A laser operation process control system based on multi-element fusion

Abstract

This invention discloses a laser operation process control system based on multi-element fusion, specifically relating to the field of automatic control. It includes a multi-element perception and fusion module, an intelligent decision-making module, a multi-actuator collaborative control module, a system overall control and human-machine interaction module, and a self-learning and optimization module. By integrating multiple sensors to collect real-time data on the geometric shape, temperature field, plasma characteristics, and environment of the operation area, a "fusion situation information map" is generated through feature extraction and fusion. This map is then combined with a high-fidelity digital twin model for forward prediction, forecasting the morphological evolution, thermodynamic behavior, and potential defects of the processing area. Based on the prediction results, the laser, motion mechanism, and auxiliary gas unit are synchronously controlled via a high-speed industrial network to achieve multi-actuator collaborative operation, coordinating task scheduling, status monitoring, and human-machine interaction. By comparing actual data with prediction results, the model parameters and decision-making strategies are continuously optimized.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology, and more specifically, to a laser operation process control system based on multi-element fusion. Background Technology

[0002] Laser processing technology, with its high precision and high efficiency, plays an increasingly important role in high-end manufacturing. However, the laser operation process is a complex dynamic process involving strong coupling of multiple physical fields such as light, mechanics, electricity, heat, and materials. The processing effect is easily affected by the comprehensive influence of multiple factors of the workpiece, which leads to severe challenges to the stability of the processing process and the quality of the finished product.

[0003] Existing technologies typically collect key physical quantities (such as position or temperature) during the processing using one or a few sensors (such as vision sensors or thermal imagers), and input this data into a pre-established digital twin model. The model calculates and simulates part of the state of the processing area (such as the temperature field), and the system adjusts individual parameters such as laser power or motion speed accordingly in order to achieve preliminary closed-loop control of the process.

[0004] However, in practical use, it still has some shortcomings, such as a single perception dimension, lack of simultaneous acquisition and deep integration of multiple elements such as geometric shape, temperature field, plasma spectrum and environmental parameters, and inability to comprehensively and accurately depict the dynamic processing situation, resulting in insufficient decision-making basis. Digital twin models usually focus on static or quasi-static simulation of physical processes, lack high-fidelity forward prediction capabilities, and are difficult to predict thermodynamic evolution trends and potential defects, resulting in lagging control behavior. The system lacks self-learning and continuous optimization mechanisms, and the model parameters and decision-making strategies are fixed, making it impossible to iteratively correct according to actual processing results. It is difficult to adapt to complex working conditions such as material property fluctuations and equipment status changes, which limits the further improvement of processing quality and system intelligence level. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide a laser operation process control system based on multi-element fusion, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a laser operation process control system based on multi-element fusion, comprising a multi-element perception fusion module: integrating multiple sensors to synchronously collect multi-element perception data of the operation area, and performing spatiotemporal registration and feature-level fusion on the data through embedded feature extraction and association networks to generate a "fusion situation information map";

[0007] Intelligent decision-making module: Receives "fusion situation information map" as real-time input, accesses the high-fidelity digital twin model of laser operation running synchronously with the physical system, performs forward prediction simulation based on real-time data, obtains prediction data, and predicts the morphology, thermodynamic evolution trend and potential defects of the processing area;

[0008] Multi-actuator collaborative control module: Receives predictive data from the intelligent decision-making module and synchronously drives the laser, motion actuator, and auxiliary gas control unit through a high-speed real-time industrial network to form a collaborative execution network;

[0009] System control and human-machine interaction module: responsible for initializing all subsystems, coordinating data and instruction flows between modules, monitoring system operation status, and providing a graphical human-machine interface for importing work tasks, presetting process parameters, visualizing real-time "fusion situation information diagrams", and managing system alarms and logs;

[0010] The self-learning and optimization module continuously collects actual operation process data and final quality assessment results, compares them with the prediction data of the digital twin model, generates model error indicators, and periodically iterates and optimizes the intrinsic parameters and decision-making strategies of the prediction data through online learning.

