Method for monitoring all-position dynamic molten pool in laser welding of pressure vessel barrel

Abstract

The invention relates to the technical field of laser welding, and provides a pressure vessel barrel laser welding all-position dynamic molten pool monitoring method which comprises the following steps: arranging a laser heat source and monitoring equipment to obtain molten pool temperature data and molten pool form data of a to-be-welded pressure vessel barrel; constructing a thermodynamic constraint based on a heat conduction equation, and training a molten pool monitoring model in combination with temperature data; constructing a hydrodynamic constraint based on a hydrodynamic equation, and fusing morphological data to further optimize the model; a target molten pool monitoring model with physical mechanism consistency is obtained by constructing thermal-fluid coupling physical constraints and combining with strengthening model training; and finally, inputting space-time coordinates of a to-be-measured position to realize inversion of the dynamic state of the molten pool in the all-position welding process. According to the method, multi-physics field constraint and data driving modeling are fused, the generalization ability and robustness of the model under the complex welding posture are improved, and the technical problem that the molten pool state is difficult to accurately monitor in all-position laser welding is effectively solved.

Description

Technical Field

[0001] The embodiments of the present invention generally relate to the field of laser welding technology, and more particularly to a method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders. Background Technology

[0002] With the continuous advancement of high-end equipment manufacturing technologies in energy, chemical, and nuclear power fields, the safety and reliability requirements for pressure vessels under extreme environments such as high temperature, high pressure, and strong corrosion are constantly increasing. The welding quality of their core structural components directly affects the overall performance and service life of the equipment. Laser welding, with its advantages of high energy density, precise and controllable heat input, and excellent weld formation, has been gradually applied to the high-precision automated welding of pressure vessel cylinders. However, in actual manufacturing processes, pressure vessel cylinders often require all-position welding (such as flat welding, vertical welding, horizontal welding, and overhead welding). The continuous changes in welding posture lead to significant alterations in the dynamic behavior of the molten pool. Under conditions of constantly changing gravity, the flow, heat transfer, and formation of the molten pool are subject to complex nonlinear coupling effects of gravity, surface tension, and recoil pressure, which can easily cause formation defects such as molten pool instability, collapse, and undercut, severely restricting the geometric quality of the weld and the mechanical properties of the joint.

[0003] Currently, research in this field faces a core contradiction: quality control in all-position welding urgently requires high-precision real-time sensing and prediction of the dynamic coupling of multiple physics fields within the molten pool. However, existing technologies have significant limitations. On the one hand, while analytical or numerical models based on purely physical mechanisms can describe the thermo-fluid coupling process, their computational efficiency is insufficient to meet real-time monitoring needs, and they lack adaptability to complex dynamic conditions. On the other hand, models purely driven by data often become "black boxes" due to the lack of physical constraints, and their predictions lack physical rationality and robustness under conditions outside the training data. On the other hand, widely used methods such as infrared thermometry or visual sensing can usually only acquire single-dimensional information such as local temperature or two-dimensional contours of the molten pool, failing to simultaneously provide complete observation data on the three-dimensional flow field and complete temperature field within the molten pool. This results in a lack of sufficient high-dimensional "true values" to support model training and validation. These factors collectively lead to the inability of existing models to accurately characterize and predict complex dynamic behaviors in all-position welding, such as abrupt changes in molten pool morphology and thermal accumulation effects caused by continuous changes in posture, thus limiting their generalization ability and prediction accuracy in engineering practice. Summary of the Invention

[0004] To address the above issues, this invention integrates the heat conduction equation and the fluid dynamics equation to construct a molten pool thermal-fluid coupling constraint with clear physical meaning. This introduces the physical laws of the welding process into the data-driven molten pool monitoring model training process, thereby achieving joint modeling and constraint optimization of the molten pool temperature field and flow field evolution behavior.

[0005] According to an embodiment of the present invention, a method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders is provided.

[0006] In a first aspect of the invention, a method for monitoring the dynamic molten pool in all positions during laser welding of a pressure vessel shell is provided. The method includes:

[0007] Step S01: Deploy laser heat source and monitoring equipment to obtain molten pool temperature data and molten pool morphology data of the pressure vessel cylinder to be welded, and determine the set of measurement points and matching points for training based on the welding area, all-position welding angle and monitoring position of the monitoring equipment.

[0008] Step S02: Construct a molten pool monitoring model based on a feedforward deep neural network, obtain temperature prediction values ​​through the molten pool monitoring model, construct molten pool thermodynamic constraints based on the heat conduction equation and temperature prediction values, and train the molten pool monitoring model based on the thermodynamic constraints to obtain the first target molten pool monitoring model;

[0009] Step S03: Obtain the predicted value of the molten pool shape through the molten pool monitoring model, construct the molten pool fluid dynamic constraints based on the fluid dynamic equation and the predicted value of the molten pool shape, and train the first target molten pool monitoring model based on the fluid dynamic constraints to obtain the second target molten pool monitoring model;

[0010] Step S04: Construct thermal-fluid coupling constraints for the molten pool, and train the second target molten pool monitoring model based on the thermal-fluid coupling constraints to obtain the third target molten pool monitoring model;

[0011] Step S05: Obtain the spatiotemporal coordinates of the test position of the pressure vessel cylinder to be welded, and determine the dynamic state of the molten pool at the test position of the pressure vessel cylinder to be welded through the third target molten pool monitoring model.

[0012] Furthermore, the set of measuring points mentioned in step S01 includes several measuring point coordinates, each measuring point corresponding to the location where the actual molten pool temperature and morphology data can be obtained; the set of matching points includes several matching point coordinates, each matching point corresponding to the location of the physical equation residual calculation, used to embed physical law constraints.

[0013] Furthermore, after determining the set of measurement points and the set of collocation points for training in step S01, the method further includes a step of normalizing the coordinates of the measurement points, the coordinates of the collocation points, and the all-position welding angle: linearly mapping the coordinates of the measurement points and the coordinates of the collocation points to the interval [0, 1]; and converting the all-position welding angle θ into a unit vector (sinθ, cosθ).

[0014] Further, step S02, which involves constructing thermodynamic constraints for the molten pool based on the heat conduction equation and predicted temperature values, and training the molten pool monitoring model based on these constraints to obtain the first target molten pool monitoring model, includes:

[0015] The temperature, thermal conductivity, and thermal diffusivity at the locations of the measuring points and the coordinates of the matching points on the pressure vessel cylinder to be welded are predicted using the molten pool monitoring model, and the predicted temperature value is obtained.

[0016] Based on the set of measuring points and the set of matching points of the pressure vessel cylinder to be welded, as well as the predicted temperature value, a thermodynamic constraint of the molten pool is constructed, and the thermodynamic constraint loss of the molten pool monitoring model is determined.

[0017] With the goal of minimizing the thermodynamic constraint loss of the molten pool monitoring model, the molten pool monitoring model is trained and its parameters are updated.

[0018] 5. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 4, characterized in that the calculation formula for the thermodynamic constraint of the molten pool is as follows:

[0019] ,

[0020] In the formula, To measure the degree of agreement between the predictions of the thermodynamic model and the physical laws and measured data; To measure the difference between the temperature field predicted by the model and the measured temperature data; To measure the degree of deviation between the temperature field predicted by the model and the heat conduction equation; The temperature field predicted by the model, including spatiotemporal coordinates. , The function output; The temperature data measured in the experiment comes from the sensor and exists only at a limited number of measuring points; Let be the partial derivative with respect to time t, describing the rate of change of temperature over time.

[0021] Further, step S03, which involves constructing molten pool fluid dynamic constraints based on fluid dynamic equations and predicted molten pool morphology, and training the first target molten pool monitoring model based on these constraints to obtain the second target molten pool monitoring model, includes:

[0022] The velocity field, pressure field, and surface morphology of the molten pool are obtained at the locations of the measuring point coordinates and the collocation point coordinates of the pressure vessel cylinder to be welded by the molten pool monitoring model, and the predicted value of the molten pool morphology is obtained.

