A method and system for online control of electric arc additive manufacturing process based on digital twin.

Abstract

This invention discloses an online control method and system for arc additive manufacturing processes based on digital twins. The method includes the following steps: real-time acquisition of sensing data from the physical printing site, updating the status of the virtual workpiece and virtual print head in the digital twin system; using an online prediction model to calculate the probability of quality anomalies occurring in the predicted area within a preset future time period, generating a risk map R; recalculating and optimizing subsequent printing paths and corresponding process parameters, generating optimization instructions; executing simulated printing to verify the optimization instructions; and after confirming successful verification, sending the optimization instructions to the physical printer for execution. This invention systematically overcomes core defects such as control lag, model isolation, disconnect between path and parameter optimization, unsmooth virtual image updates, and lack of forward-looking verification, elevating the process control mode of arc additive manufacturing from a passive, post-event corrective response to an active, pre-event preventive intelligent control.

Description

Technical Field

[0001] This invention relates to the field of electric arc additive manufacturing technology, specifically to an online control method and system for electric arc additive manufacturing processes based on digital twins. Background Technology

[0002] Arc additive manufacturing is an advanced manufacturing technology that uses an electric arc as a heat source to melt metal wire and deposit it layer by layer along a predetermined path to form metal parts. However, this process involves complex multi-physics coupling effects, and the forming quality is extremely sensitive to process factors such as heat input, path planning, and interlayer morphology, making it prone to defects such as molten pool instability, uneven layer height, and heat accumulation deformation. To improve the quality and process controllability of arc additive manufacturing, the industry has mainly explored three aspects: process monitoring and feedback control, application of digital twin architecture, and path planning.

[0003] In process monitoring and feedback control, existing technologies integrate multiple types of sensors, such as visual and temperature sensors, to monitor key indicators like molten pool state and layer height online. Some advanced research attempts to combine monitoring data with feedback control mechanisms, such as stabilizing the molten pool morphology by dynamically adjusting heat input, or using deep learning algorithms for online defect identification and triggering closed-loop adjustment actions. These technologies form the foundation of current process control. Regarding digital twin architecture applications, digital twin technology provides a framework for virtual-real mapping in arc additive manufacturing. For example, a digital twin architecture based on the OPC UA communication protocol has been proposed, enabling 3D visualization and data integration of the process. More cutting-edge research has proposed the concept of "simulation in the loop," attempting to integrate complex finite element analysis into the printing process for online assessment of structural integrity, extending the digital twin model from a simple monitoring tool to a decision-making aid. In path planning technology, traditional methods are mainly based on offline programming using static geometric models, which is difficult to adapt to dynamically changing printing environments. To address the challenge of forming complex curved surfaces, existing research has introduced planarization techniques, such as mesh unfolding, to assist in path generation. Meanwhile, to meet real-time requirements, GPU parallel acceleration technology is used to quickly process voxel data and generate printing paths, providing underlying computing power support for dynamic adjustments.

[0004] However, the aforementioned existing technologies still have many key shortcomings and bottlenecks in achieving forward-looking and complete real-time intelligent control of the entire arc additive manufacturing process. First, most existing monitoring-feedback systems belong to the "sensing-response" mode, that is, adjustments are only initiated after abnormalities such as layer height deviation and temperature exceeding the limit are detected. This response mechanism has an inherent lag and cannot effectively deal with dynamic problems such as rapid heat accumulation and stress evolution in the arc additive manufacturing process. Although some studies have attempted to integrate simulation prediction, the calculation of complex physical models is extremely time-consuming and cannot meet the real-time decision-making requirements at the second or even sub-second level. Essentially, it has failed to achieve the leap from "post-correction" to "prevention". Second, existing digital twin systems focus on state mapping and visualization, while online prediction models mostly focus on single defect identification. The two have not been deeply integrated, and it is impossible to build a lightweight model that can directly and quickly predict the comprehensive quality risk in the short term based on multi-dimensional data such as real-time geometric state, thermal cycling history and process parameters. This results in a lack of forward-looking basis for control decisions. Furthermore, path optimization and process parameter adjustment are severely disconnected, and the optimization dimension is singular. Existing research generally regards path planning as a static geometric filling or obstacle avoidance problem, while the adjustment of process parameters is based on local temperature control in isolation. Few technologies can synchronously and collaboratively replan the spatial coordinates of subsequent paths and their corresponding process parameters when quality risks are discovered. This results in one-sided control methods and often causes secondary problems such as arc instability and poor wire feeding due to neglecting the kinematic smoothness of the path.

[0005] Furthermore, the updates to the virtual model suffer from unevenness, affecting decision stability. Virtual workpieces require high-frequency updates to synchronize with their physical state, but existing methods often use direct replacement or simple interpolation to process real-time sensor point clouds and other data. This is highly susceptible to data noise interference, which can easily cause jitter or abrupt changes in the virtual model's surface mesh. This unstable virtual representation injects noise into downstream prediction and optimization modules, ultimately leading to decision fluctuations. Finally, existing control systems generally lack a look-ahead verification process for optimization instructions. Typically, the new paths generated by the algorithm are directly sent to the physical equipment for execution. However, whether the new path will cause collisions with the already printed structure and whether its expected quality improvement effect is certain are not effectively verified. This poses a significant safety hazard of sending invalid or high-risk instructions to expensive equipment. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing electric arc additive manufacturing systems, such as lack of real-time feedback, lagging control, and difficulty in coping with dynamic process changes. It proposes an online control method and system for electric arc additive manufacturing processes based on digital twin technology, which can realize real-time monitoring of the forming process, defect prediction, and dynamic optimization of path and process parameters.