[0011] The technical effects and advantages of this invention are as follows:

[0012] This invention utilizes multi-sensor fusion technology to simultaneously collect and deeply fuse geometric topography, temperature field, plasma state, and environmental parameters, generating a global "fusion situation information map." This overcomes the limitations of traditional systems with a single perception dimension. By combining a high-fidelity digital twin model for forward predictive simulation, it can predict the morphology and thermodynamic evolution trend and potential defects of the processing area in advance, enabling the control system to transform from a passive response to an active intervention, significantly improving the control accuracy and stability of the process.

[0013] This invention synchronously drives the laser, motion mechanism, and auxiliary gas unit through a high-speed real-time industrial network, forming a fast-responding collaborative execution network. It adopts a closed-loop control strategy, compares instructions and feedback in real time, and dynamically adjusts execution parameters to achieve precise synchronization between the laser focus and the motion trajectory. This effectively suppresses processing defects caused by parameter mismatch or external interference and ensures the consistency of work quality.

[0014] This invention utilizes an online learning mechanism to periodically iteratively optimize the model's intrinsic parameters and decision-making strategies, enabling the system to continuously adapt to changes in new materials, processes, and equipment conditions. This gradually reduces the discrepancy between predictions and actual results, forming a virtuous cycle of "online optimization - data accumulation - model update," thereby achieving continuous self-improvement in processing efficiency. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0016] Figure 2 This is a schematic diagram of the intelligent decision-making and digital twin prediction process of the present invention;

[0017] Figure 3 This is a flowchart illustrating the self-learning and optimization module of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] As attached Figure 1 Appendix Figure 2 and appendix Figure 3 The laser operation process control system shown includes a multi-element fusion module: integrating multiple sensors to synchronously collect multi-element sensing data of the operation area, and performing spatiotemporal registration and feature-level fusion on the data through embedded feature extraction and association networks to generate a "fusion situation information map".

[0020] It should be specifically noted that the multi-sensor includes a three-dimensional vision sensor, a multispectral thermal imager, a plasma spectral monitor, and an environmental sensor. The multi-element sensing data includes geometric topography, temperature field distribution, plasma plume characteristics, and environmental parameters.

[0021] It should be further explained that the 3D vision sensor selected is the Keyence LJ-V7000 series laser profile sensor, which uses blue laser (wavelength 405nm) to avoid interference from strong light during laser processing. The acquisition accuracy reaches ±0.5um, and the acquisition range covers 20mm×15mm (field of view). It acquires the geometric shape data of the workpiece in the working area in real time (including surface profile, dimensional deviation, and machining allowance). The reason for selection is that the 3D vision sensor has strong anti-interference ability and maintains high accuracy under high-speed movement (up to 10m / s), which meets the needs of dynamic operation scenarios such as laser cutting and welding.

[0022] The multispectral thermal imager uses the FLIR A655sc infrared thermal imager, which supports a resolution of 640×512 pixels, a spectral response range of 8-14um (mid-wave infrared), and a temperature measurement range of -20℃ to 1500℃. It collects real-time temperature field distribution data of the processing area (including hot spot temperature, temperature gradient, and heat-affected zone range). The reason for its selection is that, compared with traditional single-band thermal imagers, the multispectral design effectively eliminates the interference of smoke and dust on temperature measurement. During laser welding, it can accurately capture the temperature changes of the molten pool, avoiding the problems of workpiece burn-through due to excessively high temperature or weak welding due to excessively low temperature.

[0023] The plasma spectral monitor uses the Ocean Optics QE Pro-R high-resolution spectrometer, with a spectral acquisition range of 200-1100nm and a resolution of 0.02nm. It acquires in real time the plasma plume spectral data generated by the interaction between the laser and the material (including characteristic spectral line intensity, spectral linewidth, plasma electron temperature and density). The reason for this choice is that the plasma state directly reflects the coupling efficiency between the laser energy and the material. By monitoring its spectral characteristics, it can be determined whether the laser power matches the material properties.

