Laser galvanometer parameter adjusting method and device based on machine learning

Abstract

The invention relates to the technical field of laser processing and automatic control, and discloses a laser galvanometer parameter adjusting method and device based on machine learning. The method comprises the steps that galvanometer scanning data and real-time environment variables are obtained, simulation optimization is conducted, and a stable machining path is obtained; machining is executed according to the stable machining path, real-time deviation is monitored, the deviation and the environment variables are fused, and a dynamic parameter set is determined; and performing quality evaluation according to a processing result driven by the dynamic parameter set, and generating a closed-loop control signal. According to the method, a complete closed loop from offline optimization to online correction to afterward learning is established, so that the problem of low machining precision and efficiency caused by strategy stiffness in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of laser processing and automatic control technology, and in particular to a method and apparatus for adjusting laser galvanometer parameters based on machine learning. Background Technology

[0002] Currently, laser processing technology plays a core role in modern manufacturing. Its high efficiency and precision make it widely used in precision cutting, welding, and surface treatment, and it is crucial for promoting intelligent and high-quality industrial production. As a key component controlling the scanning path of the laser beam, the precise adjustment of the laser galvanometer's parameters directly determines the final processing quality and efficiency.

[0003] In existing technologies, industrial control software often employs fixed parameters or manual adjustments to control laser galvanometers. While these methods meet basic requirements for standardized workpieces and stable processing environments, their limitations become apparent when processing scenarios become complex, such as when workpiece geometry is irregular or material properties change. Because these control strategies lack the ability to perceive and respond to real-time changes in the processing environment, the system cannot dynamically adjust key parameters such as scanning angle and frequency to adapt to instantaneous demands, ultimately leading to decreased processing accuracy or low equipment efficiency.

[0004] In summary, existing technologies suffer from low precision and efficiency in laser processing. Summary of the Invention

[0005] This invention provides a laser galvanometer parameter adjustment method and device based on machine learning to solve the problems of low precision and efficiency in existing laser processing technologies.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a machine learning-based method for adjusting laser galvanometer parameters, comprising: Acquire real-time environmental variables and galvanometer scanning data, process them, and generate the initial motion trajectory; The processing procedure is simulated based on the initial motion trajectory, and a stable processing path is obtained through iterative optimization. The processing operation is performed according to the stable processing path, and the actual processing deviation during the processing is obtained to generate a deviation correction increment; By combining the deviation correction increment with the real-time environmental variables, a dynamic parameter set for adjusting the processing operation is determined; Images of the processing results driven by the dynamic parameter set are acquired and evaluated to obtain quality assessment results; Based on the quality assessment results, a closed-loop control signal is generated to optimize subsequent processing.

[0007] Preferably, the step of acquiring real-time environmental variables and galvanometer scanning data, processing them, and generating an initial motion trajectory includes: The real-time environmental variables are acquired, and the galvanometer scanning data is collected to form the original data sequence; The original data sequence is filtered and then fused with the real-time environmental variables to obtain a smooth data sequence. Logical control processing is performed on the smoothed data sequence to generate a continuous adjustment sequence; Key transition points are extracted from the continuous adjustment sequence, and a smooth curve path is generated based on the key transition points. Interpolation filling is then performed to determine the initial motion trajectory.

[0008] Preferably, the step of simulating the processing based on the initial motion trajectory and iteratively optimizing to obtain a stable processing path includes: The processing is simulated based on the initial motion trajectory, surface quality indicators are calculated, and predicted indicator values ​​are obtained. If the predicted index value is lower than the preset standard threshold, the initial motion trajectory will be marked as an unqualified path; For the unqualified paths, a trajectory parameter set is obtained and optimized to obtain an updated trajectory parameter set; A new motion trajectory is generated based on the updated trajectory parameter set, and a stability judgment is performed to determine the stable processing path.

[0009] Preferably, the step of performing the processing operation according to the stable processing path, obtaining the actual processing deviation during the processing, and generating the deviation correction increment includes: Based on the stable machining path, an initial control signal is generated and sent to drive the galvanometer to perform the machining operation; During the processing operation, real-time processing deviation data is collected, processed, and the deviation value is determined. If the deviation value exceeds a preset deviation threshold, a correction operation is performed to obtain the deviation correction increment.

[0010] Preferably, the step of fusing the deviation correction increment with the real-time environmental variables to determine the dynamic parameter set for adjusting the processing operation includes: The deviation correction increment is fused with the real-time environmental variables to generate an initial deviation environment dataset. The initial deviation environment dataset is filtered to obtain the filtered deviation environment dataset. Calculate the comprehensive deviation index from the filtered deviation environment dataset; If the comprehensive deviation index exceeds the preset correction threshold, parameter optimization calculation is performed to determine the dynamic parameter set.

[0011] Preferably, the acquisition of processing result images of the processing operations driven by the dynamic parameter set, and the evaluation to obtain quality evaluation results, includes: The processing results of the processing operation driven by the dynamic parameter set are image acquired, and the acquired image data is cleaned to obtain a cleaned image dataset. Quality features are extracted from the cleaned image dataset to obtain a feature vector; Based on the feature vector, a quality assessment calculation is performed to determine the quality assessment result.

