A laser cutting device for metal product manufacturing

Abstract

This invention discloses a laser cutting device for metal product manufacturing, comprising a multi-parameter sensing unit, a state estimation unit, a low-order physical model unit, a proxy prediction unit, an optimization control unit, and an execution and feedback unit. The optimization control unit adopts a hybrid control architecture that combines model predictive control with a long-term optimization module based on policy learning to achieve real-time coupled adjustment of multiple parameters such as laser power, cutting speed, auxiliary gas pressure, and focal position. Through deep integration of algorithms and parameters, this device achieves intelligent, precise, and stable control of the laser cutting process, and is suitable for automated processing and production of high-precision metal products.

Description

Technical Field

[0001] This invention belongs to the field of laser processing technology, specifically relating to a laser cutting device based on multi-source sensing and model prediction algorithms, which falls under the direction of intelligent manufacturing and precision cutting control technology. Background Technology

[0002] Laser cutting, as a high-energy-density thermal processing technology in metal product manufacturing, is widely used in industrial fields such as automobile manufacturing, shipbuilding, sheet metal processing, and electronic casings due to its high cutting speed, narrow kerf, high processing precision, and high degree of automation. The quality of laser cutting is affected by a combination of process parameters, with typical key parameters including laser power, cutting speed, focal point position (spot size and depth of focus), and type and pressure of assist gas. Simultaneously, the material's absorptivity, thermal conductivity, thickness, and surface condition (oxide layer, oil, roughness) also significantly influence the thermo-fluid-mechanical coupling behavior of the cutting process.

[0003] To improve cutting quality and suppress thermal deformation, academia and industry have proposed various control and optimization methods, including digital twin methods based on physical simulation, response surface methodology and heuristic optimization based on statistical experimental design, and data-driven machine learning or reinforcement learning methods.

[0004] A common approach is to establish a high-precision finite element thermo-mechanical coupling model or a complete digital twin system. This model predicts the temperature field and corresponding thermal deformation by simulating the laser heat source, heat conduction, convection, and radiation boundary conditions, and then formulates compensation strategies accordingly. This method offers good physical interpretation and can intuitively reflect the influence of different process variables on the thermal field and deformation, making it suitable for complex mechanism studies and offline process planning. However, detailed finite element simulations involve large computational loads, and mesh refinement and multi-physics coupling solutions are significantly time-consuming, making it difficult to meet the demands of real-time and low-latency control in industrial settings. Furthermore, high-precision simulations are highly dependent on boundary conditions, material thermophysical parameters, and the actual laser distribution. If actual operating conditions (such as fixture contact, material batches, optical path contamination, and nozzle wear) deviate from the expected parameters, the simulation results are prone to significant errors, affecting the compensation effect.

[0005] Another type of approach is based on statistical design and parameter optimization techniques such as Response Surface Methodology (RSM), orthogonal experiments, or Box-Behnken, combined with optimization algorithms such as Particle Swarm Optimization (PSO) and Genetic Algorithm (GA) to perform offline multi-objective optimization of several key process parameters. This type of approach has mature experimental design concepts, facilitating the obtaining of empirical mappings between parameters and responses, and can find the optimal process combinations for objectives such as roughness and dimensional errors within a certain operating range. However, response surface and heuristic optimization belong to static, offline optimization paradigms, relying on a large amount of experimental data, resulting in long experimental cycles, high material and energy consumption, and the established empirical models often fail to cover production line operating condition drift and material batch variations, lacking online adaptive capabilities and real-time closed-loop control capabilities.

[0006] In recent years, data-driven methods—including supervised regression models, deep neural networks, and reinforcement learning—have been increasingly used for laser cutting parameter prediction and control. Data-driven methods demonstrate advantages in modeling complex nonlinear coupling relationships and can improve adaptability to changes in the field through online learning; reinforcement learning methods show potential in policy learning and long-term performance optimization in continuous action spaces. However, data-driven methods generally face challenges such as large training data requirements, poor interpretability, insufficient robustness to out-of-distribution operating conditions, and difficulties in ensuring performance under safety constraints. Furthermore, strategies relying solely on data, in the absence of physical priors, are prone to generating unsafe control commands under extreme or unseen operating conditions.

[0007] In summary, existing technologies still face several prominent challenges in industrial laser cutting applications: while high-precision physical simulation (digital twins) can provide physically interpretable predictions, they suffer from high computational costs and synchronization lag, making them unsuitable for online short-term prediction and millisecond-level control; response surface methodology and offline empirical optimization methods rely on large amounts of experimental data and static models, resulting in poor adaptability to material batches, fixture variations, and equipment aging; pure learning methods lack uncertainty quantification and safety assurance mechanisms, posing potential risks in industrial settings; existing solutions often rely solely on temperature or visual information, lacking the fusion and utilization of multimodal signals such as spectral, slag morphology, and acoustic signals, leading to insufficient sensitivity to early cutting quality degradation; and while a few studies have achieved parameter optimization, they have not formed an engineering closed-loop system combining edge-level low-latency control, cloud-based strategy evolution, and local safety fallback, lacking deployability and robustness guarantees for production lines.

[0008] Based on the shortcomings of the existing technologies, there is an urgent need for a new technical solution that, while preserving physical interpretability, reduces the online computational burden, enhances the model's adaptability to real-world working conditions, quantifies uncertainty, and forms an engineered closed-loop control architecture. This solution would meet the comprehensive requirements of high precision, low energy consumption, real-time performance, and safety for laser cutting in metal product manufacturing environments. Summary of the Invention

[0009] To overcome the problems of poor real-time performance, weak generalization, large discrepancies between simulation and reality, lack of quantification of uncertainties, and imperfect closed-loop engineering in existing laser cutting control technologies for industrial applications, this invention aims to provide a laser cutting device for metal product manufacturing. This device combines a low-order physical model (ROM), an uncertainty surrogate model, model-based online optimization (MPC), and long-term policy learning (RL), supplemented by multi-source sensing and adaptive correction mechanisms, to achieve real-time coupled adjustment of multiple parameters such as laser power, cutting speed, focal point position, and auxiliary gas flow rate in the industrial setting. This significantly improves cutting accuracy, stability, and production efficiency, while also possessing good safety and engineering deployability.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A laser cutting device for producing metal products, characterized in that the device comprises:

[0012] A multi-parameter sensing unit is used to collect process and status parameters in real time; the process and status parameters include laser power, cutting speed, auxiliary gas pressure, focal point position, and local temperature field of the workpiece.