[0023] Based on the set of measuring points and the set of matching points of the pressure vessel cylinder to be welded, as well as the predicted value of the molten pool morphology, a molten pool fluid dynamic constraint is constructed, and the molten pool fluid dynamic constraint loss of the molten pool monitoring model is determined.

[0024] With the goal of minimizing the fluid dynamics constraint loss of the molten pool monitoring model, the first target molten pool monitoring model is trained and its parameters are updated.

[0025] Furthermore, the formula for the fluid dynamic constraint loss of the molten pool is as follows:

[0026] ,

[0027] In the formula, Used to measure the degree of agreement between the predictions of a fluid dynamics model and the measured fluid morphology and fluid dynamics equations; The difference between the fluid state predicted by the model and the measured fluid morphology data is measured, and includes three types of variables; among them For velocity field, that is, the flow velocity of liquid metal in molten pool; This refers to the pressure field, specifically the static pressure distribution within the molten pool. Surface morphology refers to the height variation of the free surface of the molten pool; This is to measure the degree of deviation between the fluid field predicted by the model and the fundamental equations of fluid mechanics.

[0028] Further, the specific steps of step S04 are as follows: determining the thermal-fluid coupling constraint loss of the molten pool based on the molten pool thermodynamic constraint loss and the molten pool fluid dynamic constraint loss of the second target molten pool monitoring model;

[0029] With the goal of minimizing the thermal-fluid coupling constraint loss of the molten pool monitoring model, the second target molten pool monitoring model is trained and its parameters are updated to obtain the third target molten pool monitoring model.

[0030] Furthermore, the formula for the thermal-fluid coupling constraint loss of the molten pool is as follows:

[0031] ,

[0032] In the formula, This is a weighted sum of heat conduction losses and fluid dynamics losses; Controlling thermal conduction constraints The proportion of contribution to the total loss; when > When the model prioritizes satisfying the fitting accuracy of the temperature field and the constraints of the heat conduction equation, it prioritizes satisfying the constraints of the temperature field fitting accuracy and the heat conduction equation. When = 0, the model degenerates into a pure fluid dynamics model, and thermal effects are ignored; Controlled fluid dynamics constraints The proportion of contribution to the total loss; when > When the model prioritizes satisfying the fitting accuracy of the velocity field, pressure field, molten pool morphology, and fluid dynamics equation constraints; when When = 0, the model degenerates into a pure heat conduction model, ignoring the influence of fluid flow on heat transport.

[0033] Furthermore, the determination of the dynamic state of the molten pool at the test location of the pressure vessel cylinder to be welded using the third target molten pool monitoring model in step S05 includes:

[0034] The spatiotemporal coordinates of the test position of the pressure vessel cylinder to be welded are input into the third target molten pool monitoring model to obtain the molten pool temperature and molten pool morphology of the pressure vessel cylinder to be welded at the test position.

[0035] Based on the molten pool temperature and molten pool morphology at the test location of the pressure vessel cylinder to be welded, the dynamic state of the molten pool at the test location of the pressure vessel cylinder to be welded is determined. The dynamic state of the molten pool includes parameters such as molten pool stability, molten depth, and molten width.

[0036] This invention integrates the heat conduction equation and the fluid dynamics equation to construct a molten pool heat-fluid coupling constraint with clear physical meaning. It introduces the physical laws of the welding process into the data-driven molten pool monitoring model training process, realizing the joint modeling and constraint optimization of the molten pool temperature field and flow field evolution behavior.

[0037] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0038] Beneficial effects:

[0039] 1. By integrating the heat conduction equation and the fluid dynamics equation, a molten pool thermal-fluid coupling constraint with clear physical meaning is constructed. The physical laws of the welding process are introduced into the data-driven molten pool monitoring model training process, realizing the joint modeling and constraint optimization of the molten pool temperature field and flow field evolution behavior.

[0040] 2. By deploying laser heat sources and multi-source monitoring equipment to obtain actual molten pool data, and combining the measurement point and matching point collaborative training strategy, the model's generalization ability and prediction accuracy for complex dynamic molten pool behavior (such as molten pool morphology changes, heat accumulation effect, gravity influence, etc.) under all-position welding conditions are improved.

[0041] 3. The target molten pool monitoring model can invert the dynamic state of the molten pool in real time based on the spatiotemporal coordinates of the location to be measured, realizing high-precision and robust online monitoring of key parameters such as weld depth, weld width, and molten pool stability during the welding process, providing reliable technical support for intelligent welding quality control of equipment with high safety requirements such as pressure vessels. Attached Figure Description

[0042] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein:

[0043] Figure 1 A flowchart illustrating a method for monitoring the dynamic molten pool in all positions during laser welding of a pressure vessel cylinder according to an embodiment of the present invention is shown. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0045] According to an embodiment of the present invention, a method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders is proposed. By integrating the heat conduction equation and the fluid dynamics equation, a molten pool heat-fluid coupling constraint with clear physical meaning is constructed. The physical laws of the welding process are introduced into the data-driven molten pool monitoring model training process, realizing the joint modeling and constraint optimization of the evolution behavior of the molten pool temperature field and flow field.

[0046] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0047] Figure 1 This is a schematic flowchart of a method for monitoring the dynamic molten pool in all positions during laser welding of a pressure vessel cylinder, according to an embodiment of the present invention. The method includes:

[0048] Step S01: Deploy laser heat source and monitoring equipment to obtain molten pool temperature data and molten pool morphology data of the pressure vessel cylinder to be welded, and determine the set of measurement points and matching points for training based on the welding area, all-position welding angle and monitoring position of the monitoring equipment.

[0049] Step S02: Construct a molten pool monitoring model based on a feedforward deep neural network, obtain temperature prediction values ​​through the molten pool monitoring model, construct molten pool thermodynamic constraints based on the heat conduction equation and temperature prediction values, and train the molten pool monitoring model based on the thermodynamic constraints to obtain the first target molten pool monitoring model;

[0050] Step S03: Obtain the predicted value of the molten pool shape through the molten pool monitoring model, construct the molten pool fluid dynamic constraints based on the fluid dynamic equation and the predicted value of the molten pool shape, and train the first target molten pool monitoring model based on the fluid dynamic constraints to obtain the second target molten pool monitoring model;

[0051] Step S04: Construct thermal-fluid coupling constraints for the molten pool, and train the second target molten pool monitoring model based on the thermal-fluid coupling constraints to obtain the third target molten pool monitoring model;

[0052] Step S05: Obtain the spatiotemporal coordinates of the test position of the pressure vessel cylinder to be welded, and determine the dynamic state of the molten pool at the test position of the pressure vessel cylinder to be welded through the third target molten pool monitoring model.

[0053] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0054] To provide a clearer explanation of the above-mentioned method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders, a specific embodiment will be used for illustration below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.

[0055] The following specific example will further illustrate the method for all-position dynamic molten pool monitoring in laser welding of pressure vessel cylinders:

[0056] Step S01: Deploy laser heat source and monitoring equipment to obtain molten pool temperature data and molten pool morphology data of the pressure vessel cylinder to be welded, and determine the set of measurement points and matching points for training based on the welding area, all-position welding angle and monitoring position of the monitoring equipment.

[0057] Among them, the laser heat source is a laser beam generating device used to provide high energy density heat input during the welding process of pressure vessel cylinder, so as to achieve high-precision, low-heat-input local melting and support the all-position laser welding process.

[0058] The monitoring equipment is a collection of multi-source sensors used to collect information on the physical state of the molten pool, providing measured data on the temperature and morphology fields of the molten pool as input for model training and validation. Specifically, the monitoring equipment can integrate infrared thermal imaging equipment, high-speed optical vision systems, and coaxial spectral monitoring devices to simultaneously capture temperature and morphology signals.