[0012] Preferably, generating a closed-loop control signal for optimizing subsequent processing based on the quality assessment result includes: Based on the quality assessment results, parameter optimization is performed to obtain a parameter adjustment scheme; The parameter adjustment scheme is compared with the preset parameter safety range, and the adjustment values ​​that exceed the parameter safety range are corrected to obtain the corrected adjustment scheme. The closed-loop control signal is generated according to the revised adjustment scheme.

[0013] Secondly, the present invention provides a laser galvanometer parameter adjustment device based on machine learning, comprising: The trajectory generation module is used to acquire real-time environmental variables and galvanometer scanning data, process them, and generate the initial motion trajectory. The path optimization module is used to simulate the processing process based on the initial motion trajectory and iteratively optimize to obtain a stable processing path; The deviation monitoring module is used to perform processing operations according to the stable processing path, obtain the actual processing deviation during the processing, and generate deviation correction increments. The parameter determination module is used to integrate the deviation correction increment with the real-time environmental variables to determine a dynamic parameter set for adjusting the processing operation; The quality assessment module is used to acquire images of the processing results of the processing operations driven by the dynamic parameter set, and to evaluate them to obtain quality assessment results. The signal generation module is used to generate a closed-loop control signal for optimizing subsequent processing based on the quality assessment results.

[0014] Thirdly, the present invention also provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the machine learning-based laser galvanometer parameter adjustment method described in any one of the above.

[0015] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the machine learning-based laser galvanometer parameter adjustment method described above.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention processes and plans the galvanometer scanning data, and performs iterative optimization by simulating the machining process. It transforms the abstract machining requirements into a precise and reliable stable machining path, which solves the problem that the existing technology relies on fixed parameters and the path planning is out of touch with the actual working conditions. This provides a reliable execution basis for subsequent high-precision online machining.

[0017] (2) By establishing a real-time feedback correction closed loop from actual processing deviation monitoring to dynamic parameter set adjustment, the present invention realizes online dynamic adjustment of the processing process, solves the problem of rigid control strategy of existing technology and inability to cope with sudden disturbances and environmental changes during processing, and improves the processing accuracy and stability of the system in dynamic and complex scenarios.

[0018] (3) By constructing a long-term learning mechanism from final product quality assessment to the next processing strategy optimization, this invention forms an intelligent processing control system that can self-iterate and continuously optimize, solving the problem of isolated single processing effect and lack of long-term adaptive learning ability in the existing technology, and realizing the continuous improvement of processing quality and efficiency. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the laser galvanometer parameter adjustment method based on machine learning provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the laser galvanometer parameter adjustment device based on machine learning provided in the second embodiment of the present invention. Detailed Implementation

[0020] 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.

[0021] Reference Figure 1 The first embodiment of the present invention provides a laser galvanometer parameter adjustment method based on machine learning, including the following steps: S11: Acquire real-time environmental variables and galvanometer scanning data, process them, and generate the initial motion trajectory; S12, Simulate the processing process based on the initial motion trajectory, and iteratively optimize to obtain a stable processing path; S13, perform processing operations according to the stable processing path, obtain the actual processing deviation during the processing, and generate deviation correction increment; S14, combine the deviation correction increment with the real-time environmental variables to determine a dynamic parameter set for adjusting the processing operation; S15, acquire the processing result image of the processing operation driven by the dynamic parameter set, evaluate it, and obtain the quality evaluation result; S16, Based on the quality assessment results, generate a closed-loop control signal for optimizing subsequent processing.

[0022] In step S11, real-time environmental variables and galvanometer scanning data are acquired, processed, and an initial motion trajectory is generated, including: The real-time environmental variables are acquired, and the galvanometer scanning data is collected to form the original data sequence; The original data sequence is filtered and then fused with the real-time environmental variables to obtain a smooth data sequence. Logical control processing is performed on the smoothed data sequence to generate a continuous adjustment sequence; Key transition points are extracted from the continuous adjustment sequence, and a smooth curve path is generated based on the key transition points. Interpolation filling is then performed to determine the initial motion trajectory.

[0023] In one implementation, a high-precision angle sensor deployed on the laser processing equipment collects the deflection angle of the galvanometer in real time at a preset sampling frequency, forming the raw data sequence. Simultaneously, physical parameters of the processing environment are acquired through sensors such as temperature and vibration sensors, serving as the real-time environmental variables.

[0024] It should be noted that the process of determining the sampling frequency includes: acquiring historical galvanometer scanning data covering various typical processing tasks (e.g., straight line filling, circular cutting, complex curve engraving); applying a Fast Fourier Transform (FFT) to each segment of historical data to transform it from the time domain to the frequency domain, generating a spectrum; identifying the frequency range where energy is concentrated from the spectrum, and determining the highest frequency point that can cover 99% of the signal energy, which is the highest effective frequency. According to the Nyquist sampling theorem, a frequency at least twice the highest effective frequency is selected as the final sampling frequency to ensure that the signal is not distorted during sampling. For example, the sampling frequency can be set to 1000Hz, that is, recording a coordinate point containing a timestamp and angle value every millisecond to ensure that the high-speed dynamic behavior of the galvanometer can be captured.