[0007] To achieve the above objectives, the present invention is specifically implemented through the following technical solutions: A method for online control of an electric arc additive manufacturing process based on digital twins includes the following steps: S1. Real-time acquisition of sensing data at the physical printing site, including the height contour point cloud of the printed surface, the temperature distribution of the molten pool area, and the real-time position and process parameters of the print head. S2. Update the status of the virtual workpiece and virtual printhead in the digital twin system based on the perceived data; S3. Using an online prediction model, based on the shape, temperature distribution, and process parameters currently in use of the virtual workpiece, calculate the probability of quality anomalies occurring in the predicted area within a preset future time period, and generate a risk map R. S4. When a quality risk area is determined to exist based on the risk map, with the goal of avoiding the quality risk area, the subsequent printing path and corresponding process parameters are recalculated and optimized to generate optimization instructions. S5. Before sending the optimization instructions to the physical printer for execution, a simulated test print is performed in the digital twin system to verify the optimization instructions; S6. After confirming that the verification is successful, the optimization instruction is sent to the physical printer for execution; after the physical printer executes the new instruction, it returns to step S2 and repeats the process until printing is complete.

[0008] Further, in step S2, updating the state of the virtual workpiece and virtual printhead in the digital twin system based on the perceived data specifically includes: Suppose that the surface of the virtual workpiece at the previous moment is represented by a set of mesh vertices, denoted as set. The newly acquired surface point cloud data is denoted as a set. For each vertex of the virtual workpiece surface mesh In the height contour point cloud Find the nearest point in the middle And update the new position of the vertex according to the formula. : ; Where λ is the smoothing factor, a constant between 0 and 1, used to control the smoothness of model updates; for If there is no corresponding old vertex in the mesh, a new mesh patch is created.

[0009] Further, in step S3, the online prediction model is used to calculate the probability of a quality anomaly occurring in the predicted area within a preset future time period based on the shape, temperature distribution, and currently used process parameters of the virtual workpiece, generating a risk map. This specifically includes: S3-1 Extracting geometric features from the current state Thermal characteristics and process characteristics The three elements are then transformed to the same scale and concatenated into a comprehensive feature vector F: ; in, , , The weight matrix is ​​a learnable matrix; the process features This includes current I, voltage U, and scanning speed v; S3-2. Input the comprehensive feature vector F into a pre-trained neural network and output the risk map R.

[0010] Furthermore, the extraction of process features from the current state includes: Process characteristics The arc voltage, current, and scanning speed parameters are initially fused using an energy density model E, the expression of which is: ; Where U is the arc voltage, I is the current, v is the scanning speed, and d is the arc diameter.

[0011] Furthermore, in step S4, the step of recalculating and optimizing the subsequent printing path and corresponding process parameters with the goal of avoiding the quality risk area specifically includes: S4-1. During the macro-planning phase, adjust the overall filling direction of subsequent layers and define a first objective function. The first objective function Used to balance the predicted total risk value and total path length of the adjusted path coverage area; a genetic algorithm is used to search for the filling direction that minimizes the first objective function; S4-2. Under the determined filling direction, perform a micro-adjustment stage, finely adjust the coordinates of each point on the path to be printed and their corresponding process parameters, and define a second objective function. The second objective function The predicted risk values ​​and process parameter offsets at each point on the balancing path are used; the gradient descent method is used to solve for the coordinates of each point and the process parameters that minimize the second objective function.

[0012] Furthermore, the first objective function The expression is: ; in, This is the total risk value of the path coverage area after adjustment, as shown in the risk map. It is the total length of the path. and These are the weighting coefficients for the two, respectively; Second objective function The expression is: ; in, It is the predicted risk value at point i on the path. It is the offset of the process characteristic parameter at that point from the standard value. The weighting factor for the offset.

[0013] Furthermore, the second objective function also includes a smoothing term S: , in, Let be the coordinate vector of the i-th point on the path; the second objective function after adding a smoothing term. for: ; in, The weighting coefficients for the smoothing term are used to constrain the coordinate changes between adjacent points on the path in order to maintain path smoothness.

[0014] Further, in step S5, verifying the optimized instructions includes: The optimized instructions are simulated and executed in a digital twin environment to check for collision interference. The online prediction model is used to assess the quality risk after executing the optimized instructions; Calculate a confidence score C, the expression for which is: ; The verification is considered successful when there is no collision interference, the confidence score C is greater than the preset threshold, and there is no collision alarm.

[0015] On the other hand, the present invention also provides an online control system for an electric arc additive manufacturing process based on digital twins, comprising: The physical printing unit includes the arc additive manufacturing equipment body and its motion and process controller; The sensing unit is configured to collect sensing data of the physical printing site in real time. The sensing data includes the height contour point cloud of the printed surface, the temperature distribution of the molten pool area, and the real-time position and process parameters of the print head. The digital twin and computing unit are configured as follows: The states of the virtual workpiece and virtual printhead are updated based on the perceived data; Using an online prediction model, based on the shape, temperature distribution, and process parameters currently in use of the virtual workpiece, the probability of quality anomalies occurring in the predicted area within a preset future time period is calculated, and a risk map is generated. When a quality risk area is determined to exist based on the risk map, the subsequent printing path and corresponding process parameters are recalculated and optimized with the goal of avoiding the quality risk area, and optimization instructions are generated. Before sending the optimization instructions to the physical printing unit for execution, a simulated test print is performed in the digital twin environment to verify the optimization instructions; The communication unit is configured to enable data exchange between the sensing unit, the digital twin and computing unit, and the physical printing unit.