[0059] In this embodiment, the laser heat source generates a focused beam through a laser, which acts on the surface of the workpiece to form a molten pool. The monitoring equipment simultaneously captures temperature and morphology signals using an infrared thermal imager, a high-speed vision camera, and a spectral sensor.

[0060] Molten pool temperature data are time-series measurements that reflect the temperature distribution inside and on the surface of the molten pool. They are used to construct thermodynamic constraints and calibrate the model's ability to predict thermal field evolution.

[0061] Molten pool morphology data is dynamic observation data that describes the geometric profile of the molten pool (such as melt width, melt depth, and surface curvature). It is used to support fluid dynamics constraint modeling and reflects the influence of gravity, surface tension, and other factors on the stability of the molten pool.

[0062] Molten pool stability refers to the ability of the molten pool to maintain its geometric and thermodynamic equilibrium during welding. It is used to assess whether instability phenomena such as spattering, collapse, or oscillation occur, and is a key criterion for weld quality. In a specific embodiment, molten pool stability includes, but is not limited to, thermal-fluid coupling stability, resistance to gravitational disturbances, and surface tension self-healing ability. Penetration depth is the maximum melting depth of the molten pool along the thickness direction of the workpiece. It directly affects the joint's load-bearing capacity and sealing performance and is a core acceptance indicator for pressure vessel welding. Weld width parameter is the maximum transverse melting width of the molten pool on the workpiece surface. It affects the weld formation coefficient and heat input distribution, and is related to residual stress and deformation control.

[0063] In this embodiment, the molten pool temperature data is collected by infrared or thermocouple sensors, representing the measured results of the heat conduction process; the molten pool morphology data is obtained by high-speed vision or laser scanning to acquire the three-dimensional morphology of the molten pool surface. By using an infrared thermal imager to collect a two-dimensional temperature field and combining it with a calibration algorithm to convert it into absolute temperature values, or by using a high-speed CCD camera with structured light projection to obtain a dynamic sequence of the molten pool's three-dimensional morphology, a measured basis is provided for the construction of physical constraints, supporting the data foundation for model training.

[0064] Specifically, the welding area and all-position welding angles of the pressure vessel cylinder to be welded are determined, and laser heat sources and monitoring equipment are deployed in the welding area to ensure that the optical path and field of view cover the molten pool area. Furthermore, the laser processing head and multi-source sensors are installed on the welding actuator to ensure that the beam focus point and the monitoring field of view coincide in the welding area. For example, deploying laser heat sources and monitoring equipment in the welding area using a follow-up coaxial monitoring module ensures that the sensor is always directly facing the center of the molten pool, or a multi-view monitoring array is arranged at a fixed position to cover the entire all-position welding process, thereby establishing the physical conditions for controllable heat input and effective data acquisition.

[0065] All-position welding angles are spatial orientation parameters describing the welding posture relative to the direction of gravity. They are used to characterize flat welding, vertical welding, horizontal welding, and overhead welding conditions. As a key variable affecting the stress state of the molten pool, they are incorporated into the design of measuring point and distribution, as well as physical constraint modeling. All-position welding angles include, but are not limited to, one or more of the following: horizontal surround angle, vertical tilt angle, and local normal angle of the weld. The monitoring position of the monitoring equipment refers to the installation coordinates of the multi-source sensing device in space and the direction of its observation field of view center. It is used to determine the spatial range in which molten pool data can be effectively collected, affecting the feasibility of measuring point coordinates and the physical representativeness of the distribution coordinates.

[0066] Determining the welding area and all-position welding angles of the pressure vessel shell to be welded involves defining the welding path for the circumferential or longitudinal seam based on the shell structure drawings and welding process specifications, and quantifying the corresponding welding posture angles for each segment. Further, determining the welding area and all-position welding angles of the pressure vessel shell to be welded is achieved by extracting the weld centerline from the CAD model and discretizing it into path points with angle labels, or by using robot kinematics inverse kinematics to solve for real-time welding posture angles to form a continuous angle sequence. This explicitly models the welding geometry and posture perturbations, providing prior information for subsequent measurement point-matching collaborative design.

[0067] Based on the welding area, all-position welding angles, and the monitoring positions of the monitoring equipment, the sets of measurement points and collocation points used for training are determined. This involves comprehensively considering geometric accessibility, attitude change characteristics, and physical field sensitivity to generate two types of coordinate sets within the welding area. Furthermore, based on the welding area, all-position welding angles, and the monitoring positions of the monitoring equipment, the sets of measurement points and collocation points used for training are determined by uniformly sampling measurement points over time within the monitoring field of view while simultaneously densifying collocation points in key areas such as the edge and bottom of the molten pool. Alternatively, this can be achieved by dividing the attitude intervals according to the all-position welding angles and independently designing measurement point and collocation point distribution strategies for each attitude type. This constructs a training sample structure that combines data authenticity and physical representativeness, supporting collaborative training of measurement points and collocation points.

[0068] The measurement point set is a collection of coordinates for several measurement points. Each measurement point corresponds to a location where actual molten pool temperature and morphology data can be obtained, providing real physical observation data as a source of supervisory signals for model training. In a specific embodiment, the measurement point set is selected based on the field of view of the monitoring equipment, the welding area, and the all-position welding angle, selecting observable and representative spatiotemporal locations. Furthermore, the measurement point set is directly related to the molten pool temperature data and molten pool morphology data, and its distribution is constrained by the monitoring position of the monitoring equipment and the all-position welding angle. The measurement point coordinates are the specific spatiotemporal location identifiers of each element in the measurement point set, used to uniquely determine the physical location corresponding to the measured data, and to align the observed values ​​with the model output.

[0069] A collocation set is a collection of collocation point coordinates. Each collocation point corresponds to the location in the residual calculation of the physical equation (heat conduction or fluid dynamics). It is used to embed physical constraints, guide the model to satisfy the governing equations, and improve generalization ability. In an exemplary embodiment, the collocation point set is deployed within the welding area according to the gradient of the physical field changes or the requirements of numerical solution, without relying on actual measurements but needing to cover key dynamic regions. For example, the collocation point set includes, but is not limited to, one or more of the following: collocation points with high thermal gradients, collocation points near free surfaces, and collocation points in gravity-sensitive areas. The collocation point coordinates are the specific spatiotemporal location identifiers of each element in the collocation point set, used as sampling points for the residual calculation of the physical equations and participating in the construction of the model loss function.

[0070] The coordinates of the measuring points, the coordinates of the collocation points, and the all-position welding angles of the pressure vessel cylinder to be welded are normalized. Normalization is a data preprocessing operation that converts the measuring point coordinates, collocation point coordinates, and all-position welding angles into a unified dimensionless scale or standardized representation. This is used to eliminate input feature distribution offsets caused by differences in spatial scale and inconsistent angle units, improving model training stability and cross-pose generalization ability. In this embodiment, the normalization process is explained in context: it linearly maps the original spatial coordinates to the [0, 1] interval according to the geometric boundary of the welding area, and converts the all-position welding angles into a standardized direction representation. Furthermore, normalization includes, but is not limited to, one or more of the following: coordinate linear scaling normalization, angle periodic mapping normalization, and direction vector normalization.

[0071] Furthermore, the coordinates of the measuring points, the coordinates of the collocation points, and the all-position welding angles of the pressure vessel cylinder to be welded are normalized by performing triaxial independent normalization on the spatial coordinates of the measuring points and collocation points according to the axial length, circumferential arc length, and radial thickness of the cylinder, and by converting the all-position welding angle θ into a two-dimensional vector representation of (sinθ, cosθ) while retaining its periodicity and directional information. This makes the input features under different welding postures comparable and enhances the model's ability to uniformly learn the thermo-fluid coupling law.