[0024] The environmental sensor integrates temperature and humidity sensors, dust sensors, and vibration sensors to collect data on temperature, humidity, dust concentration, and equipment vibration in the working environment. It was chosen because environmental factors have a significant impact on the accuracy of laser operations. For example, excessive humidity can cause the laser lens to fog up, reducing the laser energy transmission efficiency, and equipment vibration can cause the laser focus to shift.

[0025] The process for generating a "fused situational awareness map" is as follows: The raw data collected by each sensor is preprocessed to eliminate noise and redundant information. Gaussian filtering is applied to the 3D vision data to remove random noise from the contour data. The specific formula is:

[0026]

[0027] in This is the Gaussian filter kernel function. Here, the standard deviation is... , , These are the relative coordinates of the pixels within the filtering window.

[0028] Then, the RANSAC algorithm is used to fit the workpiece reference surface, and the deviation between the actual contour and the reference surface is calculated to obtain standardized geometric feature data (contour deviation value, machining allowance distribution matrix). Multispectral thermal image data is filtered using median filtering (3×3 window size) to remove isolated hot spots caused by smoke interference. Then, radiometric calibration (based on blackbody calibration data provided by FLIR) is used to convert grayscale values ​​into actual temperature values, generating a temperature field matrix (the correspondence between pixels and temperature). Plasma spectral data is subjected to wavelet transform (db4 wavelet basis, decomposition level 3) to remove background noise. Then, spectral line fitting (Lorentz function) is used to extract the peak intensity and half-width at half-maximum of characteristic spectral lines, and the plasma electron temperature (based on the Boltzmann equation) and electron density (based on the Stark broadening effect) are calculated. The specific formula for calculating the plasma electron temperature is as follows:

[0029]

[0030] in, For electron temperature, , The energy level is at the atomic level. Boltzmann's constant, , For spectral line intensity, , For statistical weighting, , The wavelength is the spectral line wavelength.

[0031] The specific formula for calculating plasma electron density is as follows:

[0032]

[0033] in, For electron density, The full width at half maximum (FWHM) of the spectral line. This is the Stark broadening factor (related to the type of element).

[0034] Environmental data were smoothed using a moving average filter (window size 10) to analyze temperature, humidity, and dust concentration data. The frequency components of vibration data were analyzed using Fourier transform to identify abnormal equipment vibrations (such as 50Hz vibration caused by spindle imbalance).

[0035] Using the clock (1µs accuracy) of the system's main control unit as a reference, timestamps are added to the data from each sensor. Linear interpolation is used to interpolate low-frequency data to the time dimension of high-frequency data, ensuring time synchronization of data at the same moment. A spatial coordinate mapping relationship between each sensor is established using a calibration board. A 9×9 checkerboard calibration board is selected, with a checkerboard size of 10mm×10mm and 81 feature points evenly distributed within a 200mm×200mm area. The workpiece coordinate system of the 3D vision sensor is used as a reference, and the coordinates of the feature points on the calibration board are collected. The coordinates of the sensors, thermal imagers, and spectral monitors are calculated using the least squares method to solve the coordinate transformation matrix (rotation matrix and translation vector). This maps the temperature field data from the thermal imager and the plasma data from the spectral monitor to the workpiece coordinate system, achieving precise spatial alignment. After solving the coordinate transformation matrix, 10 feature points that were not included in the calculation are selected, and their coordinates in each sensor are transformed to the workpiece coordinate system using the transformation matrix. The deviation between the transformed coordinates and the actual coordinates is calculated. If the average deviation is less than 0.01 mm, the calibration is considered qualified; otherwise, the calibration is repeated.