[0016] Furthermore, the sensing unit includes a laser scanner for acquiring the height contour point cloud and an infrared thermal imager for acquiring the temperature distribution.

[0017] The technical solution of this invention, through the coordinated work of the aforementioned technical steps and system components, can systematically overcome the core defects of existing technologies, such as lagging regulation, isolated models, disconnect between path and parameter optimization, uneven virtual image updates, and lack of forward-looking verification. By constructing a complete cyber-physical loop of "physical perception, smooth virtual-real mapping, online quality risk prediction, two-stage collaborative optimization, forward-looking simulation verification, and closed-loop execution," this invention fundamentally elevates the process control mode of arc additive manufacturing from a passive, reactive corrective response to an active, proactive, and preventative intelligent control.

[0018] Specifically, the introduction of the online prediction model enables the system to anticipate the emergence of quality risks from real-time multidimensional state data seconds in advance, providing a valuable time window and decision-making basis for control actions and solving the problems of control lag and model isolation. The smooth update algorithm injects noise resistance and stability into the digital twin, ensuring that the input basis for prediction and optimization is continuous and reliable. The macro-micro two-stage optimization strategy achieves coordinated optimization of spatial paths and process parameters at both macro-layout and micro-detail scales, while also considering path smoothness, solving the problems of single and disconnected optimization dimensions. Finally, the simulation trial and confidence assessment mechanism built into the digital twin environment adds an indispensable layer of protection to the entire closed loop, ensuring the safety and effectiveness of the generated control commands and eliminating major hidden dangers in the operation of the automated system. Therefore, this invention significantly improves the consistency of forming quality, process stability, and the level of system autonomy and intelligence in the arc additive manufacturing process, reducing the reliance on manual intervention and experience in the manufacturing process. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the online control method of the present invention; Figure 2 This is a connection diagram of the online control system of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0021] Before implementing the control method of the present invention, the path and parameters are first initialized in a conventional manner, that is, the designed three-dimensional model is sliced, a preset printing path and process parameters are generated, and sent synchronously to the printer controller and the digital twin system; the digital twin system loads and displays the preset path in the virtual environment.

[0022] In the embodiments, such as Figure 1 As shown, the online control method for an electric arc additive manufacturing process based on digital twins according to the present invention specifically includes the following steps: Step S1 is real-time perception of multi-source data: During the printing process, perception data of the physical printing site is collected in real time, including: using a laser scanner to obtain the height contour point cloud of the printed surface, using an infrared camera to obtain temperature images of the molten pool area, and reading process parameters such as the real-time position of the print head, arc current, voltage, and moving speed from the controller; and sending these data to the digital twin system in real time.

[0023] Step S1 is fundamental to synchronizing the digital twin system with the physical site. The sensing data comprises three dimensions: morphological, thermal, and motion. Morphological data consists of the height contour point cloud of the printed surface. This point cloud data is typically acquired by structured light or laser scanning sensors at frequencies of tens of Hz, reflecting the actual three-dimensional morphology of the deposited weld bead with sub-millimeter precision, providing accurate surface geometry information for the digital twin. Thermal data represents the temperature distribution of the molten pool region. This temperature distribution can be acquired by an infrared thermal imager and presented as a two-dimensional image matrix, with each pixel value corresponding to the temperature value of a point in space, thus completely capturing the temperature field of the molten pool and heat-affected zone. Motion data includes the real-time position of the printhead and process parameters. The printhead position is typically the spatial coordinates and orientation of the robot end effector or welding torch in a Cartesian coordinate system. Process parameters include at least arc current, arc voltage, and scanning speed. This data is read directly from the controller or sensors of the physical printer in real time. By synchronously acquiring these three categories of data, a comprehensive, multi-dimensional physical world mirror can be provided for the digital twin.

[0024] Step S2 is to smoothly synchronize the virtual and real states: The digital twin system updates the state of the virtual printhead and the virtual workpiece according to the received data and the perceived data, so that the virtual scene is consistent with the physical scene.

[0025] The core of step S2 lies in achieving a smooth and stable virtual-real mapping mechanism. In specific embodiments, due to the unavoidable measurement noise and outliers in the sensor data, directly using the acquired raw point cloud to replace or simply interpolate and update the virtual workpiece surface mesh can easily lead to high-frequency fluctuations or local abrupt changes on the model surface. This unstable virtual performance will inject noise into the downstream prediction and decision-making modules, resulting in misjudgments and unnecessary control actions. To solve this problem, this invention employs an incremental update algorithm with a smoothing factor.

[0026] More preferably, the update of the virtual workpiece state in step S2 is achieved through a real-time modeling method that combines old and new data. Specifically, it is assumed that the surface of the virtual workpiece at the previous moment is represented by a set of mesh vertices, denoted as set. The newly acquired surface point cloud data is denoted as a set. .for Each vertex in ,exist Find the nearest point in the middle Calculate the new position of the vertex : ; in, It is a smoothing factor between 0 and 1, a preset constant used to control the speed of model updates. The larger the value, the smoother the model changes and the stronger its noise resistance.

[0027] When λ approaches 1, the model tends to maintain historical states, resulting in a smoother update process and stronger suppression of transient noise, but the response speed to physical changes slows down. When λ approaches 0, the model tends to quickly follow newly acquired data, resulting in a fast response speed, but the smoothness decreases. The specific value of λ can be adjusted according to sensor noise levels, printing speed, etc., and is preferably between 0.7 and 0.85 to achieve a balance between response speed and noise resistance.