[0025] In one implementation, since the original data sequence may contain noise generated by equipment vibration or electromagnetic interference, if its signal-to-noise ratio is lower than a preset noise tolerance threshold, the original data sequence is filtered. The specific operations of this processing include: setting a sliding window with a preset window size (e.g., 5 data points); for each data point in the original data sequence, performing an arithmetic mean operation on the value of that data point and its preceding and following adjacent data points within the sliding window; and using the calculated average as the new value of that data point after filtering. After traversing the entire original data sequence, the set of all new values ​​constitutes the smoothed data sequence.

[0026] It should be noted that the window size is determined experimentally based on a trade-off between noise suppression level and signal response delay. It can be selected between 3 and 10 data points, determined through signal-to-noise ratio improvement and signal delay tests. For example, by using windows of different sizes (e.g., 3, 5, 7 data points) to perform filtering tests on historical datasets containing known characteristic signals and typical noise, a minimum window size is selected that effectively filters out most high-frequency noise while ensuring the signal's step response delay is below the system's allowable range.

[0027] It is worth noting that the process of determining the noise tolerance threshold includes: acquiring the original data sequences corresponding to "qualified products" and "defective products" from a large amount of historical processing data, and calculating their signal-to-noise ratios (SNRs) to form two sets of SNR datasets; setting a series of candidate thresholds for these two sets of SNR data; calculating the true positive rate (the proportion of correctly identified defective products) and false positive rate (the proportion of qualified products being mistakenly identified as defective products) for each candidate threshold; plotting the Receiver Operating Characteristic (ROC) curve with the false positive rate on the horizontal axis and the true positive rate on the vertical axis; determining the coordinate point on the curve that is closest to the upper left corner (0,1) in Euclidean distance, and the candidate threshold corresponding to this point is the noise tolerance threshold that can optimally distinguish between the two types of samples. The typical range of the noise tolerance threshold is 15-25dB, which can be adjusted according to the environmental noise level. For example, if the threshold determined by this method is 20dB, then any data sequence with an SNR lower than 20dB will trigger filtering processing.

[0028] In another implementation, statistical analysis is performed on the obtained smoothed data sequence within a sliding time window to calculate the instantaneous required values ​​of the galvanometer parameters. The specific calculation includes: within each time window, calculating the mean, variance, and peak deviation (the maximum absolute value of the difference between all data points and the mean within that time window) of the angle values ​​in the smoothed data sequence; comparing the calculated mean, variance, and peak deviation with a preset benchmark model to calculate the deviation. It is worth noting that the benchmark model is a parameter vector containing {ideal mean, ideal variance, ideal peak deviation}, and its establishment process is as follows: under ideal experimental conditions, a precisely calibrated laser galvanometer system is used to perform a standard processing task on a standard workpiece; galvanometer scanning data during this process is collected and smoothed; the statistical characteristics of the processed data within each time window are calculated, and their average value is taken. This average value is then solidified as the benchmark model, representing the standard of the system under undisturbed and stable operating conditions.

[0029] It should be noted that the calculation of the deviation involves treating the three statistical features of the current time window as points in a three-dimensional feature space and calculating the Euclidean distance between these points and the points represented by the benchmark model. This distance value is the deviation, which integrates deviations from multiple dimensions into a single, comparable scalar value. Then, logical control processing is performed to generate a set of target adjustment parameters for the current time point. This processing uses the deviation calculated in the previous step as continuous input, performs fuzzy inference and defuzzification calculations, and outputs a set of continuously corresponding, specific, target-adjusted parameter values ​​(e.g., target frequency, target angle). It is worth noting that the specific operations of the fuzzy inference and defuzzification calculation include: comparing the input deviation value with multiple pre-defined segmented intervals (e.g., numerical ranges representing "small deviation", "medium deviation", and "large deviation"), and calculating a matching score between 0 and 1, which represents the degree of conformity between the deviation value and each interval; traversing each rule in the fuzzy rule base, for a rule of the form 'IF deviation is "large" AND deviation change rate is "zero"', the system takes the minimum matching score corresponding to each condition in the 'IF' part of the rule as the trigger strength of the rule. Based on the trigger strength, the output fuzzy set defined by the 'THEN' part of the rule (which is itself a preset geometric shape, such as a triangle) is processed. Specifically, it is truncated along its vertical axis at a position with a height equal to the trigger strength. The lower half of the graphic retained after truncation is the output fuzzy shape of the rule. All the output fuzzy shapes generated by the triggered rules are superimposed to form a total, aggregated fuzzy output graphic. The centroid method is used to calculate the geometric center of the aggregated fuzzy output graphic, and the horizontal coordinate value of the center is the final output precise parameter adjustment amount.