[0072] For example, in the scenario of a pressure vessel cylinder undergoing a 360° circumferential welding process from 0° flat welding to full circumferential welding, the dynamic molten pool monitoring method for laser welding of the pressure vessel cylinder in this embodiment is as follows: During this process, the coordinates of the measuring points are distributed along the surface of the cylinder, spanning several meters, while the welding angle in all positions continuously changes from 0 degrees to 360°. If the original coordinates (e.g., x = 2.3m, y = 0.8m) and angles (e.g., 270°) are directly input, the model will have difficulty recognizing the symmetrical physical mechanism of overhead welding (270°) and vertical welding (90°) under the action of gravity. Through normalization, all spatial coordinates are compressed into a [0, 1]³ cube, and the angles are mapped to (sinθ, cosθ) vectors; for example, 270° becomes (−1, 0), and 90° becomes (1, 0), making it easier for the model to learn the influence of gravity direction on the molten pool. In subsequent training with thermal-fluid coupling constraints, the normalized input significantly accelerated convergence and enabled the target weld pool monitoring model to accurately predict weld depth and stability even in unseen 135-degree horizontal welding positions.

[0073] By normalizing the coordinates of the measuring points, the coordinates of the collocation points, and the all-position welding angles of the pressure vessel cylinder to be welded, the spatial position is mapped to a unified dimensionless interval, and the angle is converted into a standardized representation that retains periodicity and directionality. This eliminates the input feature distribution offset caused by differences in geometric scale and inconsistent angle units in all-position welding, enabling the data-driven model to focus on the physical essence of heat-fluid coupling rather than surface geometric differences. This achieves the technical effect of improving the training stability, convergence speed, and generalization ability across welding postures of the model that integrates the constraints of heat conduction equations and fluid dynamics equations.

[0074] The pressure vessel shell to be welded is a cylindrical pressure-bearing structural component that requires all-position laser welding. Its geometric characteristics and material properties determine the boundary conditions of the dynamic behavior of the molten pool.

[0075] The data on the temperature and shape of the molten pool at the coordinates of the measuring points on the cylinder of the pressure vessel to be welded are obtained. The spatial position is then deduced by using timestamps and welding speed to achieve spatiotemporal alignment of non-fixed-point measurements. Alternatively, a calibration plate and a 3D reconstruction algorithm can be used to map the pixel coordinates to the measuring point coordinates in the workpiece coordinate system, thereby providing the model with a measured monitoring signal that is strictly aligned with a specific spatiotemporal position.

[0076] Taking the monitoring of the transition section from 90° vertical welding to 180° overhead welding of the circumferential weld of a pressure vessel shell as an example, the dynamic molten pool monitoring method for laser welding of the pressure vessel shell in this embodiment involves a continuous change in the welding angle from vertically upward to inverted downward in this transition section, causing a drastic change in the direction of gravity's effect on the molten pool. The system first identifies this section as the critical welding area and quantifies the angle change curve. Subsequently, a coaxial high-speed camera and an infrared thermal imager are deployed on the outside of the shell, ensuring that the monitoring position covers the molten pool throughout the entire process. Based on the angle change rate and the monitoring field of view, the system densely deploys measurement point coordinates in the high dynamic zone (such as around 135°) and sets matching point coordinates in the easily collapsible area at the bottom of the molten pool. During actual welding, temperature and morphological data are collected only at the measurement point coordinates, while the matching point coordinates are not measured but participate in the calculation of the thermal-fluid equation residuals. Finally, the trained model can accurately predict the stability of the molten pool at any angle in this transition section, avoiding undercut or collapse caused by sudden changes in gravity.

[0077] By determining the welding area and all-position welding angle of the pressure vessel cylinder to be welded, and deploying laser heat sources and monitoring equipment in the welding area, the set of measurement points and the set of collocation points for training are determined based on the welding area, all-position welding angle, and monitoring positions of the monitoring equipment. By acquiring the molten pool temperature data and molten pool morphology data at the coordinates of the measurement points on the pressure vessel cylinder to be welded, welding posture disturbance factors are explicitly introduced. The set of measurement points and the set of collocation points are scientifically constructed to form a training sample structure driven by both data and physics. The measurement points provide real observations of molten pool temperature and morphology, and the collocation points embed the residuals of heat conduction and fluid dynamics equations. Thus, the model can simultaneously satisfy data fitting and physical consistency during training, effectively overcoming the shortcomings of traditional pure data-driven methods that ignore multi-field coupling mechanisms. This significantly improves the robustness and accuracy of the model in predicting molten pool morphology evolution, heat accumulation effect, and stability state in all-position welding, laying a dual foundation for high-precision online monitoring.

[0078] Step S02: Construct a molten pool monitoring model based on a feedforward deep neural network, obtain temperature prediction values ​​through the molten pool monitoring model, construct molten pool thermodynamic constraints based on the heat conduction equation and temperature prediction values, and train the molten pool monitoring model based on the thermodynamic constraints to obtain the first target molten pool monitoring model.

[0079] The molten pool monitoring model is a trainable mathematical model based on a data-driven architecture used to predict the dynamic evolution of the molten pool state during welding. This model is based on a feedforward deep neural network, and by embedding physical constraints into the training process, it achieves a fusion of data-driven methods and heat conduction mechanisms. This maintains the flexibility of the neural network while ensuring its adherence to physical laws and possessing good generalization ability. The input layer of the molten pool monitoring model receives spatiotemporal coordinate data, including spatial location and welding time; the hidden layer consists of multiple fully connected layers, each employing a nonlinear activation function, extracting high-order features of the molten pool temperature field through layer-by-layer nonlinear transformation; the output layer directly generates the predicted value of the molten pool temperature field.

[0080] The temperature, thermal conductivity, and thermal diffusivity at the locations of the measuring points and coordinate points of the pressure vessel cylinder to be welded are predicted using the molten pool monitoring model, and the predicted temperature value is obtained.

[0081] Temperature, as a thermodynamic state quantity within and at a specific spatiotemporal location of the molten pool, serves as a core variable in the heat conduction process, directly determining material phase transformation and thermal stress distribution. Thermal conductivity, a property parameter reflecting a material's ability to conduct heat, characterizes the heat flux density per unit temperature gradient and influences heat flux calculations in the heat conduction equation; it is a crucial input parameter for temperature field evolution. The thermal diffusivity, a comprehensive property parameter reflecting the rate of temperature equilibrium within the material, is equal to thermal conductivity divided by the product of density and specific heat capacity. It determines the time-scale characteristics of the heat conduction equation, affecting heat accumulation and heat dissipation dynamics. In this embodiment, the predicted temperature value is calculated by the model based on the input spatiotemporal coordinates, synchronously predicting the temperature and its related property parameters through an internal network structure, and then substituting them into the heat conduction operator. Furthermore, the predicted temperature value depends on the coordinates of the measurement point and the collocation point as input locations; its accuracy directly affects the magnitude of the molten pool's thermodynamic constraint loss. For example, this operation is achieved by using multi-output neural network branches to predict temperature, thermal conductivity, and thermal diffusivity separately, and then combining them to calculate the evolution of the temperature field; or by setting thermal conductivity and thermal diffusivity as implicit functions of temperature, and learning their nonlinear relationship end-to-end by the model, thereby achieving joint modeling of the thermal field and its physical parameters, and providing complete input for the residual calculation of the heat conduction equation.

[0082] The molten pool thermodynamic constraint loss is a scalar loss term used to quantify the deviation between the model output and the physical laws of heat conduction. In this model, a Physical Information Neural Network (PINN) framework is employed to embed the heat conduction equation as a soft constraint into the loss function, ensuring that it strictly satisfies the laws of heat conduction while fitting the data.