[0028] for For regions where there are no corresponding old vertices, new mesh patches are created directly based on the point cloud data to incrementally expand the model. In this way, the virtual workpiece surface can effectively filter out high-frequency random noise while reflecting macroscopic morphological changes, ensuring the continuity and stability of model updates and laying a reliable foundation for subsequent accurate predictions. The virtual printhead's state update directly drives the corresponding model in the virtual scene to perform synchronous rigid body motion based on the collected real-time position and attitude data.

[0029] Step S3 is online quality risk prediction: The digital twin system has a built-in online prediction model. Using the online prediction model, based on the shape, temperature distribution, and process parameters of the current virtual workpiece, it calculates the probability of quality problems that may occur in the workpiece in the near future if printing continues in the current manner. The production risk map R is shown. The quality problems include local overheating, underheating, uneven temperature, or unstable molten pool. Problematic areas are marked.

[0030] Step S3 is the core of achieving the qualitative leap from post-event detection to pre-event prevention. Preferably, the online prediction model in step S3 is a fast computational model that transforms multiple types of input data into a quality prediction map, and its construction method is as follows: S3-1, Feature Extraction and Fusion: Extracting three types of features from the current state, including geometric features. Thermal characteristics Process characteristics Geometric features Information used to describe the local topography of deposited structures may include, but is not limited to, local height, surface normal, slope, curvature, etc., calculated from a digital mesh model. Thermal characteristics. Used to describe the current thermodynamic state of a workpiece, including but not limited to local average temperature, temperature gradient amplitude and direction, maximum temperature, cooling rate, etc., calculated from temperature distribution data. Process characteristics. Used to describe the current applied energy input state, which may include, but is not limited to, direct readings of arc current, arc voltage, scanning speed, or their derivatives.

[0031] After obtaining the three types of features, they need to be transformed to the same numerical scale, for example, through normalization or standardization, to eliminate numerical biases caused by different physical dimensions and value ranges. Then, the normalized feature vectors are concatenated along the feature dimensions to form a unified comprehensive feature vector. Therefore, geometric features... Thermal characteristics Process characteristics After being converted to the same scale, they are concatenated into a single comprehensive feature vector: ; in, , , The weight matrix is ​​a learnable matrix; the process features This includes current I, voltage U, and scanning speed v.

[0032] S3-2, Quality Prediction: Establish a neural network N, input the comprehensive feature vector F into the pre-trained neural network N, and output a two-dimensional risk map R.

[0033] The neural network model here can employ a fully convolutional network or an encoder-decoder structure with spatial attention. Its input is a comprehensive feature vector or a feature map mapped from the comprehensive feature vector, and its output is a two-dimensional risk map corresponding to the current spatial range of the workpiece surface. The value of each pixel or grid point in the risk map represents the probability or risk score of any specified type of quality anomaly occurring at that spatial location within a preset future time period, such as several seconds to tens of seconds in the future.

[0034] Quality anomalies may include, but are not limited to, local over-high, local under-high, collapse, warping, porosity, unstable molten pool, and lack of fusion. The training data for this neural network N comes from time-series data recorded in previous printing or specialized experiments, including the feature input at each time step and the actual quality labels at the corresponding positions verified by non-destructive testing or destructive testing. Through supervised learning, the network learns the ability to infer future quality risks from the current state.

[0035] As a preferred approach, to more efficiently and physically represent the energy input characteristics of process parameters, when extracting process features, instead of simply concatenating parameters such as voltage, current, and speed, a simplified energy density model is first used to initially fuse these parameters. Therefore, for process features... The arc voltage, current, and scanning speed parameters are initially fused using an energy density model E, the expression of which is: ; Where is the arc voltage, is the current, is the scanning speed, and is the arc diameter. The calculated energy density E, as an independent physical quantity containing rich process information, is used as an important component of the process feature vector, replacing or supplementing the original voltage, current, and speed values. This energy density-based feature construction method reduces the input dimensionality of the neural network, and the input features themselves have more explicit physical connotations, which helps to improve the model's prediction accuracy and generalization ability.

[0036] Step S4 is the stage dynamic path optimization: When a region where quality problems may occur is predicted according to the risk map, the system automatically starts the path optimization algorithm; the algorithm takes the actual state of the current virtual workpiece as the starting point, recalculates the subsequent printing path and process parameters, and generates optimization instructions, with the goal of avoiding the quality problems predicted in step S3; the new path should ensure printing quality as much as possible while keeping the path smooth.

[0037] As a preferred approach, step S4 employs a two-stage collaborative optimization strategy combining macro-planning and micro-adjustment to address the issues of disconnect between path planning and parameter adjustment, and the single optimization dimension prevalent in traditional methods. Specifically, the recalculation and optimization of subsequent printing paths and corresponding process parameters, with the goal of avoiding the quality risk area, includes: Step S4-1: Execute the macro-planning phase. The macro-planning phase focuses on the overall filling strategy for subsequent layers, such as layers 2 to 5. In this phase, the system maintains the basic scan path pattern within a single layer, but adjusts the overall filling direction angle of that layer. By enumerating or intelligently searching for different filling direction combinations, a first objective function is calculated for each candidate solution. For predicted risk areas, the approximate filling direction of subsequent layers is adjusted. Define the first objective function. To evaluate the advantages and disadvantages of different filling directions: ; in, This is the total risk value of the path coverage area after adjustment, as shown in the risk map. It is the total length of the path. and These are the weighting coefficients that balance the two objectives. The first objective function... To balance the predicted total risk value and total path length of the adjusted path coverage area, a genetic algorithm is used to search for... Minimum fill direction.