[0013] The state estimation unit is used to filter and fuse the data collected by the multi-parameter sensing unit and estimate the implicit working condition state.

[0014] Low-order physics model unit, used for short-time prediction based on a pre-established reduced-order physics model;

[0015] The proxy prediction unit is used to construct a proxy model based on the output of the low-order physical model and real-time measurement data, and output processing quality prediction and uncertainty estimation.

[0016] An optimization control unit is configured to run a constrained online optimization control algorithm and generate parameter adjustment instructions based on the low-order physical model prediction, the surrogate model's quality prediction, and uncertainty estimation; and

[0017] The execution and feedback unit is used to send the parameter adjustment command to the laser cutting actuator and send the cutting process feedback signal back to the state estimation unit and the proxy prediction unit to realize virtual and real closed-loop control.

[0018] Preferably, the optimization control unit adopts a hybrid control architecture that combines model predictive control with a long-term optimization module based on strategy learning to achieve real-time coupled adjustment of multiple parameters such as laser power, cutting speed, auxiliary gas pressure, and focal position.

[0019] Preferably, the multi-parameter sensing unit includes: a laser power sensor, a beam analyzer, an infrared / near-infrared temperature imager, a kerf profile camera, an auxiliary gas pressure sensor, and an online optical or acoustic sensor for detecting kerf roughness.

[0020] Preferably, the low-order physical model unit is constructed by extracting basis functions from offline finite element simulation data using generalized POD or SVD and then constructing a reduced-order dynamic system, the reduced-order state of which satisfies the formal expression:

[0021] Preferably, the proxy prediction unit adopts a Gaussian process regression model. The proxy prediction unit takes the output of the low-order physical model, real-time measurement data and historical processing data as input, and outputs the predicted value of the processing quality index and the corresponding confidence interval.

[0022] Preferably, the optimization control unit includes:

[0023] The short-term safety optimization module implements hard-constrained model predictive control (MPC) based on the uncertainty estimation of the low-order physical model and the surrogate model to generate safe execution instructions under device and physical constraints; and

[0024] The long-term strategy optimization module performs periodic or batch-level optimization of the strategy based on reinforcement learning (RL), and sends the reference strategy or strategy parameters generated by reinforcement learning to the short-term safety optimization module to improve process efficiency and quality.

[0025] Preferably, the long-term policy optimization module employs one or more algorithms among Deep Deterministic Policy Gradient (DDPG), Soft Behavior Policy (SAC), or Proximal Policy Optimization (PPO) to perform mixed training in a simulation environment and with a small number of real samples, and uses the policy parameters obtained from offline training to guide the objective function weights of the MPC.

[0026] Preferably, the device adopts an edge-cloud collaborative deployment architecture: the low-order physical model and MPC run in real time on the edge computing unit to meet the control cycle requirements; the long-term strategy optimization module (RL training), historical data management and large-scale retraining are performed in the cloud or on the factory server, and the strategy or model updates are periodically distributed to the edge computing unit.

[0027] Preferably, it also includes a visualization monitoring module, which is used to display the temperature field prediction cloud map, the surrogate model prediction confidence interval, the MPC optimization trajectory, the RL strategy iteration information and historical data traceability records in real time, and provides a manual intervention entry point and a model retraining trigger mechanism.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] Significantly improved real-time performance and meets industrial closed-loop control requirements: By compressing the key information of high-precision finite element analysis into a reduced-order model and running prediction and MPC solution in real time at the edge, parameter adjustment can be completed within a controllable low latency to meet the production line cycle time requirements.

[0030] Combining physical interpretability and data-driven adaptive capabilities: ROM retains physical priors and dynamic understanding, while surrogate models and RL provide data-driven performance enhancements and policy learning. The combination of the two can obtain more reliable and interpretable control behavior.

[0031] Uncertainty Quantification and Safety Assurance: The confidence interval provided by the proxy model is used to guide the conservative processing and proactive data collection of MPC, thereby prioritizing safety and quality under uncertain or abnormal operating conditions.

[0032] The project boasts strong deployability and scalability: the edge-cloud collaborative architecture supports iterative upgrades in the factory environment, facilitating migration across materials and machines, as well as long-term model maintenance.

[0033] Improved cutting quality and efficiency: By adjusting multiple parameters such as laser power, cutting speed, focus and gas flow rate in real time, thermal deformation can be effectively suppressed, kerf roughness and dimensional error can be reduced, while energy consumption can be optimized and the yield rate can be improved. Attached Figure Description

[0034] Figure 1 This is an implementation framework diagram of a laser cutting device for metal product manufacturing;

[0035] Figure 2 This is a micrograph of a cut surface processed by a laser cutting device used in the production of metal products. Detailed Implementation

[0036] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0037] The purpose of this invention is to provide a laser cutting device for metal product manufacturing. This device combines a low-order physical model (ROM), an uncertainty surrogate model, model-based online optimization (MPC), and long-term policy learning (RL), supplemented by multi-source sensing and adaptive correction mechanisms, to achieve real-time coupled adjustment of multiple parameters such as laser power, cutting speed, focal point position, and auxiliary gas flow rate in the industrial setting. This significantly improves cutting accuracy, stability, and production efficiency, while also possessing good safety and engineering deployability.