[0083] The constraint can be constructed using the Navier-Stokes equations, the continuity equation, and the free surface dynamics model. The residuals between the theoretical flow field and the model prediction are calculated at the measurement and collocation points to form physical constraint expressions. At the measurement points, data fitting terms are constructed using measured molten pool temperature data; at the collocation points, physical constraint terms are formed by calculating the difference between the theoretical heat flux of the heat conduction equation and the spatiotemporal derivative of the model-predicted temperature. Alternatively, for different all-position welding angles, the weights of the surface tension and gravity terms in the equations can be dynamically adjusted to enhance attitude adaptability. This allows the fundamental laws of fluid mechanics to be embedded into the data-driven framework, ensuring that the model output conforms to the conservation of mass and momentum. Specific loss terms are calculated by weighted summation of the squared residuals of the heat conduction equation at all measurement and collocation points (such as velocity divergence, the imbalance of density multiplied by acceleration and pressure gradient, and the sum of viscous and volume forces). These losses can take the form of one or more of the following: measurement point thermal residual loss, collocation point thermal residual loss, and boundary heat flux mismatch loss. The loss function is fully differentiable and serves as a key optimization objective in parameter updates during training. It directly guides the model to learn temperature field evolution behavior that conforms to real physical laws, thereby ensuring the thermophysical consistency of the output.

[0084] Specifically, the formula for calculating the heat conduction equation is as follows: In the formula The temperature field of the molten pool describes the temperature distribution at various points within the molten pool during the welding process. The temperature field is the core variable of the molten pool thermodynamics, directly affecting the melting and solidification process of the material and the welding quality, such as the depth and width of the weld. For time, it represents the dynamic stages of the welding process; The thermal diffusivity reflects a material's ability to transfer heat. It is a spatial second derivative operator, which mathematically represents the spatial curvature of the temperature field and physically reflects the direction and rate of heat diffusion; These are spatial coordinates, representing the position of any point within the molten pool; For heat source terms, it represents the heat input power density per unit volume, i.e., the energy contribution of the heat source to the molten pool during laser welding.

[0085] The calculation formula for the thermodynamic constraint of the molten pool is as follows:

[0086] ;

[0087] In the formula, To measure the degree of agreement between the predictions of the thermodynamic model and the physical laws and measured data; To measure the difference between the temperature field predicted by the model and the measured temperature data, and to ensure that the model conforms to actual observations; To measure the degree of deviation between the temperature field predicted by the model and the heat conduction equation, and to ensure that the model conforms to the physical laws of heat conduction; The temperature field predicted by the model, including spatiotemporal coordinates. , The function output; The temperature data measured in the experiment comes from the sensor and exists only at a limited number of measuring points; Let be the partial derivative with respect to time t, describing the rate of change of temperature over time.

[0088] This constraint, working in conjunction with the measured molten pool temperature data, ensures that the model output both fits the observed data and strictly satisfies energy conservation and thermophysical laws. Simultaneously, it complements the structural design of the molten pool monitoring model, serving as a key physical constraint during the training phase, effectively enhancing the model's generalization ability and physical reliability.

[0089] During training, molten pool temperature data is used as a supervision signal, and the residual of the heat conduction equation is used as a regularization term to optimize model parameters. The model not only minimizes the error between the predicted and measured temperatures but also introduces the heat conduction equation as a physical constraint term, calculating its residual and incorporating it into the total loss function. Automatic differentiation techniques are used to calculate the required partial derivatives, ensuring that the physical constraints are differentiable and participate in gradient backpropagation. The total loss function is a weighted sum of the data fitting term and the physical constraint term, and the network parameters are updated synchronously through optimization algorithms.

[0090] This model can accurately predict the dynamic changes of the molten pool temperature field based on measured temperature data obtained from a limited number of sensors, combined with physical constraints acting on the entire computational domain, and strictly follows the laws of heat conduction. This architecture maintains the expressive power of neural networks while enhancing the physical reliability of the model, making it suitable for monitoring the state of molten pools under complex operating conditions.

[0091] To minimize the thermodynamic constraint loss of the molten pool monitoring model, the training and parameter updates of the model employ a gradient descent-type optimization algorithm, backpropagating the thermodynamic constraint loss to adjust the model parameters. For example, this operation achieves dual-objective optimization of data and physics by jointly minimizing the thermodynamic constraint loss and the mean square error of the measured temperature data; alternatively, a learning-by-course strategy can be used, first optimizing the measurement point fitting loss and then gradually introducing the collocational point physical constraint loss, thereby enabling the model to gradually satisfy the physical laws of heat conduction during training, improving the extrapolation ability and robustness of thermal field prediction.

[0092] Taking the modeling of heat accumulation effect in the overhead welding section of a pressure vessel cylinder as an example, the dynamic molten pool monitoring method for laser welding of pressure vessel cylinders in this embodiment is based on the fact that in the overhead welding position, the molten pool is prone to sinking due to gravity, and multiple laser scans lead to significant local heat accumulation. The system deploys measuring points and auxiliary points in this area: the measuring points acquire actual temperatures using infrared cameras, while the auxiliary points are densely distributed at the junction of the bottom of the molten pool and the heat-affected zone. The molten pool monitoring model receives these coordinates and simultaneously predicts the temperature, thermal conductivity, and thermal diffusivity of each point. Since the thermal conductivity decreases due to the increased material temperature during overhead welding, ignoring this nonlinearity would overestimate the heat dissipation rate. By constructing the residual of the heat conduction equation and calculating the thermodynamic constraint loss of the molten pool, the training process forces the model to learn a positive feedback mechanism of "high temperature, low thermal conductivity, and increased heat accumulation." Ultimately, the model can accurately predict the local overheating risk in the overhead welding section, providing a basis for real-time laser power control.

[0093] This embodiment provides a method for monitoring the dynamic molten pool at all positions during laser welding of pressure vessel cylinders. The method uses a molten pool monitoring model to jointly predict the temperature, thermal conductivity, and thermal diffusivity at the coordinates of the measuring points and collocation points, generating physically consistent temperature prediction values. Based on the set of measuring points, collocation points, and temperature prediction values, the method constructs molten pool thermodynamic constraints using the residuals of the heat conduction equation and quantifies them as molten pool thermodynamic constraint losses. By training the model with the goal of minimizing these losses, the model can strictly adhere to the physical laws of heat conduction while fitting measured temperature data. This effectively overcomes the shortcomings of traditional pure data-driven methods that ignore the thermal field evolution mechanism, significantly improving the model's accuracy and robustness in modeling key thermal behaviors such as heat accumulation effects and local thermal gradients. This provides a high-fidelity thermal field foundation for high-precision online inversion of molten pool states.

[0094] Step S03: Obtain the predicted value of the molten pool shape through the molten pool monitoring model, construct the molten pool fluid dynamic constraints based on the fluid dynamic equation and the predicted value of the molten pool shape, and train the first target molten pool monitoring model based on the fluid dynamic constraints to obtain the second target molten pool monitoring model.

[0095] By using a molten pool monitoring model, the velocity field, pressure field, and surface morphology of the molten pool at the locations of the measuring points and collocation points of the pressure vessel cylinder to be welded are obtained, thus yielding a predicted value for the molten pool morphology.

[0096] The velocity field describes the distribution of velocity vectors of molten metal at various points in space within the molten pool. It reflects the dynamic characteristics of convection, thermocapillary flow, and gravity-driven flow within the molten pool and is a core variable in fluid dynamics modeling. The pressure field is the pressure distribution state within the molten pool and at the gas-liquid interface. It influences the evolution of the free surface morphology of the molten pool, especially determining the depth and stability of the molten pool depression under laser recoil pressure. The molten pool surface morphology is the three-dimensional geometric contour of the free surface of the molten pool, including morphological features such as weld width, weld depth, and edge curvature. It is directly related to the welding quality and is the external manifestation of defects such as undercut and collapse. The predicted molten pool morphology is the flow field state estimate output by the molten pool monitoring model after jointly extrapolating the velocity field, pressure field, and molten pool surface morphology at the measuring point coordinates and collocation point coordinates. In this embodiment, the predicted molten pool morphology is predicted by the model based on the input spatiotemporal coordinates, simultaneously predicting velocity, pressure, and surface geometry, and maintaining physical consistency through the fluid control equation structure. Furthermore, the predicted molten pool morphology depends on the coordinates of the measuring points and collocation points as input locations; its completeness directly affects the accuracy of the molten pool fluid dynamics constraint loss. For example, this operation can be achieved by using a multi-head neural network to output velocity components, pressure, and surface height separately, and then combining them into a unified flow field representation; or by introducing a level set or VOF implicit function to represent the free surface, allowing the model to directly predict the interface function and its evolution, thereby achieving joint modeling of multiple physical quantities of the molten pool and providing complete state variables for the residual calculation of the fluid dynamics equations.