[0038] In the embodiment, the first objective function The objective function consists of two weighted components: the first is the predicted total risk value within the coverage area of ​​the adjusted new path, obtained by integrating or summing the risk probability values ​​within the corresponding path area in the risk map; the second is the total length of the new path. These two components are balanced using preset weighting coefficients. When the need for quality risk avoidance is higher, the weight of the first component can be increased; when the requirements for printing efficiency are more stringent, the weight of the second component can be increased. To efficiently find the optimal filling direction that minimizes the first objective function within a continuous and complex search space, a genetic algorithm is used in the macro-planning stage. The genetic algorithm encodes each filling direction combination as an individual, and through selection, crossover, and mutation operations, it evolves across generations, approaching the global optimum in fewer iterations, thus meeting the real-time requirements of online control.

[0039] Step S4-2: After determining the optimal layer filling direction through the macro-planning stage, the micro-adjustment stage begins. The micro-adjustment stage focuses on the next path segment that will be actually executed, typically a path segment from several seconds to tens of seconds in the future. In this stage, the spatial coordinates of each path point on the path segment and the corresponding process parameters, such as arc voltage, current, or scanning speed or their derivatives, are finely adjusted over a continuous range.

[0040] Specifically, the coordinates of each point on the next segment of the path to be printed and its corresponding process parameters are finely adjusted, and a more refined second objective function is defined. :

[0041] in, It is the predicted risk value at point i on the path. It is the deviation of the process parameter (such as energy density E) at that point from the standard value. These are weights. The second objective function. The offsets of predicted risk values ​​and process parameters at each point on the balancing path are used; the gradient descent method is used to solve for the second objective function. Minimize the coordinates and parameters of each point.

[0042] The second objective function also consists of two weighted terms: the first term is the predicted risk value at the i-th point on the path under the current adjustment scheme, which is obtained by re-inputting the new path into the online prediction model or by sampling and interpolating directly on the risk map; the second term is the offset between the adjusted process parameters at the i-th point and the preset standard process parameters, used to constrain the range of change in process parameters and prevent the introduction of new unstable factors due to drastic adjustments. The micro-adjustment stage employs gradient descent or a variant optimizer such as Adam to solve the problem. By calculating the gradient of the second objective function with respect to the coordinates of each path point and the process parameters, and iteratively updating these variables along the gradient descent direction, a set of path and parameter configurations that locally or globally minimizes the second objective function is finally found.

[0043] As a further preferred option, to maintain the kinematic smoothness of the optimized path and avoid drastic acceleration and deceleration during robot movement caused by sudden local jagged changes in the path, which could lead to secondary problems such as arc instability, wire feeding difficulties, or molten pool disturbances, a smoothing term S is introduced into the second objective function of the micro-adjustment stage:

[0044] in, Let be the coordinate vector of the i-th point on the path. The smoothing term is defined as the sum of the norms of the coordinate changes between all adjacent points on the path, such as the sum of the squares of the Euclidean distances between adjacent points or the second-order difference norm. This smoothing term is multiplied by a weighting coefficient and then added to the second objective function. By adjusting this weighting coefficient, a configurable balance can be achieved between the path's ability to accurately avoid risk areas and the smoothness of the path itself, ensuring that the optimization results meet both quality requirements and the motion control constraints of actual printing.

[0045] The objective function after adding the smoothing term is: ; in, These are the weighting coefficients for the smoothing term.

[0046] S5. Optimization Scheme Forward Verification and Execution: Before sending the new optimization instruction path to the physical printer, a simulated print test is performed in the digital twin system to verify the optimization instruction, check whether the new path will cause a collision, and evaluate the effect of the new path again using the predictive model.

[0047] The purpose of step S5 is to establish a safety valve to prevent automatically generated, unverified path instructions from being directly issued to expensive physical equipment, causing collisions or ineffective control. The verification process includes three parallel steps. The first step is geometric collision detection. In the digital twin environment, the virtual printhead is driven to simulate movement along the trajectory defined by the optimization instructions, and interference and collision checks are performed in real time between the virtual printhead model and the virtual workpiece model, virtual workbench model, and other peripheral equipment models. This step can be completed quickly using a mature physics engine or collision detection library, typically in the millisecond range. The second step is effect pre-evaluation. The virtual state after simulating the execution of the optimization instructions, or directly based on the coordinates of the area covered by the new path, is called again to calculate the expected quality risk map after implementing the new path. The third step is quantitative confidence assessment, calculating a confidence score C:

[0048] The maximum risk value predicted after optimization refers to the highest global risk probability value in the new risk map generated in the second step; the maximum risk value predicted before optimization refers to the highest global risk probability value in the original risk map that triggered this optimization. This confidence score directly reflects the degree to which the optimization operation eliminates or reduces the most serious quality risk. Theoretically, its value range can cover negative infinity to 1; a positive and larger score indicates a more significant optimization effect. The system presets a threshold, such as 0.2 or 0.3. Only when no collision interference is detected during the simulated printing process, the calculated confidence score C is greater than the preset threshold, and there is no collision alarm, is the optimization command considered verified. Only verified optimization commands are allowed to be sent to the physical printer for execution. This mechanism fundamentally prevents the accidental issuance of high-risk or invalid commands, ensuring the safety of the physical manufacturing process and the effectiveness of control actions.

[0049] In the display interface of the digital twin system, different colors are used to visually distinguish the paths: the preset path is blue, the actual mapped path is green, the predicted high-risk area is covered by red semi-transparent layer, and the optimized new path is a yellow flashing line.