[0036] The feature vectors of each data are extracted by a CNN network: the contour deviation feature vector (64 dimensions) is extracted from the geometric shape data, the temperature gradient feature vector (64 dimensions) is extracted from the temperature field data, the spectral line intensity feature vector (64 dimensions) is extracted from the plasma data, and the interference feature vector (32 dimensions) is extracted from the environmental data.

[0037] An attention mechanism is introduced to calculate the weights of each feature vector. The multiple feature vectors are then fused into a unified fused feature vector (128 dimensions) through weighted summation. The fused feature vector is then mapped to a two-dimensional image space, with the workpiece coordinate system as the horizontal and vertical axes and the fused feature value (such as the combined value of "contour deviation + temperature deviation + plasma state") as the gray depth, to generate a gradient "fused situation information map".

[0038] Intelligent decision-making module: Receives "fusion situation information map" as real-time input, accesses the high-fidelity digital twin model of laser operation running synchronously with the physical system, performs forward prediction simulation based on real-time data, obtains prediction data, and predicts the morphology, thermodynamic evolution trend and potential defects of the processing area.

[0039] It should be specifically noted that the digital twin model simulates the interaction process between laser and materials, and the model construction includes a geometric twin model, a physical twin model, and a process twin model.

[0040] The intelligent decision-making module includes real-time data input, forward prediction simulation, and decision generation, wherein the forward prediction simulation includes morphological prediction, thermodynamic evolution prediction, and potential defect prediction.

[0041] It should be further explained that, based on the workpiece geometric data collected by the 3D vision sensor, the 3D geometric model of the workpiece is constructed using BIM technology, including the workpiece material, size, and processing path parameters, and the geometric shape of the workpiece is updated in real time (such as the contour changes after cutting).

[0042] A physical model of laser-material interaction was constructed based on finite element analysis (FEA) and molecular dynamics (MD): The thermodynamic model was established using ANSYS APDL, with input parameters including laser power and scanning speed, to simulate the formation of the molten pool, heat conduction, and the evolution of the heat-affected zone. Specific analysis formulas are as follows:

[0043]

[0044] in, For material density, For specific heat capacity, For temperature, For time, Thermal conductivity, The intensity of the laser heat source.

[0045] The plasma model is built based on COMSOL Multiphysics. The laser wavelength and material composition are input to simulate the generation, expansion and radiation process of plasma and predict the spectral intensity and electron temperature. The mechanical model is built using ABAQUS to simulate the residual stress distribution of the workpiece after laser processing and predict the amount of deformation (such as warping deformation after welding).

[0046] Based on 1000 sets of historical laser welding parameters and quality data, a random forest algorithm was used to establish a mapping model between process parameters (laser power, scanning speed, auxiliary gas flow rate) and processing quality (weld strength, surface roughness). Initial process parameters were recommended according to the workpiece material and processing requirements. The 1000 sets of data covered three common materials: stainless steel (thickness 1-10mm), aluminum alloy (thickness 2-8mm), and carbon steel (thickness 3-12mm). Each material included parameter combinations with different thicknesses and different welding joint types (butt joint, corner joint). The training set and validation set were divided in a 7:3 ratio. The training set was used for model parameter learning, and the validation set was used to evaluate model performance. The mean absolute error (MAE) was used to evaluate the accuracy of process parameter prediction. The requirement was that the MAE be less than 5%. If the requirement was not met, the amount of data was increased and the model was retrained.

[0047] The geometric shape deviation, temperature field data, plasma data, and environmental data from the "fusion situational information map" are input into the digital twin model to update the model's real-time status.

[0048] Based on the updated digital twin model, forward prediction with a time step of 1ms is performed to predict the state changes of the processing area within the next 50ms: including morphological prediction, predicting the evolution of workpiece contour deviation (e.g., if the current deviation is 0.2mm and the laser power is too high, the deviation is predicted to expand to 0.3mm after 50ms), thermodynamic evolution prediction, predicting the diffusion trend of the temperature field (e.g., the heat-affected zone expands from 2mm to 2.5mm), determining whether the material's critical temperature is exceeded, and potential defect prediction, predicting potential defects (e.g., burn-through caused by excessive temperature, porosity caused by plasma instability), and calculating the probability of defect occurrence.