[0038] To achieve the above objectives, see Appendix Figure 1 A schematic diagram of an implementation framework for a laser cutting device for metal product manufacturing. The present invention adopts the following technical solution:

[0039] A laser cutting device for producing metal products, characterized in that the device comprises:

[0040] A multi-parameter sensing unit is used to collect process and status parameters in real time; the process and status parameters include laser power, cutting speed, auxiliary gas pressure, focal point position, and local temperature field of the workpiece.

[0041] The state estimation unit is used to filter and fuse the data collected by the multi-parameter sensing unit and estimate the implicit working condition state.

[0042] Low-order physics model unit, used for short-time prediction based on a pre-established reduced-order physics model;

[0043] The proxy prediction unit is used to construct a proxy model based on the output of the low-order physical model and real-time measurement data, and output processing quality prediction and uncertainty estimation.

[0044] An optimization control unit is configured to run a constrained online optimization control algorithm and generate parameter adjustment instructions based on the low-order physical model prediction, the surrogate model's quality prediction, and uncertainty estimation; and

[0045] The execution and feedback unit is used to send the parameter adjustment command to the laser cutting actuator and send the cutting process feedback signal back to the state estimation unit and the proxy prediction unit to realize virtual and real closed-loop control.

[0046] Specifically, the multi-parameter sensing unit provided by this invention constitutes the core sensing layer of the intelligent laser processing system, aiming to achieve multi-dimensional real-time monitoring and dynamic feature extraction of the laser cutting process. This unit integrates multiple types of sensors and signal fusion algorithms to achieve high-precision acquisition, dynamic calibration, and timing synchronization of key process and state parameters, thereby providing data support for subsequent power control, adaptive focus adjustment, and processing quality prediction.

[0047] The multi-parameter sensing unit is used to collect process and status parameters in real time. The parameters include at least: laser power, cutting speed, auxiliary gas pressure, focal position, and local temperature field of the workpiece.

[0048] Preferably, the multi-parameter sensing unit includes: a laser power sensor, a beam analyzer, an infrared / near-infrared temperature imager, a kerf profile camera, an auxiliary gas pressure sensor, and an online optical or acoustic sensor for detecting kerf roughness.

[0049] Specifically, the multi-parameter sensing unit includes the following modules:

[0050] A laser power sensor, positioned in the laser output path, is used to detect the laser's output power in real time. It employs a high-frequency response sensor based on a thermopile or photodiode array, achieving a measurement accuracy better than ±0.5%. The detected data is used for real-time feedback control to stabilize the laser output, and combined with a control algorithm to dynamically compensate for power fluctuations.

[0051] A beam analyzer is used to monitor the spatial energy distribution, mode quality (M² value), and focus position drift of a laser beam. Beam cross-section reconstruction is achieved through beam splitting sampling and high-resolution CCD imaging to evaluate beam uniformity and focusing status. The data results are linked with the focus adjustment mechanism to dynamically correct focus shift and machining depth errors.

[0052] An infrared / near-infrared temperature imager, positioned at the side or top of the cutting area, is used to acquire the temperature field distribution on the workpiece surface and in the cut area. Dual-band (3–5μm and 8–12μm) imaging technology is employed to reduce the influence of emissivity, enabling real-time calibration and dynamic thermal image reconstruction of the temperature field. By analyzing the temperature gradient and molten pool boundary characteristics, the uniformity of heat input and the risk of overheating during the cutting process can be determined.

[0053] The kerf contour camera, employing a combination of a high-frame-rate industrial camera and structured light or stripe illumination, is used for real-time detection of kerf shape and width variations. The image processing module can extract kerf geometric feature parameters, such as kerf width, taper, and edge roughness indices, supporting online quality assessment.

[0054] An auxiliary gas pressure sensor, installed in the gas supply path, detects the instantaneous pressure and flow rate of the auxiliary gas (such as O2, N2, or Ar). A high-response pressure sensor and a digital signal output module are used to monitor nozzle status, gas stability, and abnormal blockages. Data is analyzed in conjunction with kerf temperature changes to determine the efficiency of gas injection and slag removal.

[0055] Online optical / acoustic roughness detection sensors utilize optical methods such as online detection based on scattered light intensity distribution or laser confocal principles, while acoustic methods assess surface roughness by utilizing changes in acoustic signals generated during the cutting process. The detection results, combined with machine learning models, enable intelligent judgment and trend prediction of cutting quality.

[0056] The state estimation unit is connected to the aforementioned multi-parameter sensing unit and is used to filter, fuse, and infer implicit states from the collected multi-source data. The design goal of this unit is to acquire key implicit operating condition variables that cannot be directly measured but are crucial to process stability and processing quality in real time through data layer modeling and time-series dynamic estimation, thereby achieving an intelligent transition from "signal perception" to "state cognition".

[0057] Specifically, the state estimation unit receives raw signal data from the multi-parameter sensing unit, including laser power, cutting speed, auxiliary gas pressure, focal point position, and local temperature field of the workpiece. Since these signals commonly suffer from random noise interference, sampling lag, and nonlinear coupling during actual processing, data from a single sensor is insufficient to accurately reflect the real operating conditions.

[0058] Therefore, this invention employs the Kalman Filter (KF) algorithm to construct a state estimation framework. The system first establishes a state-space model of the laser processing process:

[0059] Equations of state describe the dynamic evolution of process parameters (such as focus drift and temperature distribution);

[0060] The observation equation corresponds to the measurement relationship of the multi-sensor output.

[0061] Within each time step, the state estimation unit is based on a prediction-correction mechanism:

[0062] Prediction phase: Based on the previous estimate and the system model, predict the current operating condition.

[0063] Correction phase: Using real-time measurement data from the sensing unit, the predicted value is corrected to suppress noise interference and minimize the estimation error covariance.