[0030] It should be noted that the preset process of the fuzzy rule base is an automated rule mining process based on historical data. This process first acquires a training dataset containing tens of thousands of historical processing data snapshots. Each snapshot contains a set of precise, numerical input states (e.g., {deviation = 5.2, deviation change rate = 0.1}) and a corresponding, validated, optimal numerical control output (e.g., {frequency adjustment = +3.5Hz}). Before rule mining, these continuous numerical values ​​are discretized, mapped to category labels according to preset intervals (e.g., mapping deviation 5.2 to "large", and frequency adjustment +3.5Hz to "positive large"), forming a transaction database. Then, an association rule mining algorithm is applied to this transaction database. Through multiple iterative scans, the support of all individual 'state-output' itemsets (i.e., their frequency in the dataset) is calculated, and items below a preset support threshold are pruned. Next, longer candidate combinations are generated based on the remaining frequent itemsets, and this support calculation and pruning process is repeated until no new frequent itemsets can be generated. From the final frequent itemset, candidate association rules in the form 'IF{state combination}THEN{output}' are generated, and the confidence score of each rule (i.e., the conditional probability that 'output' also occurs when 'state combination' occurs) is calculated. Finally, only rules with confidence scores higher than a preset confidence threshold are retained. This set of rules with high support and high confidence constitutes the final fuzzy rule base. It is worth noting that the support and confidence thresholds are determined through cross-validation experiments during the model building phase. The determination process is as follows: the training dataset is randomly divided into k subsets; k-1 subsets are used alternately as training data, and the remaining subset is used as validation data. Rule mining and model building are performed under a series of candidate threshold combinations (e.g., support from 0.05 to 0.5, confidence from 0.6 to 0.95); the threshold combination that minimizes the average processing error of the final control system under the simulated test environment in k validations is selected as the final configuration.

[0031] In one implementation, to ensure the smoothness of the galvanometer motion, a smooth and continuous path needs to be constructed based on the continuous adjustment sequence. This construction process includes: first, calculating the second-order difference of the angle values ​​in the continuous adjustment sequence. The instantaneous angular acceleration is approximated, and key transition points where the second-order difference exceeds a preset turning threshold are extracted. These points define the basic shape and turning points of the path. Between every two adjacent key transition points, a series of high-density intermediate coordinate points are generated by solving a cubic polynomial equation that accurately passes through these two points and ensures that the velocity and acceleration remain continuous (i.e., the first and second derivatives are continuous) at the connection point. All intermediate points together constitute the smooth curved path. By calculating the Euclidean distance between all adjacent points on the smooth curved path, two points whose distance exceeds a preset gap distance threshold are identified. This region is the path gap. Between the two endpoints of the path gap, new coordinate points are generated equidistantly along the connecting line with a preset filling density. These newly generated points are inserted into the path sequence to complete the filling process, and finally, the initial motion trajectory is determined.

[0032] It should be noted that the specific operation of solving the cubic polynomial equation is as follows: For any two key transition points P1(x1,y1) and P2(x2,y2), and the known tangent slope (i.e., velocity) k1 at point P1 and the tangent slope k2 at point P2, establish a system of four linear equations with four unknown coefficients (a,b,c,d) constituting the following four constraints: the curve must pass through point P1; the curve must pass through point P2; the derivative of the curve at point P1 must be equal to k1; the derivative of the curve at point P2 must be equal to k2. By solving this system of equations, the cubic polynomial connecting P1 and P2 can be uniquely determined. The coefficient.

[0033] It should be noted that the preset turning threshold is determined based on physical modeling and analysis of the maximum permissible angular acceleration of the galvanometer reflector. This threshold is determined by the maximum permissible angular acceleration. Together with the sampling period T, the maximum permissible angular acceleration is determined. The value of the maximum allowable angular acceleration is multiplied by the square of the sampling period T, and the product is the final threshold value; wherein the value of the maximum allowable angular acceleration is derived from the physical specifications and mechanical design limits of the galvanometer system.

[0034] For example, if the maximum permissible angular acceleration of the system is At a sampling frequency of 1000Hz (T=0.001s), the corresponding turning threshold can be calculated to be 0.12 radians.

[0035] It should be noted that the preset filling density is determined by ensuring that the distance between any two adjacent points on the filled path is less than or equal to the distance the galvanometer can move within one control cycle at the maximum processing speed. This ensures that the drive system can interpret the filled path sequence as a smooth and continuous physical motion. It should also be noted that the preset gap distance threshold is determined based on research into laser processing technology. Experiments have shown that if the distance the laser spot moves per unit time exceeds a certain critical value, it will cause discontinuous processing marks or insufficient energy density on the workpiece surface. Process experiments have shown that if the laser spot moves more than 5 μm on the workpiece surface, a visible processing breakpoint will occur. Therefore, the preset gap distance threshold can be set to 5 μm. This physical distance can be converted into the galvanometer's deflection angle based on the galvanometer's scanning field and optical magnification.

[0036] In step S12, the processing process is simulated based on the initial motion trajectory, and a stable processing path is obtained through iterative optimization, including: The processing is simulated based on the initial motion trajectory, surface quality indicators are calculated, and predicted indicator values ​​are obtained. If the predicted index value is lower than the preset standard threshold, the initial motion trajectory will be marked as an unqualified path; For the unqualified paths, a trajectory parameter set is obtained and optimized to obtain an updated trajectory parameter set; A new motion trajectory is generated based on the updated trajectory parameter set, and a stability judgment is performed to determine the stable processing path.