[0097] The molten pool hydrodynamic constraint loss is a scalar loss term that quantifies the deviation between the output of the primary target molten pool monitoring model and the physical laws of hydrodynamics. In an exemplary embodiment, the molten pool hydrodynamic constraint loss is a weighted sum of the squared residuals of the mass conservation equation, momentum equation (such as Navier-Stokes), and free surface boundary conditions at the measurement and collocation points. Furthermore, the molten pool hydrodynamic constraint loss includes, but is not limited to, one or more of the following: momentum equation residual loss, continuity equation residual loss, and free surface curvature-pressure matching loss. In a specific embodiment, this operation is achieved by constructing data fitting terms at the measurement points using the measured molten pool surface morphology, and constructing physical constraint terms at the collocation points using only the fluid equation residuals; or by dynamically adjusting the weights of the surface tension and gravity terms in the equations for different all-position welding angles to enhance attitude adaptability, thereby embedding the fundamental laws of hydrodynamics into the data-driven framework and ensuring that the model output conforms to the conservation of mass and momentum.

[0098] In this embodiment, the formula for the molten pool hydrodynamic constraint loss is as follows:

[0099]

[0100] In the formula, Used to measure the degree of agreement between the predictions of a fluid dynamics model and the measured fluid morphology and fluid dynamics equations; The difference between the fluid state predicted by the model and the measured fluid morphology data is measured, and includes three types of variables; among them For velocity field, that is, the flow velocity of liquid metal in molten pool; This refers to the pressure field, specifically the static pressure distribution within the molten pool. Surface morphology refers to the height variation of the free surface of the molten pool; This is to measure the degree of deviation between the fluid field predicted by the model and the fundamental equations of fluid mechanics.

[0101] The molten pool fluid dynamics constraint is a joint physical constraint loss function formed by integrating the molten pool thermodynamic constraint loss and the molten pool fluid dynamics constraint loss. As a unified optimization objective, it guides the model to simultaneously satisfy the physical laws of the thermal field and the flow field and their interaction mechanism.

[0102] Molten pool hydrodynamic constraints are implemented by calculating flow field residuals using fluid control equations and incorporating them into the model optimization objective. The first-target molten pool monitoring model is trained based on these hydrodynamic constraints, using molten pool morphology data as a supervisory signal and fluid equation residuals as constraints to update model parameters. This enhances the model's ability to identify and predict dynamic changes in molten pool morphology (such as collapse and edge biting). Furthermore, the molten pool hydrodynamic constraints and molten pool morphology data collaboratively provide experimental evidence of the flow field and work in synergy with the first-target molten pool monitoring model to optimize its flow field prediction branch.

[0103] Taking the prediction of molten pool collapse in the transition zone from vertical to overhead welding of pressure vessel cylinders as an example, the dynamic molten pool monitoring method for all positions in laser welding of pressure vessel cylinders in this embodiment is as follows: in this region, the direction of gravity gradually changes from vertically downward to upward, and the bottom of the molten pool is prone to collapse due to loss of support. The system deploys densely spaced points near the free surface at the bottom of the molten pool in this section, and sets measuring points for high-speed cameras to collect the actual morphology. The first target molten pool monitoring model receives these coordinates and simultaneously predicts the velocity field (displaying downward backflow), pressure field (displaying the low-pressure area at the bottom), and surface morphology (predicting the downward trend). Based on the Navier-Stokes equations and the free surface curvature-pressure relationship, the fluid dynamic constraint loss is calculated; if the model does not correctly capture the gravity-surface tension balance, the loss will increase significantly. During training, the model is forced to learn the physical links of "reversal of gravity direction, insufficient bottom pressure, and surface curvature instability," ultimately enabling it to provide early warning of collapse risks and trigger adaptive adjustments to laser power or welding speed.

[0104] This embodiment provides a method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders. The method uses a molten pool monitoring model to jointly predict the velocity field, pressure field, and molten pool surface morphology at the coordinates of the measuring points and collocation points, generating physically consistent predicted values ​​for the molten pool morphology. Based on the set of measuring points, collocation points, and the predicted molten pool morphology, the method constructs molten pool fluid dynamic constraints using the residuals of the fluid dynamics equations and quantifies them as molten pool fluid dynamic constraint losses. The model is then trained with the goal of minimizing these losses, ensuring that it fits the measured molten pool morphology data while strictly adhering to physical laws such as mass conservation, momentum conservation, and free surface dynamics. This effectively overcomes the shortcomings of traditional methods that neglect changes in gravity direction and the coupling effects of multiple force fields. It significantly improves the model's ability to model complex fluid phenomena such as drastic molten pool deformation and flow instability in all-position welding, and enhances its generalization robustness. This provides a reliable flow field basis for high-precision online inversion of weld depth, weld width, and stability, thereby supporting the technical effect of intelligent welding closed-loop control.

[0105] Step S04: Construct a thermal-fluid coupling constraint for the molten pool, and train the second target molten pool monitoring model based on the thermal-fluid coupling constraint to obtain the third target molten pool monitoring model.

[0106] The molten pool thermal-fluid coupling constraint is a multi-physics joint constraint mechanism formed by integrating heat conduction and fluid dynamics equations. It is used to achieve collaborative modeling of the temperature field and flow field, improving the model's adaptability to complex molten pool behavior in all-position welding. In this embodiment, the molten pool thermal-fluid coupling constraint embeds the thermal field and flow field control equations into the model loss function, forming coupled physical residual terms. This operation simultaneously solves the energy equation and momentum equation within a unified computational domain, constructing the coupled loss function, or introducing thermocapillary convection terms to explicitly model the influence of temperature gradients on surface tension and flow field. This enables mutual feedback modeling of the temperature field and flow field, improving the state inversion capability under complex working conditions. Furthermore, the molten pool thermal-fluid coupling constraint includes, but is not limited to, one or more of the following: thermally induced buoyancy coupling constraint, surface tension-temperature gradient coupling constraint, and gravity-thermal accumulation interaction constraint.

[0107] The determination of the molten pool thermal-fluid coupling constraint loss based on the molten pool thermodynamic constraint loss and molten pool hydrodynamic constraint loss of the second target molten pool monitoring model involves mathematically combining the calculated thermodynamic constraint loss and hydrodynamic constraint loss to form a single joint loss term. In an exemplary embodiment, this operation is achieved by using a weighted summation method, whereby the joint loss is the sum of the thermal loss and the fluid loss multiplied by their respective weights, where the weights are dynamically adjusted according to the welding posture. For example, cross-coupling terms can also be introduced, such as substituting the temperature gradient into the surface tension gradient term before calculating the fluid residual, forming a nonlinear joint loss. This achieves a unified expression of the physical constraints of the thermal field and the flow field, enabling the model training process to synchronously optimize the consistency of the multiphysics field.

[0108] In this embodiment, the formula for the thermal-fluid coupling constraint loss of the molten pool is: In the formula, This is a weighted sum of heat conduction losses and fluid dynamics losses; Controlling thermal conduction constraints The proportion of contribution to the total loss; when > When the model prioritizes satisfying the fitting accuracy of the temperature field and the constraints of the heat conduction equation, it prioritizes satisfying the constraints of the temperature field fitting accuracy and the heat conduction equation. When = 0, the model degenerates into a pure fluid dynamics model, and thermal effects are ignored; Controlled fluid dynamics constraints The proportion of contribution to the total loss; when > When the model prioritizes satisfying the fitting accuracy of the velocity field, pressure field, molten pool morphology, and fluid dynamics equation constraints; when When = 0, the model degenerates into a pure heat conduction model, ignoring the influence of fluid flow on heat transport.