[0064] This unit employs a hierarchical fusion strategy tailored to different signal characteristics and temporal resolutions:

[0065] Low-level fusion (physical quantity level) aligns the data from temperature imagers, pressure sensors, and power sensors in time and models the noise, using adaptive Kalman filtering to eliminate high-frequency interference. The output is a smoothed version of the underlying physical quantities (such as average power, temperature baseline, and steady-state pressure).

[0066] Mid-level fusion (dynamic feature level) jointly analyzes the infrared temperature field sequence and the slit contour camera image to extract the temperature gradient change rate and the slit deformation trend. Using a KF state extension model, these change rates are input as dynamic system variables, enhancing the ability to characterize nonlinear coupling relationships.

[0067] High-level fusion (implicit state level): After integrating the above information, the filter outputs implicit operating conditions, including: local average temperature (Ta): reflecting the overall heating level of the workpiece and used to determine the uniformity of heat input; temperature gradient (∇T): used to analyze the local stress distribution and the range of the heat-affected zone; focus deviation estimation (Δf): inverting the focus drift amount through beam analyzer data and kerf geometry parameters; real-time kerf width estimation (We): inferring the real-time kerf width change by fusing image features and acoustic signals.

[0068] This fusion process not only achieves spatiotemporal consistency of data from different modalities (optical, thermal, and mechanical), but also enables the system to maintain continuous and reliable state estimation through model prediction even when some sensor signals are lost or noise is too high.

[0069] The low-order physical model unit is connected to the state estimation unit and the optimization control unit. It is used to make short-time predictions of the thermal behavior of the laser cutting process based on the pre-established reduced-order model (ROM), providing physical prior support for real-time compensation and control strategies.

[0070] During laser cutting, the localized heating, melting, and cooling of the material are accompanied by complex heat conduction, convection, and phase transition behaviors, and its transient temperature field evolution exhibits strong nonlinearity and multi-temporal scale characteristics. Although the traditional finite element method (FEM) can accurately describe these processes, directly calling the complete finite element model in the real-time control system leads to a huge computational load and excessive time delay, making it unsuitable for online compensation control.

[0071] To this end, the present invention establishes a physical benchmark dataset based on a high-precision finite element model in the offline stage, and extracts the main dynamic characteristics of the system through a reduced-order modeling method, thereby forming a low-order physical model unit that combines computational efficiency and physical interpretability.

[0072] Preferably, the low-order physical model unit is constructed by extracting basis functions from offline finite element simulation data using generalized POD or SVD and then constructing a reduced-order dynamic system, the reduced-order state of which satisfies the formal expression:

[0073] Specifically, the establishment of a reduced-order physical model mainly includes the following three steps:

[0074] S1: Offline finite element simulation and sample generation. First, under typical processing parameters (laser power, cutting speed, gas pressure, focal position, etc.), multiple sets of finite element thermo-mechanical coupling simulations are performed to obtain multi-dimensional time-series temperature field and deformation response data.

[0075] S2: Generalized POD / SVD reduction processing. The simulation data matrix above is processed using generalized orthogonal decomposition (POD) or singular value decomposition (SVD) to extract basis functions. Modes with a major energy percentage exceeding 95% are selected as the reduction basis. These basis functions characterize the main spatial distribution features of heat conduction and thermal stress changes, forming the basis of the reduced state space.

[0076] S3: Construction of the reduced-order dynamic system. In the reduced-order space, the original high-dimensional partial differential equation system is transformed into a low-dimensional ordinary differential equation system through pseudo-Galerkin projection, forming the reduced-order dynamic equations:

[0077]

[0078] Through this formal expression, the complex high-dimensional heat conduction problem is mapped to a low-dimensional dynamic system that can be solved by real-time integration.

[0079] Short-term prediction and online invocation mechanism

[0080] The low-order physical model unit forms a closed-loop coupling relationship with the state estimation unit during system operation: the state estimation unit provides the current implicit operating conditions (such as local average temperature, focal deviation, etc.); the low-order physical model uses this state as initial conditions, combined with the current process inputs. It can quickly predict the temperature field and deformation trend in the short future (e.g., 0.1–0.5 seconds); the prediction results are input into the optimization control unit in real time to correct the cutting speed, power or focus position in advance, realizing a collaborative control mode of "feedforward prediction + closed-loop correction". In terms of hardware implementation, this process can run in the edge computing module, with a typical single prediction latency of less than 5ms, which can meet the real-time requirements of high-speed laser cutting.

[0081] During laser cutting, although the reduced-order physical model can accurately predict the thermal field evolution in a short time, its prediction accuracy is still affected by non-modeling factors such as material inhomogeneity, changes in surface reflectivity, and airflow disturbances. To further improve the prediction capability of actual processing quality (such as surface roughness, kerf width, and heat-affected zone size), this invention introduces a surrogate prediction unit on top of the physical model to model and compensate for the deviation between physical prediction and actual observation, and uses probabilistic regression to achieve prediction confidence assessment.

[0082] The proxy prediction unit is used to construct a proxy model based on the output of the low-order physical model and real-time measurement data, and output the processing quality prediction and uncertainty estimate.

[0083] Preferably, the proxy prediction unit adopts a Gaussian process regression model. The proxy prediction unit takes the output of the low-order physical model, real-time measurement data and historical processing data as input, and outputs the predicted value of the processing quality index and the corresponding confidence interval.

[0084] Specifically, the inputs to the proxy prediction unit include: prediction outputs from the low-order physical model unit (such as reduced-order state variables like local temperature field, temperature gradient, and stress distribution); real-time measurement data from the multi-parameter sensing unit and state estimation unit (such as laser power, focus deviation, local temperature, kerf width, and optical / acoustic roughness signals); and process parameter-quality mapping samples from the historical processing database, used for continuous model updates and confidence interval calibration.