[0049] If the prediction results have no potential defects, the current process parameters are maintained; if defects exist, optimization decisions are generated based on the process twin model and reinforcement learning algorithm.

[0050] Multi-actuator collaborative control module: Receives predictive data from the intelligent decision-making module and synchronously drives the laser, motion actuator, and auxiliary gas control unit through a high-speed real-time industrial network to form a collaborative execution network.

[0051] It should be specifically noted that the collaborative execution network realizes synchronous control and dynamic adjustment of the actuators. The actuators are divided into selection and hardware connection. The design of the collaborative execution network includes synchronous control, dynamic adjustment, and fault redundancy.

[0052] It should be further explained that the laser used is an IPG YLR-10000 fiber laser, which supports analog (0-10V) and digital (EtherCAT) control, and adjusts the laser power, pulse frequency and duty cycle in real time. The motion actuator adopts Delta DVP series servo system, including X / Y / Z three-axis linear module (positioning accuracy ±0.005mm) and A-axis rotary module, supports EtherCAT bus control, and has a maximum motion speed of 1m / s. The auxiliary gas control unit uses SMC ITV series electro-proportional valve to control the flow rate (0-50L / min) and pressure (0-0.8MPa) of auxiliary gas (such as argon, oxygen), with a response time of less than 10ms. Hardware connection: Each actuator is connected to the system control unit through EtherCAT bus (1ms cycle). EtherCAT supports distributed clock synchronization to ensure that the control commands of each actuator are sent and executed synchronously.

[0053] Based on the predictive data from the intelligent decision-making module, control commands are generated for each actuator (e.g., laser power 8kW, X-axis speed 0.5m / s, gas flow rate 20L / min). The central control unit synchronously sends these commands to each actuator via the EtherCAT bus, ensuring precise alignment between the laser focus and the motion mechanism's trajectory. Real-time feedback data from the actuators (e.g., actual laser power, actual motion mechanism position) is received and compared with the control commands. A PID controller is used for closed-loop control. The specific analysis formula is as follows:

[0054]

[0055] in, For controller output, This is the proportionality coefficient. The integral coefficient is... These are the differential coefficients. This represents the deviation between the set value and the actual value.

[0056] The system master control and human-machine interaction module is responsible for initializing all subsystems, coordinating the data and instruction flows between modules, monitoring the system's operating status, and providing a graphical human-machine interface for importing work tasks, presetting process parameters, visualizing real-time "fusion situation information diagrams," and managing system alarms and logs.

[0057] It should be specifically noted that the graphical human-machine interface is combined with the monitoring system to achieve overall control, which specifically includes subsystem initialization, data flow and instruction flow coordination, and system status monitoring.

[0058] The system's central control and human-machine interaction module features a graphical user interface (HMI) divided into five functional areas: a task import area, a parameter preset area, a real-time visualization area, an alarm and log management area, and a system settings area. The task import area parses the processing path and dimensional requirements from the drawings to generate a work order. The parameter preset area provides preset templates for process parameters, automatically verifying their rationality after setting. The real-time visualization area displays the geometry, temperature field, and plasma state of the work area in real time, allowing users to intuitively observe dynamic changes during processing through multi-view switching. The alarm and log management area displays system alarm information in real time with alarm handling suggestions and automatically records system operation logs. The system settings area assists users in configuring system parameters and supports user permission management.

[0059] It should be further explained that when the system starts up, the central control unit automatically initializes each module, checks the sensor connection status, calibrates the actuator, and loads the digital twin model parameters. If the initialization fails, an audible and visual alarm is triggered (red LED flashing + buzzer alarm), and the cause of the fault is displayed on the human-machine interface.