[0037] In one implementation, the simulation of the machining process is completed through a pre-constructed physics-based simulation model. It should be noted that the process of pre-constructing the simulation model is a systematic project that includes model selection, parameterization, and experimental calibration. The specific operations are as follows: Select a physical model that can describe the laser-material interaction, such as a transient heat conduction model based on finite element analysis (FEA); Obtain a series of initial physical parameters required for the model from the technical manuals provided by material suppliers and the laser specifications, including the absorption rate of the material to laser of a specific wavelength, thermal conductivity, specific heat capacity, and the beam quality and power distribution of the laser; Conduct model calibration. This process is carried out by performing a series of benchmark physical experiments (for example, single scribing on standard materials at different powers and speeds), and using high-precision measuring instruments (such as confocal microscopes) to obtain the actual machining results (such as scribe depth, width); By continuously fine-tuning the key parameters in the model (especially those parameters that change significantly with temperature, such as the absorption rate), until the average error between the prediction results of the simulation model and the measurement results of all benchmark physical experiments is less than a preset convergence tolerance (for example, 0.01 mm), the model at this time is solidified and used for subsequent online simulation. It should be noted that the convergence tolerance refers to the mean absolute error (MAE) between the machining dimensions predicted by the simulation (such as scribe width) and the measured values of the benchmark physical experiments, and its value should be determined according to the dimensional tolerance of the final product. Generally, it is required that the MAE is less than 1 / 3 to 1 / 5 of the tolerance; Exemplarily, if the product tolerance is ±0.05 mm, the convergence tolerance can be set to 0.01 mm (i.e., 10 μm).

[0038] The model takes the parameters of the initial motion trajectory (including the coordinates, speeds, and curvature changes of each node on the trajectory) and laser parameters (such as power) and material properties (such as thermal conductivity) as inputs, and numerically simulates the machining process by solving the transient heat conduction equation that describes the laser energy absorption, heat conduction, and phase change processes, so as to predict the surface topography of the machined workpiece, and calculate the surface quality index according to this topography. Exemplarily, the surface quality index can be the surface roughness Ra value.

[0039] In another implementation, the calculated predicted index value is compared with a preset standard threshold. It should be noted that the standard threshold is derived from the engineering drawing of the target workpiece or the industry quality standard that it must follow (for example, the ISO surface finish grade required in a specific application scenario). It represents the minimum quality requirement that the product must meet to be judged as qualified. If the predicted index value fails to reach this standard, for example, the predicted surface roughness Ra value is greater than the standard threshold, the current initial motion trajectory is marked as a non-conforming path, and optimization processing is initiated.

[0040] It is worth noting that the specific operations of the optimization process include: defining a cost function whose output value is set as the square of the difference between the predicted index value and the standard threshold; using the trajectory parameter set constituting the initial motion trajectory (e.g., the coordinates of the control points of the spline curve) as the optimization variable, calculating the gradient of the cost function with respect to each parameter; updating each parameter by a small step along the negative direction of its corresponding gradient, the step size being determined by a preset learning rate; and finally, all the updated parameters together constitute the updated trajectory parameter set.

[0041] It should be noted that the learning rate is a key hyperparameter determined through offline experiments during the system debugging phase. The determination process is as follows: using a grid search method, optimization tests are performed on a set of standard validation trajectories from a series of candidate learning rate values ​​(e.g., 0.001, 0.01, 0.1); an optimal learning rate value is selected that allows the cost function to converge quickly without causing oscillations or divergence in the optimization process due to excessively large step sizes, and this value is chosen as the final configuration.

[0042] In one implementation, a new motion trajectory is regenerated based on the updated trajectory parameter set. Subsequently, the stability of this new trajectory is assessed to ensure it is physically stable and executable. This assessment process includes: in the time domain, calculating the maximum amplitude fluctuation of the new trajectory relative to an ideal smooth curve; if this fluctuation is below a preset amplitude fluctuation threshold, it passes time domain verification; in the frequency domain, performing a Fast Fourier Transform (FFT) on the acceleration curve of the new trajectory to calculate the sum of its high-frequency energy; if this energy is below a preset high-frequency energy threshold, it passes frequency domain verification. Only trajectories that pass both time-domain and frequency-domain verifications are considered stable.

[0043] It should be noted that both the amplitude fluctuation threshold and the high-frequency energy threshold are derived from the analysis of the physical characteristics of the galvanometer system. The former corresponds to the minimum lens vibration amplitude that can cause perceptible defects on the workpiece surface; the latter corresponds to the vibration energy level that may induce mechanical resonance in the system or lead to accelerated fatigue of the components. For example, through experimental calibration, the amplitude fluctuation threshold can be determined to be 0.05°, and the high-frequency energy threshold is a normalized value corresponding to 1% of the total energy.