[0085] The output of the proxy prediction unit includes: the predicted values ​​and confidence intervals of the processing quality indicators output by the proxy prediction unit, mainly including: the predicted surface roughness (Ra, Rz); the geometric error of the kerf (width, taper); the width or deformation of the heat-affected zone (HAZ); and the corresponding confidence intervals or variance estimates, which are used to reflect the uncertainty estimation of the model.

[0086] The proxy prediction unit uses a Gaussian process regression (GPR) model, which has excellent nonparametric regression capabilities with low sample data and can naturally provide the predicted mean and variance.

[0087] The GPR model uses a kernel function to measure the correlation between input samples and achieves physically consistent quality prediction by jointly modeling the output of the low-order physical model and the measurement signal.

[0088]

[0089] The optimization control unit is used to run a constrained online optimization control algorithm and generate parameter adjustment instructions based on the predictions of the low-order physical model, the quality predictions of the surrogate model, and the uncertainty estimation.

[0090] The optimization control unit is connected to the low-order physical model unit, the surrogate prediction unit, and the multi-parameter sensing unit. It is used to integrate the real-time status, short-term prediction, and quality prediction results of the processing process, and run a constrained online optimization algorithm to dynamically adjust key parameters such as laser power, cutting speed, auxiliary gas pressure, and focal position, thereby achieving the optimal balance between processing quality and energy efficiency.

[0091] The core objective of optimizing the control unit is to minimize the surface roughness, kerf size error, and heat-affected zone while ensuring equipment safety and cutting stability, and to maximize energy utilization.

[0092] Preferably, the optimization control unit adopts a hybrid control architecture that combines model predictive control with a long-term optimization module based on strategy learning to achieve real-time coupled adjustment of multiple parameters such as laser power, cutting speed, auxiliary gas pressure, and focal position.

[0093] This unit integrates the interpretable predictive capabilities of low-order physical models and the uncertainty feedback capabilities of surrogate predictive models to construct an "intelligent self-optimizing control system under safety constraints," realizing a collaborative control mechanism from model-driven predictive control (MPC) to data-driven reinforcement learning (RL) policy optimization.

[0094] Preferably, the optimization control unit includes the following two functional modules:

[0095] The short-term safety optimization module implements hard-constrained model predictive control (MPC) based on the uncertainty estimation of the low-order physical model and the surrogate model to generate safe execution instructions under device and physical constraints; and

[0096] The long-term strategy optimization module performs periodic or batch-level optimization of the strategy based on reinforcement learning (RL), and sends the reference strategy or strategy parameters generated by reinforcement learning to the short-term safety optimization module to improve process efficiency and quality.

[0097] Specifically, the short-term safety optimization module constructs an objective function based on the principle of Model Predictive Control (MPC) and uses the processing quality prediction output by the surrogate prediction unit as the optimization index.

[0098]

[0099] Specifically, the above optimization problem is solved under the following constraints:

[0100] Equipment physical constraints: Control variables such as laser power, motion speed, focal position, and gas pressure must not exceed the allowable range of the equipment;

[0101] Process safety constraints: Local temperature gradients, heat accumulation rates, and kerf widths must be kept within process safety thresholds to prevent material collapse or gas disturbance instability;

[0102] Model confidence interval constraint: When there is uncertainty in the quality prediction result output by the surrogate prediction unit, its upper limit of confidence interval is used as a safety boundary to ensure that the optimization solution does not fall into the high-risk region.

[0103] During operation, the MPC module uses joint predictions from the low-order physics model and the surrogate model to continuously update the optimization window in real time. Within each cycle, only the first control variable... The first part is implemented, and the rest is retained for recalculation in the next cycle, thus achieving constrained dynamic optimal control. Through the above design, the short-term safety optimization module can take into account processing quality, dimensional accuracy and thermal stability during laser cutting, avoiding the lag and uncertainty problems of traditional empirical parameter adjustment. At the same time, it has adaptive response capability and strong safety robustness, providing a stable and controllable execution foundation for the long-term strategy optimization module (policy update based on reinforcement learning).

[0104] Preferably, the long-term policy optimization module employs one or more algorithms among Deep Deterministic Policy Gradient (DDPG), Soft Behavior Policy (SAC), or Proximal Policy Optimization (PPO) to perform mixed training in a simulation environment and with a small number of real samples, and uses the policy parameters obtained from offline training to guide the objective function weights of the MPC.

[0105] Specifically, the long-term strategy optimization module is the high-level strategy decision-making part of the optimization control unit. Under the premise of ensuring system safety constraints, it dynamically adjusts the parameter configuration and objective function weight of the short-term safety optimization module by learning the long-term process evolution law and energy efficiency-quality trade-off relationship, thereby achieving global performance optimization and continuous adaptive optimization.

[0106] The long-term strategy optimization module adopts a reinforcement learning (RL) framework, using the cumulative quality index and energy consumption index of the processing system as the reward function. Through interaction with the simulation environment or the real system, it gradually learns the optimal process parameter adjustment strategy.

[0107] Specifically, this module can employ one or more deep reinforcement learning algorithms, such as Deep Deterministic Policy Gradient (DDPG), Soft Behavior Policy (SAC), or Proximal Policy Optimization (PPO), to achieve policy learning according to different application scenarios:

[0108] DDPG (Deep Deterministic Policy Gradient) is suitable for precise parameter adjustment in continuous control space, such as continuous variation control of laser power and focus position;

[0109] SAC (Soft Actor-Critic) introduces an entropy regularization term, which can improve exploration capabilities and stability, making it suitable for handling cutting processes with large operating disturbances or strong measurement noise.

[0110] PPO (Proximal Policy Optimization) maintains the stability of policy updates by pruning the objective function, making it suitable for periodic batch training and policy transfer optimization.

[0111] In terms of training methods, the long-term strategy optimization module adopts a hybrid training mechanism of simulation and field testing:

[0112] First, offline reinforcement learning training is performed using a high-fidelity simulation environment built on a low-order physical model (ROM) and a proxy prediction unit to generate an initial policy or policy parameters.