[0109] The goal is to minimize the thermal-fluid coupling constraint loss of the second target molten pool monitoring model. Training and parameter updates for this model utilize a gradient descent algorithm to backpropagate the joint loss, simultaneously updating model parameters to reduce physical residuals in both thermal and fluid aspects. In one specific embodiment, this operation is achieved through an alternating optimization strategy: first, the flow field branch is fixed to optimize the thermal field, then the thermal field is fixed to optimize the flow field, iterating repeatedly until convergence. Furthermore, a higher learning rate can be applied to the coupling loss in all key attitude ranges (such as the overhead welding section) to accelerate the physical consistency convergence of sensitive areas. This allows the model to automatically learn the nonlinear coupling relationship between the temperature field and the flow field during training, improving its ability to model complex dynamic molten pool behavior.

[0110] The second target molten pool monitoring model, trained based on thermo-fluid coupling constraints, jointly minimizes the error of measured data and the residuals of multiphysics equations to achieve final model optimization, thereby obtaining a target molten pool monitoring model with strong physical consistency and high generalization ability. The third target molten pool monitoring model is used to invert the dynamic state of the molten pool in real time based on spatiotemporal coordinates, supporting online quality monitoring. In an exemplary embodiment, the target molten pool monitoring model receives the spatiotemporal coordinates of the location to be measured as input and outputs the dynamic state of the molten pool; it also relies on various previously constructed physical constraints to complete training.

[0111] Taking the thermal-fluid coupling instability early warning in all-position circumferential welding of pressure vessel shells as an example, the dynamic molten pool monitoring method for all-position laser welding of pressure vessel shells in this embodiment is as follows: during the transition from horizontal welding to overhead welding, the molten pool simultaneously experiences increased heat accumulation and loss of gravity support. Traditional models, due to the lack of coupled thermal-fluid interaction, may underestimate the surface tension-driven convection intensity or misjudge the surface tension equilibrium point. In this scheme, the target molten pool monitoring model, during the training phase, forces the learning of a positive feedback loop of "local high temperature, reduced surface tension, accelerated backflow at the edge of the molten pool, and deepened central depression" through the thermal-fluid coupling constraint loss of the molten pool. When the spatiotemporal coordinates of a certain overhead welding position are input, the model not only outputs the temperature distribution but also simultaneously predicts the velocity field driven by the thermal gradient and the resulting free surface collapse trend; if the coupled prediction shows a sudden increase in molten depth and abnormal edge flow velocity, the system determines it as a precursor to instability and triggers process intervention. This capability stems from the explicit encoding of the multiphysics interaction mechanism by the thermal-fluid coupling constraint loss.

[0112] This embodiment provides a method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders. It determines the thermal-fluid coupling constraint loss of the molten pool by considering both the thermodynamic and hydrodynamic constraint losses based on the molten pool monitoring model. The molten pool monitoring model is trained and its parameters are updated with the goal of minimizing this thermal-fluid coupling constraint loss, resulting in a target molten pool monitoring model. By integrating thermodynamic and hydrodynamic constraints into a unified joint loss to synchronously guide model parameter updates, the model simultaneously satisfies the laws of heat conduction and fluid motion conservation during training and automatically learns the nonlinearity between them. This effectively overcomes the weak generalization ability and poor robustness caused by traditional methods neglecting the thermal-fluid coupling mechanism. Especially in all-position welding, it maintains high-precision state inversion capability even under extreme dynamic conditions such as continuous changes in gravity vectors and drastic fluctuations in molten pool morphology. Ultimately, the target molten pool monitoring model can accurately predict key quality indicators such as weld depth, weld width, and molten pool stability in real time based solely on the spatiotemporal coordinates of the measured position, providing online monitoring capabilities with both physical consistency and data adaptability for intelligent welding of high-safety-level pressure vessels.

[0113] Step S05: Obtain the spatiotemporal coordinates of the test position of the pressure vessel cylinder to be welded, and determine the dynamic state of the molten pool at the test position of the pressure vessel cylinder to be welded through the third target molten pool monitoring model.

[0114] The spatiotemporal coordinates of the location to be measured are the three-dimensional coordinates of a certain position on the welding path in space and its corresponding timestamp. These are used as input variables for the target molten pool monitoring model to locate the state of the molten pool at a specific time and location.

[0115] Obtaining the spatiotemporal coordinates of the test position of the pressure vessel cylinder to be welded involves extracting the spatial position and time information of the current weld point from the welding path planning system or motion controller, thereby providing a uniquely determined input query point for the third target model.

[0116] Determining the dynamic state of the molten pool at the target location using a third-target molten pool monitoring model involves inputting spatiotemporal coordinates into a trained model and outputting the corresponding molten pool temperature field, flow field, and geometric parameters. This enables real-time inversion of the molten pool state without direct sensing, supporting online quality assessment. Furthermore, this operation can be achieved through a real-time data channel between the welding control system and the model interface, thus realizing a sensorless state query mechanism and reducing reliance on real-time hardware acquisition.

[0117] Obtaining the molten pool temperature and morphology of the pressure vessel cylinder at the test location is achieved by the third-target molten pool monitoring model, which outputs corresponding predicted values ​​of the molten pool temperature and morphology fields based on the input spatiotemporal coordinates through internal mapping relationships. Furthermore, this operation replaces direct measurement with a data-driven approach, enabling the acquisition of high spatiotemporal resolution estimates of the molten pool's physical state.

[0118] Based on the molten pool temperature and morphology of the pressure vessel shell at the test location, the dynamic state of the molten pool at that location is determined. This involves mapping the temperature and morphological data output by the model into engineering-interpretable quality indicators (melt depth, weld width, and stability). Furthermore, this operation extracts melt depth and weld width from the morphological data using preset geometric rules (such as molten pool cross-section fitting), or constructs a stability scoring function based on temperature gradients and morphological fluctuation characteristics to quantify the risk of molten pool instability. This completes the semantic transformation from physical field prediction to process quality parameters, supporting closed-loop control decisions.

[0119] Molten pool stability refers to the ability of the molten pool to maintain its geometric and thermodynamic equilibrium during welding. It is used to assess whether instability phenomena such as spattering, collapse, or oscillation occur, and is a key criterion for weld quality. In a specific embodiment, molten pool stability includes, but is not limited to, thermal-fluid coupling stability, resistance to gravitational disturbances, and surface tension self-healing ability. Penetration depth is the maximum melting depth of the molten pool along the thickness direction of the workpiece. It directly affects the joint's load-bearing capacity and sealing performance and is a core acceptance indicator for pressure vessel welding. Weld width parameter is the maximum transverse melting width of the molten pool on the workpiece surface. It affects the weld formation coefficient and heat input distribution, and is related to residual stress and deformation control.

[0120] Taking the online quality assessment of the overhead welding section of a pressure vessel shell as an example, the dynamic molten pool monitoring method for all positions in laser welding of the pressure vessel shell in this embodiment can be used in the overhead welding position. Due to the opposite direction of gravity to the normal of the molten pool, the molten pool is prone to sagging and incomplete fusion. In this embodiment, the welding system acquires the spatiotemporal coordinates of the current weld point in real time and inputs them into the trained target molten pool monitoring model. Based on the embedded thermo-fluid coupling physical law, the model outputs the temperature distribution and three-dimensional shape of the molten pool at this position. The system further analyzes the shape to obtain a weld depth of 4.2 mm and a weld width of 6.8 mm, and judges that the molten pool is in a critical stable state based on the temperature field uniformity. Based on this, the laser power fine-tuning command is triggered to avoid collapse defects and achieve high-quality closed-loop control without sensor intervention.