[0044] It is worth noting that S12 is an iterative process. If the new motion trajectory generated based on the updated trajectory parameter set fails to pass the stability judgment, or if the predicted index value obtained after simulating the processing process again still does not reach the standard threshold, the system will return to perform optimization processing until a trajectory that can both pass the stability judgment and whose predicted index value meets the standard threshold is generated. This trajectory is then finally determined as the stable processing path.

[0045] In step S13, a processing operation is performed according to the stable processing path, and the actual processing deviation during the processing is obtained, generating a deviation correction increment, including: Based on the stable machining path, an initial control signal is generated and sent to drive the galvanometer to perform the machining operation; During the processing operation, real-time processing deviation data is collected, processed, and the deviation value is determined. If the deviation value exceeds a preset deviation threshold, a correction operation is performed to obtain the deviation correction increment.

[0046] In one implementation, the specific operation of generating and issuing the initial control signal based on the stable machining path includes: formatting and encapsulating the stable machining path (a sequence containing information such as timestamps and target angle coordinates) according to a preset low-latency, high-bandwidth communication protocol suitable for real-time industrial control; streaming the encapsulated data packets to the galvanometer drive system via a data interface; parsing the received data packets by the digital signal processor within the drive system and converting them into analog voltage or current signals that can directly drive the galvanometer deflection, which is the initial control signal. This process realizes the driving of the galvanometer to perform the machining operation. For example, the communication protocol can be an industrial Ethernet protocol (such as EtherCAT) to ensure that control commands are transmitted and executed within microsecond-level latency.

[0047] In another implementation, the acquisition of real-time machining deviation data during the machining operation is achieved through a machine vision-based feedback system. The specific operations of this system include: first, continuously capturing real-time images of the laser spot on the workpiece surface at a high frame rate using a high-speed camera deployed synchronously with the machining head; second, for each captured frame, locating the center point of the laser spot in the image coordinate system using an image processing algorithm (e.g., based on grayscale centroid calculation or edge detection); then, comparing the actual pixel coordinates of this center point with the theoretical pixel coordinates that should be present at the current timestamp according to the stable machining path, and calculating the deviation vector between the two at the pixel level; finally, converting the deviation vector into a deviation value expressed in physical units (e.g., millimeters) according to a pre-calibrated pixel-to-physical size conversion coefficient.

[0048] It should be noted that the pre-calibration process of the pixel-to-physical size conversion coefficient is a precision measurement operation completed during the system installation and debugging phase. The specific operation includes: first, placing a calibration plate with a known precision pattern (e.g., a ceramic plate printed with a dot matrix grid with a spacing of 1 mm) on the same focal plane as the workpiece; second, acquiring one or more clear images of the calibration plate using the high-speed camera; then, automatically identifying and extracting the center pixel coordinates of all grid points in the image using an image processing algorithm; finally, accurately calculating the conversion coefficient by calculating the average pixel distance between adjacent points in the image and comparing it with the known physical distance.

[0049] For example, if the average pixel distance between adjacent grid points in the horizontal direction is measured to be 150 pixels in the calibration image, and the physical distance is known to be 1 millimeter, then the conversion coefficient in the horizontal direction is determined to be 150 pixels / millimeter. Similarly, the conversion coefficient in the vertical direction can be determined, thereby achieving accurate conversion of any pixel deviation vector to a physical deviation vector.

[0050] It should be noted that, since a single measurement may contain random errors, the system will smooth the series of deviation values ​​calculated continuously within a very short time window (e.g., the most recent 10 data points), for example, by using the moving average method to calculate its mean, which is determined as the stable and reliable deviation value at the current moment.

[0051] It is worth noting that if the deviation value exceeds a preset deviation threshold, a correction operation is performed. This operation specifically includes multiplying the currently determined deviation value by a preset proportional gain coefficient, and the result is the deviation correction increment. This increment is a vector containing angle and frequency adjustments, used to compensate for the initial control signal in the next control cycle.

[0052] It should be noted that the preset deviation threshold is determined based on the dimensional tolerance requirements of the final product, through direct experimental calibration. The specific steps of this method include: in a series of controlled experiments, injecting a constant, artificially set offset into the control system, causing the galvanometer to intentionally generate a known, continuous real-time machining deviation when executing a standard machining path; gradually increasing the magnitude of this offset (e.g., starting from 0.01mm and increasing by 0.01mm each time), and performing high-precision dimensional measurements on the finished product after each experiment; comparing the measurement results with the tolerance range indicated on the product engineering drawings to find the maximum real-time machining deviation value that can produce a "qualified" product; multiplying this maximum deviation value by a safety factor less than 1 (e.g., 0.9), and the result is set as the preset deviation threshold.

[0053] It should be noted that the proportional gain coefficient is determined during the system commissioning phase using standard control system tuning methods (e.g., the Ziegler-Nichols tuning method). By conducting step response tests on the actual physical system, the value of this coefficient is experimentally adjusted until the system can correct deviations in the shortest possible time without overshooting or oscillations during the correction process. This value is then determined as the optimal configuration. For example, for a laser galvanometer system with a moderate response speed, the initial setting of this proportional gain coefficient can be tested within the range of [0.5, 2.0], and then fine-tuned based on the system response.