[0060] A real-time database is used to store the data streams of each module (such as "fusion situation information map", prediction data, and actuator feedback data). The data storage period is configurable (such as 1 year for critical data storage and 3 months for general data storage). The instruction stream is managed through a message queue to ensure that the decision and control instructions sent by the central control unit are transmitted to each module according to priority, avoiding instruction congestion.

[0061] The system monitors the operating status of each module in real time (such as sensor acquisition frequency, actuator temperature, and network bandwidth) and displays the status indicators (green for normal, yellow for warning, and red for fault) on the human-machine interface. For example, when the laser temperature exceeds 50°C (warning threshold), the interface displays a yellow warning and prompts "Laser temperature is too high, it is recommended to reduce power". When the temperature exceeds 60°C (fault threshold), a red fault is displayed, and the laser is automatically shut down.

[0062] The self-learning and optimization module continuously collects actual operation process data and final quality assessment results, compares them with the prediction data of the digital twin model, generates model error indicators, and periodically iterates and optimizes the intrinsic parameters and decision-making strategies of the prediction data through online learning.

[0063] It should be specifically noted that the data collection scope of the self-learning and optimization module includes the collection of actual operation process data and the final quality assessment data. The online learning includes real-time error detection, decision strategy adjustment and optimization result feedback, wherein decision strategy adjustment includes constructing a reward function and strategy iteration.

[0064] It should be further explained that the system continuously collects real-time data from each module, including key parameters in the "fusion situational information map" at the same frequency as the "fusion situational information map" generation frequency (10fps), prediction data and generated optimization decisions from the intelligent decision-making module, actual operating parameters of multiple actuators at a frequency of 100fps (based on actuator feedback data), and real-time data from environmental sensors at a frequency of 10fps.

[0065] After the work is completed, the quality data of the workpiece is obtained through professional testing equipment as a benchmark for evaluating the system performance. Geometric accuracy testing uses a coordinate measuring machine to detect the key dimensions of the workpiece (length, hole diameter, flatness) and compares them with the design dimensions to obtain the geometric accuracy error. Surface quality testing uses a surface roughness meter to detect the surface roughness of the workpiece and takes surface images with an industrial camera to identify surface defects (scratches, pores, cracks). Mechanical property testing is used for welding, cladding and other operations, and a universal testing machine is used to test the mechanical properties of the workpiece (tensile strength, yield strength, elongation).

[0066] The data storage adopts a distributed database architecture to store data. The database stores structured data (such as geometric accuracy error, surface roughness, and mechanical performance parameters), supporting fast query and statistics. The distributed file system stores unstructured data (such as "fusion situation information map" images, surface defect images, and prediction simulation logs), meeting the long-term storage needs of large-capacity data (approximately 1GB of data per job).

[0067] Using the task ID as a unique identifier, the predicted data of the digital twin model is aligned with the actual task data in time and space. Time alignment is based on timestamps, mapping the predicted data (time step 1ms) and the actual data to the same time node. For example, the predicted temperature and the actual temperature are extracted at 10s after the start of the task. Spatial alignment is based on the workpiece coordinate system, aligning the spatial position of the predicted data with the spatial position of the actual data to ensure the spatial consistency of error calculation.

[0068] For different types of data, corresponding error calculation methods are used:

[0069] Numerical data (such as profile deviation, temperature, and laser power) use absolute error. With relative error Evaluate.

[0070] For distributed data (such as temperature field distribution and plasma spectral line intensity distribution), the similarity between the two distributions is calculated using the Barcol distance. The smaller the Barcol distance, the closer the predicted distribution is to the actual distribution. The specific analysis formula is as follows:

[0071]

[0072] in, For Bach distance, , These are the probability vectors of two distributions, respectively.

