer. The basic control layer is used to generate the three-dimensional spatial distribution function of the process parameters at the current moment; The meta-optimization layer utilizes a lightweight temporal prediction network to predict the evolution trend of the physical field in the next few welding steps based on current and historical physical field data, and dynamically adjusts the key parameters of the three-dimensional spatial distribution function of the process parameters generated by the basic control layer accordingly.

7. The online identification and control method for welding defects as described in claim 6, characterized in that: For the welding current, it is designed as a distribution function that decays exponentially with the heat accumulation effect in the weld length direction, a Gaussian distribution function with the weld centerline as the axis of symmetry in the weld cross-section direction, and a distribution function that decreases linearly with the depth in the plate thickness direction.

8. The online identification and control method for welding defects as described in claim 6, characterized in that: The lightweight time-series prediction network used in the meta-optimization layer is a gated recurrent unit network. The input of the gated recurrent unit network is the feature vector of the molten pool image, the temperature field gradient vector, and the welding current value of the past N time steps, and the output is the adjustment amount of the key control points of the process parameter field function for the next M time steps.

9. The online identification and control method for welding defects as described in claim 1, characterized in that, Step S5 includes: The process adjustment strategy is optimized based on a multi-objective optimization algorithm, and the final process parameters are output to drive welding execution. The multi-objective optimization algorithm adopts a non-dominated sorting genetic algorithm with an elite strategy, and the optimization objectives include at least welding defect rate, welding efficiency, energy consumption, process robustness index and model prediction uncertainty. Simultaneously, the data from the entire welding process is used to perform federated causal learning-based incremental updates on the embedded intelligent model. Specifically, this includes: uploading the statistical summary of the model gradient and contribution vector to the edge server after homomorphic encryption; the edge server aggregating encrypted information from multiple welding devices to iteratively update and verify the shared causal knowledge graph of welding defects; and distributing the updated causal knowledge graph to each local controller to constrain and guide the incremental learning of the local diagnostic model.

10. A welding quality control system, characterized in that, The online identification and control method for welding defects according to any one of claims 1 to 9 includes: The personalized benchmark and knowledge graph construction module is used to drive sensors and fuse data to build a personalized pre-weld digital twin benchmark and initialize a causal knowledge graph of welding defects. The physical information spatiotemporal graph neural network reconstruction module is used to realize the synchronous acquisition of multi-source data and run the physical information spatiotemporal graph neural network to reconstruct the three-dimensional field under physical constraints of abnormal data; An embedded causal diagnosis module is used to load and execute the embedded intelligent model to realize welding defect identification and root cause contribution quantification. The 3D process field generation and two-layer optimization module is used to generate a 3D spatial distribution function of process parameters based on the contribution quantification results, and to perform basic control and meta-optimization. The Federated Causal Learning and Evolution module is used to perform multi-objective optimization decisions and manage federated learning processes based on causal knowledge graph co-evolution.

Images