[0054] In step S14, the deviation correction increment and the real-time environmental variables are combined to determine a dynamic parameter set for adjusting the processing operation, including: The deviation correction increment is fused with the real-time environmental variables to generate an initial deviation environment dataset. The initial deviation environment dataset is filtered to obtain the filtered deviation environment dataset. Calculate the comprehensive deviation index from the filtered deviation environment dataset; If the comprehensive deviation index exceeds the preset correction threshold, parameter optimization calculation is performed to determine the dynamic parameter set.

[0055] In one implementation, the specific operations of the data fusion processing include: quantifying and normalizing the deviation correction increment and each physical quantity (such as temperature and vibration frequency) in the real-time environmental variables to eliminate dimensional differences; and concatenating these processed values ​​into a multi-dimensional feature vector, which is the initial deviation environmental dataset.

[0056] It should be noted that the numerical conversion process refers to converting non-numerical environmental variables (e.g., category labels such as "low," "medium," and "high" describing vibration states) into numerical values ​​according to a pre-established mapping table. For example, this mapping table can be defined as: {"low": 1, "medium": 2, "high": 3}. The normalization process can employ a max-min normalization method to map the original value X to a uniform, pre-defined interval (e.g., [0,1]).

[0057] In another implementation, to eliminate potential instantaneous fluctuations in the fused data, the initial bias environment dataset is filtered. This filtering can be implemented using a multidimensional moving average filter. Specifically, the process includes: maintaining a fixed-size queue to store the initial bias environment dataset (i.e., N feature vectors) from the most recent N time steps; at each new time step, adding the latest feature vector to the queue and removing the oldest feature vector; and performing an element-wise arithmetic average on all the current N feature vectors in the queue to generate a new feature vector, which is the filtered bias environment dataset.

[0058] It should be noted that the queue size N was determined experimentally during the system debugging phase. The determination process was as follows: Offline simulation was used to test the impact of different queue sizes (e.g., N=5, 10, 15) on the final control stability of the system; the smallest N value was selected as the final configuration, which effectively smoothed noise while ensuring that the system's response delay to state changes remained within an acceptable range. For example, if simulations show that when the queue size is 10, the system's response delay is below the acceptable range in 99% of operating conditions, then N can be set to 10.

[0059] It is worth noting that the specific operation of calculating the comprehensive deviation index from the filtered deviation environment dataset includes: multiplying each dimension component in the vector of the filtered deviation environment dataset by a preset risk weight coefficient; then, summing the results of multiplying all components by their weights, and obtaining a single scalar value, which is the comprehensive deviation index.

[0060] It should be noted that the risk weight coefficients were determined through offline feature importance analysis. The determination process includes: acquiring a dataset containing massive amounts of historical data, where each data point includes deviation correction increments during processing, real-time environmental variables, and the quality inspection results of the final product; and randomly dividing this dataset into training, validation, and test sets according to a preset ratio (e.g., 70% / 15% / 15%). A gradient boosting decision tree model is then trained on the training set. This training process iteratively constructs a series of decision trees, each new tree aiming to correct the residuals of all previous trees, and uses gradient descent at an adjustable learning rate to minimize the mean squared error loss function of the model on the training set. It is worth noting that, to obtain optimal model performance, the key hyperparameters of the model are determined through optimization on the validation set using a grid search combined with k-fold cross-validation. Specifically, the search is performed within a predefined hyperparameter grid space, which can be exemplarily set as follows: number of trees {100, 200, 300}, learning rate {0.01, 0.05, 0.1}, and maximum tree depth {3, 5, 7}. All hyperparameter combinations within this space are systematically tested, and the set that minimizes the average error of the model on the validation set is selected as the final configuration to prevent overfitting. Then, after model training, the feature permutation importance method is used to calculate the contribution of each input feature to the prediction result. The calculated contributions of each feature are normalized by summing all contribution values ​​and then dividing the contribution of each feature by this sum. The processed result is used as its corresponding risk weight coefficient.

[0061] In one implementation, if the calculated comprehensive deviation index exceeds a preset correction threshold, parameter optimization calculation is performed. This calculation is completed through a pre-trained parameter correction model, which takes the comprehensive deviation index as input and outputs a set of adjustments to the current galvanometer control parameters. These adjustments are then superimposed on the currently used parameters to obtain the dynamic parameter set.

[0062] It should be noted that the parameter correction model (e.g., a radial basis function network) is trained offline. The radial basis function network model structure includes: an input layer for receiving the comprehensive bias index; a hidden layer composed of multiple radial basis function neurons, each neuron having a center vector at its core, the activation value of which depends on the Euclidean distance between the input vector and this center vector; and an output layer for linearly combining the outputs of the hidden layer neurons to generate the final parameter adjustment. The training process is a two-stage process: First, unsupervised learning is performed on all input samples (i.e., the "question" sample set) in the historical dataset using the K-means clustering algorithm, dividing them into K clusters; the centroid of each cluster is set as the "center" vector of a radial basis function neuron in the hidden layer; after the hidden layer center is determined, the optimal connection weights from the hidden layer to the output layer are calculated by solving a system of linear equations, minimizing the mean squared error between the network output parameter adjustment and the labeled "answer" (i.e., the optimal parameter adjustment) in the dataset for each input sample in the training set.