[0073] For categorized data (such as defect type "porosity" or "crack", and whether a defect has occurred), a confusion matrix is ​​used to calculate accuracy (number of correctly predicted samples / total number of samples × 100%), precision (number of samples predicted as positive and actually positive / number of samples predicted as positive × 100%), and recall (number of samples predicted as positive and actually positive / number of samples actually positive × 100%).

[0074] Set thresholds for error assessment indicators to determine whether the model needs optimization. For numerical data, the relative error threshold is set to 10%. If the relative error of a certain parameter exceeds 10% for 5 consecutive operations, the model part corresponding to that parameter needs optimization. For distributed data, the Parshall distance threshold is set to 0.2. If the Parshall distance of the temperature field distribution exceeds 0.2 for 3 consecutive operations, the thermodynamic model needs optimization. For categorized data, the accuracy threshold for defect prediction is set to 90%. If the accuracy is lower than 90%, the defect prediction model needs optimization.

[0075] In actual operation, if the error of a certain parameter suddenly exceeds the threshold (the absolute value of the rate of change of the value within 1 second is 1 or more), online learning optimization is triggered to adjust the decision-making strategy in real time to ensure the quality of the operation. Real-time error detection compares the predicted data with the actual data in real time. When the error of a certain parameter exceeds the threshold, the online learning process is automatically triggered. For example, in laser welding operation, the relative error between the actual molten pool temperature (1900℃) and the predicted molten pool temperature (1700℃) is 11.8%, which exceeds the threshold of 10%, triggering online optimization.

[0076] The decision-making strategy adjustment is based on a reinforcement learning algorithm (DQN algorithm, experience replay pool size 1000, exploration rate ε=0.1). The decision-making strategy of the intelligent decision-making module is adjusted in real time. This includes constructing a reward function with "error reduction degree", "processing quality compliance rate" and "operation efficiency" as indicators. For every 5% reduction in error, the reward is +10; for every compliance in processing quality, the reward is +20; and for every 5% reduction in operation efficiency, the penalty is -5. The strategy iteration continuously updates the Q-network parameters through interaction with the environment (actual operation process) and generates new decision-making strategies. The optimization result feedback applies the online-learned and optimized decision-making strategy to the current operation and records the optimization process (error change curve, decision-making strategy adjustment record), which is fed back to the database, forming a closed loop of "online optimization - data accumulation - model update - online optimization" to ensure the continuous evolution of the system.

[0077] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0078] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A laser operation process control system based on multi-element fusion, characterized in that, include: Multi-element perception fusion module: Integrates multiple sensors, including a 3D vision sensor, a multispectral thermal imager, a plasma spectral monitor, and an environmental sensor, to synchronously collect multi-element perception data of the work area. The multi-element perception data includes geometric shape, temperature field distribution, plasma plume characteristics, and environmental parameters. Through embedded feature extraction and association networks, the data is spatiotemporally registered and feature-level fused to generate a "fusion situation information map". The spatiotemporal registration includes adding timestamps to the data of each sensor based on the clock of the system's central control unit, using linear interpolation to ensure data time synchronization, and establishing the spatial coordinate mapping relationship of each sensor through a calibration board, mapping the temperature field data and plasma data to the workpiece reference coordinate system determined by the 3D vision sensor. Intelligent decision-making module: Receives "fusion situation information map" as real-time input, accesses the high-fidelity digital twin model of laser operation running synchronously with the physical system. The digital twin model includes a thermodynamic model and a plasma model simulating the interaction between laser and materials, constructed based on finite element analysis (FEA) and molecular dynamics (MD). Based on real-time data, it performs forward prediction simulation to obtain prediction data and predicts the morphology, thermodynamic evolution trend and potential defects of the processing area. Multi-actuator collaborative control module: Receives predictive data from the intelligent decision-making module and synchronously drives the laser, motion actuator, and auxiliary gas control unit through a high-speed real-time industrial network to form a collaborative execution network; The system master control and human-machine interaction module is responsible for initializing all subsystems, coordinating the data and instruction flow between modules, monitoring the system's operating status, and providing a graphical human-machine interface for importing work tasks, presetting process parameters, visualizing real-time "fusion situation information diagrams," and managing system alarms and logs. The self-learning and optimization module continuously collects actual operation process data and final quality assessment results, compares them with the prediction data of the digital twin model, generates model error indicators, and performs periodic iterative optimization of the intrinsic parameters and decision-making strategies of the prediction data through online learning. The online learning includes constructing a reward function based on reinforcement learning algorithms and iterating the strategy. The reward function is constructed based on the degree of error reduction, the processing quality compliance rate, and the operation efficiency indicators.