[0121] This embodiment provides a method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders. By inputting the spatiotemporal coordinates of the location to be measured into the target molten pool monitoring model, the molten pool temperature and morphology at that location are directly obtained. Based on these two parameters, the dynamic state of the molten pool, including weld depth, weld width, and molten pool stability, is determined. The target molten pool monitoring model integrates the heat conduction equation and the fluid dynamics equation to construct a thermo-fluid coupling constraint during the training phase. A collaborative optimization strategy of measurement points and collocation points is adopted. The model internally encodes the evolution law of multi-physics coupling in all-position welding. Key quality parameters can be accurately inverted using only spatiotemporal coordinates. This achieves online, non-sensory, and robust monitoring of core welding quality indicators without relying on real-time synchronous acquisition of multiple sensors. This provides a quantifiable and predictable state feedback basis for the intelligent welding closed-loop control of high-safety-level pressure vessels.

[0122] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders, characterized in that, The method includes: Step S01: Deploy laser heat source and monitoring equipment to obtain molten pool temperature data and molten pool morphology data of the pressure vessel cylinder to be welded, and determine the set of measurement points and matching points for training based on the welding area, all-position welding angle and monitoring position of the monitoring equipment. Step S02: Construct a molten pool monitoring model based on a feedforward deep neural network, obtain temperature prediction values ​​through the molten pool monitoring model, construct molten pool thermodynamic constraints based on the heat conduction equation and temperature prediction values, and train the molten pool monitoring model based on the thermodynamic constraints to obtain the first target molten pool monitoring model; Step S03: Obtain the predicted value of the molten pool shape through the molten pool monitoring model, construct the molten pool fluid dynamic constraints based on the fluid dynamic equation and the predicted value of the molten pool shape, and train the first target molten pool monitoring model based on the fluid dynamic constraints to obtain the second target molten pool monitoring model; Step S04: Construct thermal-fluid coupling constraints for the molten pool, and train the second target molten pool monitoring model based on the thermal-fluid coupling constraints to obtain the third target molten pool monitoring model; Step S05: Obtain the spatiotemporal coordinates of the test position of the pressure vessel cylinder to be welded, and determine the dynamic state of the molten pool at the test position of the pressure vessel cylinder to be welded through the third target molten pool monitoring model.

2. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 1, characterized in that, The set of measuring points mentioned in step S01 contains several measuring point coordinates, and each measuring point corresponds to the location where the actual molten pool temperature and morphology data can be obtained; the set of matching points contains several matching point coordinates, and each matching point corresponds to the location for calculating the physical equation residual, which is used to embed physical law constraints.

3. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 1, characterized in that, After determining the set of measurement points and the set of collocation points for training as described in step S01, the method further includes a step of normalizing the coordinates of the measurement points, the coordinates of the collocation points, and the all-position welding angle: linearly mapping the coordinates of the measurement points and the coordinates of the collocation points to the interval [0, 1]; and converting the all-position welding angle θ into a unit vector (sinθ, cosθ).

4. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 1, characterized in that, Step S02, which involves constructing thermodynamic constraints for the molten pool based on the heat conduction equation and predicted temperature, and training the molten pool monitoring model based on these constraints to obtain the first target molten pool monitoring model, includes: The temperature, thermal conductivity, and thermal diffusivity at the locations of the measuring points and the coordinates of the matching points on the pressure vessel cylinder to be welded are predicted using the molten pool monitoring model, and the predicted temperature value is obtained. Based on the set of measuring points and the set of matching points of the pressure vessel cylinder to be welded, as well as the predicted temperature value, a thermodynamic constraint of the molten pool is constructed, and the thermodynamic constraint loss of the molten pool monitoring model is determined. With the goal of minimizing the thermodynamic constraint loss of the molten pool monitoring model, the molten pool monitoring model is trained and its parameters are updated.

5. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 4, characterized in that, The calculation formula for the thermodynamic constraint of the molten pool is as follows: ， In the formula, To measure the degree of agreement between the predictions of the thermodynamic model and the physical laws and measured data; To measure the difference between the temperature field predicted by the model and the measured temperature data; To measure the degree of deviation between the temperature field predicted by the model and the heat conduction equation; The temperature field predicted by the model, including spatiotemporal coordinates. , The function output; The temperature data measured in the experiment comes from the sensor and exists only at a limited number of measuring points; Let be the partial derivative with respect to time t, describing the rate of change of temperature over time.

6. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 1, characterized in that, Step S03, which involves constructing molten pool fluid dynamic constraints based on fluid dynamic equations and predicted molten pool morphology, and training the first target molten pool monitoring model based on these constraints to obtain the second target molten pool monitoring model, includes: The velocity field, pressure field, and surface morphology of the molten pool are obtained at the locations of the measuring point coordinates and the collocation point coordinates of the pressure vessel cylinder to be welded by the molten pool monitoring model, and the predicted value of the molten pool morphology is obtained. Based on the set of measuring points and the set of matching points of the pressure vessel cylinder to be welded, as well as the predicted value of the molten pool morphology, a molten pool fluid dynamic constraint is constructed, and the molten pool fluid dynamic constraint loss of the molten pool monitoring model is determined. With the goal of minimizing the fluid dynamics constraint loss of the molten pool monitoring model, the first target molten pool monitoring model is trained and its parameters are updated.

7. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 6, characterized in that, The formula for the molten pool hydrodynamic constraint loss is as follows: ， In the formula, Used to measure the degree of agreement between the predictions of a fluid dynamics model and the measured fluid morphology and fluid dynamics equations; The difference between the fluid state predicted by the model and the measured fluid morphology data is measured, and includes three types of variables; among them For velocity field, that is, the flow velocity of liquid metal in molten pool; This refers to the pressure field, specifically the static pressure distribution within the molten pool. Surface morphology refers to the height variation of the free surface of the molten pool; This is to measure the degree of deviation between the fluid field predicted by the model and the fundamental equations of fluid mechanics.

8. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 1, characterized in that, The specific steps of step S04 are as follows: Determine the thermal-fluid coupling constraint loss of the molten pool based on the thermodynamic constraint loss and fluid dynamic constraint loss of the molten pool of the second target molten pool monitoring model; With the goal of minimizing the thermal-fluid coupling constraint loss of the molten pool monitoring model, the second target molten pool monitoring model is trained and its parameters are updated to obtain the third target molten pool monitoring model.

9. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 8, characterized in that, The formula for the thermal-fluid coupling constraint loss of the molten pool is: ， In the formula, This is a weighted sum of heat conduction losses and fluid dynamics losses; Controlling thermal conduction constraints The proportion of contribution to the total loss; when > When the model prioritizes satisfying the fitting accuracy of the temperature field and the constraints of the heat conduction equation, it prioritizes satisfying the constraints of the temperature field fitting accuracy and the heat conduction equation. When = 0, the model degenerates into a pure fluid dynamics model, and thermal effects are ignored; Controlled fluid dynamics constraints The proportion of contribution to the total loss; when > When the model prioritizes satisfying the fitting accuracy of the velocity field, pressure field, molten pool morphology, and fluid dynamics equation constraints; when When = 0, the model degenerates into a pure heat conduction model, ignoring the influence of fluid flow on heat transport.

10. The method for monitoring the dynamic molten pool in all positions during laser welding of pressure vessel cylinders according to claim 1, characterized in that, The step S05, which involves determining the dynamic state of the molten pool at the test location of the pressure vessel cylinder to be welded using the third target molten pool monitoring model, includes: The spatiotemporal coordinates of the test position of the pressure vessel cylinder to be welded are input into the third target molten pool monitoring model to obtain the molten pool temperature and molten pool morphology of the pressure vessel cylinder to be welded at the test position. Based on the molten pool temperature and molten pool morphology at the test location of the pressure vessel cylinder to be welded, the dynamic state of the molten pool at the test location of the pressure vessel cylinder to be welded is determined. The dynamic state of the molten pool includes parameters such as molten pool stability, molten depth, and molten width.

Images