[0050] Finally, step S6 is executed. After confirming that the verification is correct, the first part of the new path instruction is sent to the physical printer for execution. In this step, a closed loop is formed. After the physical printer executes the new instruction, it returns to step S2 and starts the next round of data collection, prediction and optimization, forming a continuous "monitoring-prediction-adjustment" closed loop, which is executed in a loop until printing is completed.

[0051] On the other hand, such as Figure 2 As shown, the present invention provides an online control system for implementing the above-described method in arc additive manufacturing, comprising: The physical printing unit 100 includes the arc additive manufacturing equipment body and its motion and process controller; The sensing unit 200 is configured to collect sensing data of the physical printing site in real time. The sensing data includes the height contour point cloud of the printed surface, the temperature distribution of the molten pool area, and the real-time position and process parameters of the print head. Preferably, it includes a laser scanner for three-dimensional topography scanning and acquiring the height contour point cloud, and an infrared thermal imager for temperature field monitoring and acquiring the temperature distribution. The digital twin and computing unit 300, composed of a high-performance computer, runs a digital twin virtual scene developed based on engines such as Unity, and integrates the online prediction model, path optimization algorithm, and database; it is configured as follows: The states of the virtual workpiece and virtual printhead are updated based on the perceived data; Using an online prediction model, based on the shape, temperature distribution, and process parameters currently in use of the virtual workpiece, the probability of quality anomalies occurring in the predicted area within a preset future time period is calculated, and a risk map is generated. When a quality risk area is determined to exist based on the risk map, the subsequent printing path and corresponding process parameters are recalculated and optimized with the goal of avoiding the quality risk area, and optimization instructions are generated. Before sending the optimization instructions to the physical printing unit for execution, a simulated test print is performed in the digital twin environment to verify the optimization instructions; The communication unit 400 is configured to enable full-duplex, real-time data exchange between the sensing unit, the digital twin and computing unit, and the physical printing unit.

[0052] Preferably, in the digital twin and computing unit, the online prediction model and path optimization algorithm are encapsulated as independent software service modules, which interact efficiently with the virtual scene engine through predefined API interfaces; the communication unit adopts a real-time communication protocol based on gigabit Ethernet to ensure that the end-to-end latency from data acquisition to instruction issuance is less than 50 milliseconds.

[0053] The sensing unit, acting as the system's sensor, is responsible for bridging the signal from the physical world to the information world. Specifically, it consists of a laser scanner for acquiring high-precision height contour point clouds and an infrared thermal imager for acquiring two-dimensional temperature distribution. Both are installed near the print head for follow-up measurements to ensure spatial coverage and synchronization of data acquisition.

[0054] The digital twin and computing unit is the central brain of the system, hosted by a high-performance computer. Running on this computer is a digital twin virtual scene developed using a virtual scene engine, such as Unity or Unreal Engine. This scene renders and displays virtual workpieces, virtual printheads, workbenches, and predictive risk maps with high fidelity. Core intelligent algorithms, including online prediction models and path optimization algorithms, are encapsulated as independent software service modules. Each module exposes predefined API interfaces, enabling efficient and loosely coupled data interaction with the virtual scene engine. This modular encapsulation architecture facilitates independent algorithm updates, elastic scaling of computing power, and rapid integration with different virtual scene engines. The communication unit is the system's neural network, responsible for connecting the three physical and computational entities.

[0055] To ensure the stringent real-time requirements of high-frequency sensing data uploads and control command issuance, the communication unit employs an industrial-grade real-time communication protocol based on Gigabit Ethernet and optimizes the network topology and data processing pipeline. This guarantees that the entire end-to-end latency, from data acquisition through feature extraction, prediction, optimization, verification, and finally the issuance of the optimized command to the printer controller, can be strictly controlled within 50 milliseconds. This consistent low-latency performance is a prerequisite for achieving real-time closed-loop control, enabling the system to proactively intervene in the rapidly evolving arc additive manufacturing process.

[0056] Example: Online Control Process of Arc Additive Manufacturing Based on Digital Twin This embodiment illustrates the complete implementation process of the method of the present invention. Taking the arc additive manufacturing of a large marine bulkhead component as an example, the specific steps are as follows: Step 1: First, import the 3D CAD model of the bulkhead component into the offline programming software. Set the slice layer thickness to 3mm according to the process requirements, and generate a preset printing path with a zigzag fill. The preset process parameters are: arc current 280A, voltage 32V, and print head movement speed 10mm / s. Convert the preset path into a robot-executable .src program file and send it simultaneously via Ethernet to the robot controller (KUKA KR 500 R2830) and the digital twin system deployed on the high-performance workstation. The digital twin system loads the preset path, which is clearly displayed as blue lines in the virtual scene.

[0057] Step Two (S1): Initiate Physical Printing. During the printing process, a laser line scanning sensor (Keyence LJ-V7080) installed at the robot's end effector scans the printed surface at a frequency of 50Hz to acquire a high-precision 3D point cloud. Simultaneously, an infrared thermal imager (FLIR A655sc) captures images of the molten pool and surrounding area at a frame rate of 60Hz to obtain temperature distribution images. The robot controller packages the real-time spatial coordinates (X, Y, Z) of the welding torch end effector, as well as the actual feedback values ​​of current and voltage, into data packets via a real-time communication interface. All of the above data is transmitted in real-time to the digital twin system via a gigabit industrial Ethernet switch with a latency of less than 30ms.