2. The laser operation process control system based on multi-element fusion according to claim 1, characterized in that: The feature-level fusion includes: extracting feature vectors from geometric topography, temperature field, plasma and environmental data through a CNN network, introducing an attention mechanism to calculate the weights of each feature vector, and fusing multiple feature vectors into a unified fusion feature vector through weighted summation.

3. A laser operation process control system based on multi-element fusion according to claim 1, characterized in that: The digital twin model simulates the interaction process between laser and materials. The model construction includes a geometric twin model, a physical twin model, and a process twin model. It also includes a process parameter and processing quality mapping model established using a random forest algorithm based on historical process parameters and quality data.

4. A laser operation process control system based on multi-element fusion according to claim 1, characterized in that: The intelligent decision-making module includes real-time data input, forward prediction simulation, and decision generation, wherein the forward prediction simulation includes morphological prediction, thermodynamic evolution prediction, and potential defect prediction.

5. A laser operation process control system based on multi-element fusion according to claim 1, characterized in that: The collaborative execution network enables synchronous control and dynamic adjustment of actuators. The actuators are divided into selection and hardware connection. The design of the collaborative execution network includes synchronous control, dynamic adjustment, and fault redundancy. The collaborative execution network uses EtherCAT bus to realize synchronous control of each actuator, and uses PID controller for closed-loop dynamic adjustment based on the deviation between the actuator feedback data and the control command.

6. A laser operation process control system based on multi-element fusion according to claim 1, characterized in that: The graphical human-machine interface (HMI) is integrated with the monitoring system to achieve overall control. Specifically, it includes subsystem initialization, data flow and command flow coordination, and system status monitoring. The HMI includes a task import area, parameter preset area, real-time visualization area, alarm and log management area, and system settings area. The parameter preset area provides process parameter preset templates and performs automatic rationality verification. The real-time visualization area displays geometric morphology, temperature field, and plasma status in real time and supports multi-view switching.

7. A laser operation process control system based on multi-element fusion according to claim 1, characterized in that: The system's central control unit, along with the human-machine interface module, develops a graphical user interface. The interface is divided into five functional areas: a task import area, a parameter preset area, a real-time visualization area, an alarm and log management area, and a system settings area. The task import area parses the processing path and dimensional requirements in the drawings to generate a work order. The parameter preset area provides process parameter preset templates, and automatically verifies the rationality of parameter settings. The real-time visualization area displays the geometry, temperature field, and plasma state of the work area in real time, allowing users to intuitively observe the dynamic changes during processing through multi-view switching. The alarm and log management area displays system alarm information in real time with alarm handling suggestions and automatically records system operation logs. The system settings area assists users in configuring system parameters and supports user permission management.

8. A laser operation process control system based on multi-element fusion according to claim 1, characterized in that: The self-learning and optimization module collects data including actual operation process data and final quality assessment data. The online learning includes real-time error detection, decision strategy adjustment, and optimization result feedback. Real-time error detection uses different error calculation methods for numerical data, distributed data, and categorical data, including absolute error / relative error, Bach distance, and accuracy, precision, and recall calculated by the confusion matrix. Decision strategy adjustment includes constructing a reward function and strategy iteration.

Images