[0113] Subsequently, the strategy was fine-tuned or verified on actual equipment with a very small number of real processing samples to ensure the feasibility and safety of the strategy under real physical constraints.

[0114] After training is complete, the long-term policy optimization module uses the policy or policy parameters obtained from offline training as prior knowledge to dynamically guide the operation of the short-term security optimization module, specifically including:

[0115] Adaptive adjustment of objective function weights: Based on the system state value assessment output by the reinforcement learning policy, the weight coefficients in the MPC objective function are adjusted. Periodic recalibration is performed to achieve the optimal global balance between processing efficiency, surface quality, and thermal stability;

[0116] Reference trajectory generation and update: Based on the optimal state trajectory predicted by the reinforcement learning policy (such as temperature distribution and kerf width evolution path), a reference trajectory or boundary condition is provided for the short-term MPC module to enhance the foresight and stability of the rolling optimization.

[0117] Through the above design, the long-term strategy optimization module provides strategy guidance and weight adjustment for short-term MPC on a macro time scale, enabling the system to have a dual-layer intelligent control feature of "short-term precise control + long-term self-learning optimization", thereby maintaining high consistency in cutting quality and energy efficiency under different materials, thicknesses and environmental conditions.

[0118] In addition, the device further includes an adaptive correction module during operation, which is used to dynamically update and calibrate the low-order physical model or surrogate prediction unit based on small batches of real processing data collected online, so as to continuously reduce the "Sim-to-Real" error between the simulation environment and the actual processing process.

[0119] The adaptive correction module includes adaptive updates of the basis functions of the low-order physics model (ROM) and online correction of the surrogate prediction unit.

[0120] Specifically, the adaptive update of the basis functions of the low-order physical model (ROM) is implemented as follows: During the operation of the device, the system periodically collects a small number of real processing samples (including temperature field distribution, kerf width variation, surface roughness and dimensional deviation, etc.) and compares and analyzes them with the simulation prediction results of the low-order physical model; when the model output error or prediction residual exceeds a preset threshold, an incremental property orthogonal decomposition (POD) or dynamic basis function replacement algorithm is triggered to locally update the basis function matrix of the ROM without complete retraining; the updated basis functions are more in line with the heat conduction and material response characteristics under the current process conditions, thereby improving short-term prediction accuracy and model stability.

[0121] Specifically, the online correction implementation method of the proxy prediction unit is as follows:

[0122] For the surrogate prediction unit (Gaussian process regression GPR model), the system uses the latest real processed data to perform online Bayesian updates or kernel function reparameterization correction; for the GPR model, the incremental Gaussian process regression (Online GPR) method can be used to correct the mean and covariance structure of the model in real time by incrementally updating the kernel matrix.

[0123] Through the aforementioned adaptive correction mechanism, the system can achieve continuous self-learning and accuracy maintenance of the model during long-term operation without relying on large-scale retraining or manual calibration. This module effectively eliminates model drift caused by environmental changes, material batch differences, or equipment aging, significantly improving the robustness and generalization of the laser cutting device under complex working conditions.

[0124] The execution and feedback unit is used to send the parameter adjustment command to the laser cutting actuator and send the cutting process feedback signal back to the state estimation unit and the proxy prediction unit to realize virtual and real closed-loop control.

[0125] Specifically, the execution and feedback unit includes a servo drive module, a laser power control module, a gas flow regulating valve, a focal length servo controller, and a motion control system. The multi-dimensional parameter adjustment commands output by the optimization control unit, such as laser power, cutting speed, auxiliary gas pressure, and focal position, are coordinated by this unit and then applied to the corresponding physical actuators. This achieves synchronous adjustment of the laser emitter, CNC worktable, and gas jet system, thereby completing real-time adaptive control of the processing process.

[0126] During execution, the execution and feedback unit continuously collects real-time feedback signals from the sensor array, including laser output power, air pressure changes, motion speed, temperature field distribution, and visual images of the kerf, and transmits these signals back to the state estimation unit and the surrogate prediction unit. The state estimation unit updates implicit state variables (such as local average temperature, focus deviation, and instantaneous kerf width) accordingly, while the surrogate prediction unit corrects processing quality predictions and uncertainty estimates based on the latest measurement data, thus forming a virtual-real closed-loop control chain that runs through "perception—estimation—modeling—optimization—execution—feedback".

[0127] Furthermore, to ensure the system's operational safety and robustness, the execution and feedback unit is equipped with a safety rollback strategy: when the uncertainty or confidence interval of the output of the proxy prediction unit or state estimation unit exceeds a preset threshold, the system automatically triggers the safety rollback mechanism. This mechanism includes two modes:

[0128] (1) Conservative parameter mode: The execution and feedback unit automatically switches to a set of preset safe process parameters (such as reducing laser power, reducing cutting speed, and increasing gas flow) to avoid workpiece overheating, cutting interruption or equipment damage;

[0129] (2) Safety shutdown alarm mode: When the uncertainty exceeds the second threshold or abnormal fluctuations are detected, the system immediately stops automatic control and issues an alarm signal, suspends the cutting operation until manual confirmation or model recalibration is completed.

[0130] Through the above design, the execution and feedback unit not only achieves high-precision physical execution of optimized control results, but also ensures the stability and safety of the entire system under uncertain environments, forming an intelligent laser cutting control closed loop that can learn, correct itself, and has anomaly defense capabilities.

[0131] Preferably, the device adopts an edge-cloud collaborative deployment architecture: the low-order physical model and MPC run in real time on the edge computing unit to meet the control cycle requirements; the long-term strategy optimization module (RL training), historical data management and large-scale retraining are performed in the cloud or on the factory server, and the strategy or model updates are periodically distributed to the edge computing unit.

[0132] The device adopts an edge-cloud collaborative deployment architecture to balance the real-time requirements of the laser cutting process with the high-complexity training needs of the algorithm model, thereby improving the overall system's computational efficiency and scalability while ensuring control accuracy and safety.