[0058] Step 3 (S2): After receiving the data, the digital twin system immediately drives the virtual scene update. The update of the virtual workpiece model adopts the incremental fusion algorithm of this invention: the surface mesh vertex set of the virtual workpiece in the previous frame is... With the new point cloud dataset Perform matching. Set the smoothing factor λ = 0.75, and for each pair of matching points, use the formula... = 0.75 + 0.25 The calculation of new vertex positions ensures smooth and stable model updates, effectively suppressing measurement noise. The virtual welding torch model moves synchronously based on the received real-time coordinate data. Users can observe a real-time comparison between the actual trajectory represented by the green line and the preset blue trajectory on the twin system interface.

[0059] Step 4 (S3): The online prediction model built into the digital twin system starts every 0.5 seconds. The model extracts features from the current state: geometric features (e.g., local height fluctuations in the latest layer exceed ±0.8mm), thermal features (e.g., temperature gradient in a certain area reaches 220°C / cm), and process features (current actual linear energy). E = (31.5V 275A) / (9.5mm / s 3.2mm) ≈ 298 J / mm). These features, after normalization and weighted fusion using pre-trained weight matrices, are input into a lightweight neural network. This network, trained with hundreds of sets of historical printing data, is capable of predicting quality risks within the next 3 seconds. At this point, the network outputs a risk map and marks a red semi-transparent area at the corner of the component, indicating a high risk of overheating and collapse due to heat accumulation (predicted probability > 70%).

[0060] Step 5 (S4): When a high-risk area is detected, the system automatically calls the two-stage path optimization algorithm.

[0061] Phase One (Macro Planning): The primary objective is to avoid high-risk areas around corners. Define the objective function. ,in α= 1.2 , β=0.8 A genetic algorithm was used to search for filling strategies for the subsequent five layers. After approximately 50 iterations, the algorithm determined a new additive path that effectively dispersed the heat input in the region, while increasing the total path length by only 5%. minimize.

[0062] Phase Two (Micro-adjustment): Under the established macro-strategy, fine-tune the path points to be printed in the next 10 seconds. Define the objective function. ,in γ =0.15, η =0.1. Using gradient descent, after 15 iterations, the optimization result is: shifting the 20 path points at the edge of the high-risk area outward by 0.3-0.5mm, increasing the moving speed of the corresponding points from 10mm / s to 12mm / s, and slightly reducing the current to 270A. The optimized new path is displayed as a flashing yellow line on the interface.

[0063] Step Six (S5): Before issuing the optimization command, the system performs a millisecond-level simulation test in the digital twin environment. The simulation results show that the new path has no collision interference, and after rerunning the prediction model, the risk value of the original high-risk area decreased by 45%. Calculate the confidence score. C = 1 - (55 / 100) = 0.45, which is higher than the preset threshold of 0.3. After the verification is successful, the system immediately sends the optimized first segment trajectory coordinates and process parameter instructions to the robot controller.

[0064] Step 7 (S6): The robot executes the new instruction, and the welding torch moves along the optimized yellow path. Simultaneously, the system immediately returns to Step 2 to begin collecting new point cloud, temperature, and position data, entering the next "monitor-predict-optimize" cycle. This closed-loop process continues until the entire component is printed.

[0065] This invention achieves autonomous decision-making and control based on digital twins in the field of arc additive manufacturing by constructing a complete closed loop of "physical perception-virtual mapping-online prediction-real-time optimization". The system can predict defects and risks such as molten pool instability, abnormal layer height, and excessive heat accumulation several seconds to tens of seconds in advance, and proactively adjust the path and process parameters, transforming post-event remediation into pre-event prevention, significantly improving the consistency and reliability of the forming process.

[0066] This invention utilizes a digital twin system to integrate information such as the geometric shape, temperature distribution, real-time trajectory, and predicted risks of a physical entity into a single virtual scene for high-fidelity rendering and dynamic display. Operators can remotely and intuitively monitor the entire printing process, clearly identify deviations and risk areas, and achieve a digital and immersive upgrade in monitoring methods, reducing the safety risks and reliance on experience associated with manual observation.

[0067] To address the challenge of balancing efficiency and accuracy in arc additive manufacturing path optimization, this invention proposes a two-stage algorithm: first optimizing the filling direction, then fine-tuning the points and process parameters. This strategy rapidly avoids macroscopic risk areas using a genetic algorithm, and then performs local fine-tuning of the path using gradient descent combined with smoothing constraints. This ensures both real-time computational efficiency and the smoothness and process stability of the optimized path, effectively preventing problems such as unstable wire feeding or arc drift caused by drastic path changes.

[0068] The online prediction model of this invention is continuously trained using historical data, and its prediction accuracy can be continuously improved over time. Simultaneously, the trajectory data, process parameters, prediction results, and final quality data throughout the entire process are all stored in a structured manner, forming valuable digital assets of the process. This not only supports the traceability analysis of defect causes but also provides data-driven optimal initial parameters for the process planning of subsequent parts, accelerating the process development cycle and demonstrating the long-term value of the system's self-evolution and knowledge accumulation.

[0069] The specific embodiments described in this invention are merely illustrative of the invention and are not intended to limit it. Those skilled in the art can make modifications to these embodiments without contributing any inventive step after reading this specification, but such modifications are protected by patent law as long as they fall within the scope of the claims of this invention.

Claims

1. A method for online control of an electric arc additive manufacturing process based on digital twins, characterized in that, Includes the following steps: S1. Real-time acquisition of sensing data at the physical printing site, including the height contour point cloud of the printed surface, the temperature distribution of the molten pool area, and the real-time position and process parameters of the print head. S2. Update the status of the virtual workpiece and virtual printhead in the digital twin system based on the perceived data; S3. Using an online prediction model, based on the shape, temperature distribution, and process parameters currently in use of the virtual workpiece, calculate the probability of quality anomalies occurring in the predicted area within a preset future time period and generate a risk map. S4. When a quality risk area is determined to exist based on the risk map, with the goal of avoiding the quality risk area, the subsequent printing path and corresponding process parameters are recalculated and optimized to generate optimization instructions. S5. Before sending the optimization instructions to the physical printer for execution, a simulated test print is performed in the digital twin system to verify the optimization instructions; S6. After confirming that the verification is successful, the optimization instruction is sent to the physical printer for execution; after the physical printer executes the new instruction, it returns to step S2 and repeats the process until printing is complete.