[0133] Specifically, this collaborative architecture comprises a two-tier structure: edge computing units located on the production floor and a cloud computing platform located on remote servers or in the factory data center.

[0134] Edge Computing Unit: The edge side is primarily responsible for executing time-sensitive real-time control tasks and fast-response logic. Both the low-order physical model (ROM) and model predictive control (MPC) algorithms run in real-time on the edge side. The ROM can make short-term predictions of key states such as temperature field and thermal deformation within millisecond time steps, providing immediate physical constraint support for MPC. The MPC controller, based on feedback from the proxy prediction unit and real-time data from the state estimation unit, generates adjustment commands for laser power, cutting speed, auxiliary gas pressure, and focal point position within control cycles of tens of milliseconds, ensuring high dynamic response and high-precision control of the processing. The edge side also includes a data caching and anomaly detection module to maintain safe operation in the event of network latency or communication interruption, and to switch to conservative process parameters when necessary.

[0135] Cloud computing platform: The cloud or factory server side undertakes high-computing, non-real-time tasks, including reinforcement learning (RL) policy training, historical data management, and model retraining. The cloud collects historical processing data, sensor records, and MPC execution logs from multiple production batches and multiple devices to build a large-scale experience database. Based on this database, the cloud periodically runs reinforcement learning algorithms (such as DDPG, PPO, or SAC) to train or retrain the long-term policy optimization module to improve the policy's generalization ability under different materials, thicknesses, and environmental conditions. Simultaneously, the cloud can recalibrate and reconstruct parameters for low-order physical models and surrogate prediction models, for example, through large-scale finite element simulations or incremental GPR updates, to correct model drift or improve prediction accuracy.

[0136] After training is completed, the policy parameters, model weights, or MPC objective function weight sets generated in the cloud will be distributed to the edge computing unit in a versioned manner and will be smoothly updated after the system is under low load or after manual confirmation.

[0137] The edge-cloud collaborative mechanism is as follows: the cloud is responsible for "learning" and "optimization," while the edge is responsible for "execution" and "response." The cloud continuously optimizes the decision-making capabilities of the edge controller through periodic updates and differential synchronization, achieving a closed-loop learning system that learns from historical experience and dynamically adapts to new operating conditions. Real-time operational data generated by the edge during execution (including model prediction errors, uncertainty metrics, control deviations, etc.) is transmitted back to the cloud as feedback input for the next round of policy updates and model retraining, realizing cloud-edge collaborative evolution.

[0138] In summary, this edge-cloud collaborative architecture achieves a harmonious balance between real-time control and algorithmic intelligence by moving real-time decision-making to the edge and centralizing policy learning and model updates in the cloud. This design not only significantly improves the system's response speed and robustness but also enables the device to possess self-learning, self-optimization, and cross-condition migration capabilities, making it particularly suitable for laser cutting scenarios involving multiple materials, multiple batches, and frequent changes in metal products. See the microscopic image of the cut surface processed by the laser cutting device for metal product manufacturing of this invention. Figure 2 As can be seen, the cut is smooth and free of slag, indicating a good cutting effect.

[0139] Preferably, it also includes a visualization monitoring module, which is used to display the temperature field prediction cloud map, the surrogate model prediction confidence interval, the MPC optimization trajectory, the RL strategy iteration information and historical data traceability records in real time, and provides a manual intervention entry point and a model retraining trigger mechanism.

[0140] The device also includes a visualization monitoring module, used to enhance the observability of the entire laser cutting process, make the decision-making process transparent, and enable human-machine collaborative intervention. This module constitutes the human-machine interaction and state interpretability layer of the system. By visually presenting the operating status of the core model and control unit, operators can monitor the dynamic characteristics of the processing process, the confidence level of model predictions, and the evolution trend of control decisions in real time.

[0141] Specifically, the visualization monitoring module includes the following functional components and working mechanisms:

[0142] 1. Temperature field prediction contour map display:

[0143] This module connects to the low-order physical model unit and state estimation unit data interface to render the temperature field distribution and thermal gradient changes of the workpiece in real time. The displayed content includes: the instantaneous temperature distribution of the processing area, the heat diffusion trajectory, and the identification of local high-temperature areas; it supports switching between two-dimensional cross-sections and three-dimensional cloud maps, with temperature values ​​displayed in pseudo-color mapping; operators can interactively zoom in on local areas or track the temperature change curves of key points through the interface. This function helps operators intuitively understand the spatial distribution of heat input and thermal deformation, facilitating the assessment of potential overheating or heat accumulation risks.

[0144] 2. Display of Confidence Intervals for Surrogate Model Prediction

[0145] Connected to the data output of the proxy prediction unit, this system displays the predicted values ​​and uncertainty ranges of various quality indicators (such as surface roughness, dimensional error, and thermal deformation) in real time. While displaying the prediction results, the system presents the confidence interval as a shaded band or error bar; when the confidence interval width exceeds a preset threshold, the interface automatically highlights it in red and triggers a prompt. Operators can view the sources of uncertainty (data sparsity, parameter drift, model degradation, etc.) and decide whether to perform model recalibration or manual intervention. This function significantly improves the interpretability of model predictions and operational safety.

[0146] 3. Visualization of MPC optimization trajectory and RL policy iteration information

[0147] This module can simultaneously display the operation process and policy evolution of both Model Predictive Control (MPC) and Reinforcement Learning (RL) modules: MPC part: displays the optimization target value, constraint boundary and actual control trajectory in the form of dynamic curves, which makes it easy for operators to judge the convergence and execution stability of the control policy; RL part: displays the reward value changes, policy update frequency and reference policy weight adjustment information during the reinforcement learning training process in the form of iterative curves or heatmaps; the system supports the synchronous display of "real-time running status at the edge" and "policy training progress in the cloud", so that manual monitoring and model iteration keep information consistent.

[0148] 4. Historical data tracing and trend analysis

[0149] The visualization module has an embedded historical data management interface, which can retrieve and replay the segmented records for any time period, including raw sensor data, model prediction output, control command sequences, and abnormal event markers. The system can automatically generate trend analysis reports and support cross-comparison of multiple parameters (such as the relationship between power change and temperature field response). Historical data is used to assist in offline retraining, process backtracking, and quality tracking.

[0150] 5. Manual intervention entry point and model retraining trigger mechanism

[0151] This module provides a human-machine collaborative control interface. When the system experiences high uncertainty, parameter drift, or policy anomalies, the operator can directly intervene: the operator can manually switch to a conservative parameter set or pause the automatic control mode through the interface; the operator can also manually trigger the "model retraining" command to upload the current batch of data to the cloud for incremental updates or policy re-optimization; the module supports an event logging mechanism to automatically mark the time of manual intervention and the corresponding system status for subsequent model comparison and improvement.

[0152] Compared with the prior art, the present invention has the following beneficial effects:

[0153] Significantly improved real-time performance and meets industrial closed-loop control requirements: By compressing the key information of high-precision finite element analysis into a reduced-order model and running prediction and MPC solution in real time at the edge, parameter adjustment can be completed within a controllable low latency to meet the production line cycle time requirements.

[0154] Combining physical interpretability and data-driven adaptive capabilities: ROM retains physical priors and dynamic understanding, while surrogate models and RL provide data-driven performance enhancements and policy learning. The combination of the two can obtain more reliable and interpretable control behavior.

[0155] Uncertainty Quantification and Safety Assurance: The confidence interval provided by the proxy model is used to guide the conservative processing and proactive data collection of MPC, thereby prioritizing safety and quality under uncertain or abnormal operating conditions.

[0156] The project boasts strong deployability and scalability: the edge-cloud collaborative architecture supports iterative upgrades in the factory environment, facilitating migration across materials and machines, as well as long-term model maintenance.

[0157] Improved cutting quality and efficiency: By adjusting multiple parameters such as laser power, cutting speed, focus and gas flow rate in real time, thermal deformation can be effectively suppressed, kerf roughness and dimensional error can be reduced, while energy consumption can be optimized and the yield rate can be improved.

[0158] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A laser cutting device for metal product production, characterized by, The device comprises: a multi-parameter sensing unit for real-time acquisition of process and state parameters; the process and state parameters include laser power, cutting speed, auxiliary gas pressure, focal point position, and local temperature field of the workpiece; a state estimation unit for filtering and fusing the data collected by the multi-parameter sensing unit and estimating the implicit working condition state; a low-order physical model unit for short-time prediction based on a pre-established reduced-order physical model; an agent prediction unit for constructing an agent model based on the low-order physical model output and real-time measurement data and outputting processing quality prediction and uncertainty estimation; an optimization control unit for running a constrained online optimization control algorithm based on the low-order physical model prediction, quality prediction and uncertainty estimation of the agent model, and generating parameter adjustment instructions; and an execution and feedback unit for issuing the parameter adjustment instructions to the laser cutting execution mechanism and feeding back the cutting process signals to the state estimation unit and the agent prediction unit to realize virtual-real closed-loop control; wherein the low-order physical model unit extracts basis functions by using generalized POD or SVD on offline finite element simulation data and constructs a reduced-order dynamic system, the reduced-order state of which satisfies the formal expression:

2. The laser cutting device according to claim 1, characterized in that: the optimization control unit adopts a hybrid control architecture in which a model predictive control and a long-term optimization module based on policy learning work cooperatively to realize real-time coupled adjustment of multiple parameters of laser power, cutting speed, auxiliary gas pressure, and focal point position.

3. The laser cutting device according to claim 1, characterized in that: the multi-parameter sensing unit includes a laser power sensor, a beam analyzer, an infrared / near-infrared temperature imager, a slit profile camera, an auxiliary gas pressure sensor, and an online optical or acoustic sensor for detecting kerf roughness.

4. The laser cutting apparatus of claim 1, wherein, The agent prediction unit adopts a Gaussian process regression model, and the agent prediction unit takes the low-order physical model output, real-time measurement data, and historical processing data as input and outputs the predicted value and corresponding confidence interval of the processing quality index.

5. The laser cutting apparatus of claim 2, wherein, The optimization control unit includes a short-term safety optimization module that implements a hard-constrained model predictive control (MPC) based on the uncertainty estimation of the low-order physical model and the agent model, for generating safe execution instructions under the constraints of equipment and physical limitations, and a long-term policy optimization module that periodically or batch-wise optimizes the policy based on reinforcement learning (RL) and issues the reference policy or policy parameters generated by reinforcement learning to the short-term safety optimization module to improve process efficiency and quality. The long-term policy optimization module adopts one or more algorithms such as deep deterministic policy gradient (DDPG), soft behavior policy (SAC), or proximal policy optimization (PPO) for hybrid training in a simulation environment and a small amount of real samples, and uses the policy parameters obtained by offline training to guide the weight of the objective function of the model predictive control.

6.

7. The laser cutting apparatus of claim 5, wherein, ​ 8. The laser cutting apparatus of claim 7, wherein, The laser cutting device adopts an edge-cloud collaborative deployment architecture: the low-order physical model and model predictive control are run in real time in an edge computing unit to meet control cycle requirements; the long-term strategy optimization module, historical data management, and large-scale retraining are performed on a cloud or factory server, and strategy or model updates are periodically issued to the edge computing unit.

9. The laser cutting apparatus of claim 5, wherein, A visualization monitoring module is also included for real-time display of temperature field prediction cloud maps, agent model prediction confidence intervals, model predictive control optimization trajectories, reinforcement learning strategy iteration information, and historical data traceability records, and for providing a manual intervention entry and a model retraining triggering mechanism.

Images