2. The online control method for arc additive manufacturing process based on digital twins according to claim 1, characterized in that, In step S2, updating the state of the virtual workpiece and virtual printhead in the digital twin system based on the perceived data specifically includes: Suppose that the surface of the virtual workpiece at the previous moment is represented by a set of mesh vertices, denoted as set. The newly acquired surface point cloud data is denoted as a set. For each vertex of the virtual workpiece surface mesh ; in the height contour point cloud Find the nearest point in the middle And update the new position of the vertex according to the formula. : ； Where λ is the smoothing factor, a constant between 0 and 1, used to control the smoothness of model updates; for If there is no corresponding old vertex in the mesh, a new mesh patch is created.

3. The online control method for arc additive manufacturing process based on digital twins according to claim 1, characterized in that, In step S3, the online prediction model is used to calculate the probability of a quality anomaly occurring in the predicted area within a preset future time period based on the shape, temperature distribution, and currently used process parameters of the virtual workpiece, generating a risk map. This specifically includes: S3-1 Extracting geometric features from the current state Thermal characteristics and process characteristics The three elements are then transformed to the same scale and concatenated into a comprehensive feature vector F: ； in, , , The weight matrix is ​​a learnable matrix; the process features This includes current I, voltage U, and scanning speed v; S3-2. Input the comprehensive feature vector F into a pre-trained neural network and output the risk map R.

4. The online control method for arc additive manufacturing process based on digital twin according to claim 3, characterized in that, The extraction of process features from the current state includes: Process characteristics The arc voltage, current, and scanning speed parameters are initially fused using an energy density model E, the expression of which is: ； Where U is the arc voltage, I is the current, v is the scanning speed, and d is the arc diameter.

5. The online control method for arc additive manufacturing process based on digital twin according to claim 1, characterized in that, In step S4, the step of recalculating and optimizing the subsequent printing path and corresponding process parameters with the goal of avoiding the quality risk area specifically includes: S4-1. During the macro-planning phase, adjust the overall filling direction of subsequent layers and define a first objective function. The first objective function Used to balance the predicted total risk value and total path length of the adjusted path coverage area; a genetic algorithm is used to search for the filling direction that minimizes the first objective function; S4-2. Under the determined filling direction, perform a micro-adjustment stage, finely adjust the coordinates of each point on the path to be printed and their corresponding process parameters, and define a second objective function. The second objective function The predicted risk values ​​and process parameter offsets at each point on the balancing path are used; the gradient descent method is used to solve for the coordinates of each point and the process parameters that minimize the second objective function.

6. The online control method for arc additive manufacturing process based on digital twin according to claim 5, characterized in that, The first objective function The expression is: ； in, This is the total risk value of the path coverage area after adjustment, as shown in the risk map. It is the total length of the path. and These are the weighting coefficients for the two, respectively; Second objective function The expression is: ； in, It is the predicted risk value at point i on the path. It is the offset of the process characteristic parameter at that point from the standard value. The weighting factor for the offset.

7. The online control method for arc additive manufacturing process based on digital twin according to claim 6, characterized in that, The second objective function also includes a smoothing term S: , in, Let be the coordinate vector of the i-th point on the path; the second objective function after adding a smoothing term. for: ； in, The weighting coefficients for the smoothing term are used to constrain the coordinate changes between adjacent points on the path in order to maintain path smoothness.

8. The online control method for arc additive manufacturing process based on digital twin according to claim 1, characterized in that, In step S5, verifying the optimized instructions includes: The optimized instructions are simulated and executed in a digital twin environment to check for collision interference. The online prediction model is used to assess the quality risk after executing the optimized instructions; Calculate a confidence score C, the expression for which is: ; The verification is considered successful when there is no collision interference, the confidence score C is greater than the preset threshold, and there is no collision alarm.

9. An online control system for an electric arc additive manufacturing process based on digital twins, characterized in that, include: The physical printing unit includes the arc additive manufacturing equipment body and its motion and process controller; The sensing unit is configured to collect sensing data of the physical printing site in real time. The sensing data includes the height contour point cloud of the printed surface, the temperature distribution of the molten pool area, and the real-time position and process parameters of the print head. The digital twin and computing unit are configured as follows: The states of the virtual workpiece and virtual printhead are updated based on the perceived data; Using an online prediction model, based on the shape, temperature distribution, and process parameters currently in use of the virtual workpiece, the probability of quality anomalies occurring in the predicted area within a preset future time period is calculated, and a risk map is generated. When a quality risk area is determined to exist based on the risk map, the subsequent printing path and corresponding process parameters are recalculated and optimized with the goal of avoiding the quality risk area, and optimization instructions are generated. Before sending the optimization instructions to the physical printing unit for execution, a simulated test print is performed in the digital twin environment to verify the optimization instructions; The communication unit is configured to enable data exchange between the sensing unit, the digital twin and computing unit, and the physical printing unit.

10. The online control system for arc additive manufacturing process based on digital twins according to claim 9, characterized in that, The sensing unit includes a laser scanner for acquiring the height contour point cloud and an infrared thermal imager for acquiring the temperature distribution.

Images