[0063] It should be noted that the core hyperparameter of the model, namely the number of hidden layer neurons K, is determined during the model construction phase through testing on a validation set. The determination process is as follows: a series of candidate K values ​​are selected (e.g., K={10,20,50,100}), and the model is trained completely once for each K value; the K value that achieves the lowest prediction error on the validation set is selected as the final model configuration.

[0064] It is worth noting that the determination of the correction threshold is based on a quantitative analysis of the adjustment cost and performance gain. Too frequent small adjustments increase system instability (i.e., adjustment cost), while overly slow adjustments lead to the accumulation of correctable errors (i.e., performance loss). The determination process involves: quantitatively evaluating the long-term performance of the system under different thresholds through numerous simulation experiments. This evaluation specifically includes: first, setting a set of correction thresholds containing multiple candidate values; second, for each candidate threshold, running the complete processing simulation thousands of times in a standardized simulation environment containing various typical complex processing tasks, and recording two key performance indicators generated by each candidate threshold: the average quality error between the final simulated product and the ideal model, and the total rate of change of the control parameter adjustments.

[0065] After obtaining the performance index pairs (i.e., {mean quality error, total rate of change}) corresponding to all candidate thresholds, the two sets of heterogeneous data are subjected to min-max normalization to map all performance index values ​​to a unified [0,1] interval, eliminating dimensional differences. Subsequently, these normalized index pairs are plotted as points on a two-dimensional plane, and the "Pareto optimal front" is identified. This front consists of a series of points where one index cannot be optimized without sacrificing another. Finally, on this optimal front, the point with the closest Euclidean distance to the ideal target point (0,0) is selected. This point represents the optimal balance between the two objectives of "quality error" and "system instability." The candidate threshold corresponding to this optimal balance point is ultimately set as the corrected threshold.

[0066] In step S15, the processing result image of the processing operation driven by the dynamic parameter set is acquired and evaluated to obtain a quality evaluation result, including: The processing results of the processing operation driven by the dynamic parameter set are image acquired, and the acquired image data is cleaned to obtain a cleaned image dataset. Quality features are extracted from the cleaned image dataset to obtain a feature vector; Based on the feature vector, a quality assessment calculation is performed to determine the quality assessment result.

[0067] In one implementation, the image acquisition of the processing results is achieved by deploying a high-resolution industrial camera at a downstream workstation on the production line to photograph the processed workpiece area. Since the acquired raw image data may be affected by factors such as uneven ambient lighting or surface dust, it needs to undergo a cleaning process. This process specifically includes applying a Gaussian filter kernel to the acquired image data for convolution, smoothing out drastic fluctuations in pixel values ​​caused by random noise.

[0068] It should be noted that the size and standard deviation of the Gaussian filter kernel are determined during the system calibration phase. The determination process includes: acquiring a set of test images containing standard processing features and typical noise patterns; filtering the test images using different parameter combinations; subsequently, performing subsequent feature extraction algorithms such as Sobel edge detection on the filtered images, and calculating the measurement repeatability of the extracted features (such as line width) (e.g., calculating the standard deviation of multiple measurements); selecting the parameter combination that optimizes feature measurement repeatability (i.e., minimizes the standard deviation) as the final configuration, ensuring the best balance between effective noise reduction and preservation of key image details.

[0069] In another implementation, multiple quality features capable of quantifying processing quality are extracted from the cleaned image dataset to form a feature vector. It should be noted that the specific operations for extracting quality features include quantifying line width, edge sharpness, and surface uniformity. To determine the line width, the Sobel edge detection operator is applied to the cleaned image dataset to identify the left and right boundaries of the processed lines; along the line normal direction, the pixel distance between the left and right boundaries is calculated at multiple locations, and the average of all distance values ​​is multiplied by a pre-calibrated pixel-to-physical size conversion coefficient to obtain the final line width value. It is worth noting that the pixel-to-physical size conversion coefficient is a calibration parameter specifically used for the high-resolution camera performing post-processing quality assessment in S15, and it is independent of the conversion coefficient used by the feedback camera for real-time deviation monitoring in S13. Its determination method is the same as that described in S13, i.e., by photographing a calibration board with known physical dimensions and calculating the number of pixels corresponding to known physical distances in the image to accurately determine the coefficient. After identifying the line boundaries, the system further calculates edge sharpness by calculating the gradient magnitude of each boundary pixel and using the average gradient magnitude of all boundary pixels as a quantification of edge sharpness. Simultaneously, to evaluate surface uniformity, the system calculates the standard deviation of all pixel grayscale values ​​within the identified processing area. This standard deviation serves as a quantification of surface uniformity; a smaller value indicates a more uniform surface.

[0070] It is worth noting that the quality assessment calculation based on the feature vector is performed using a pre-trained machine learning model. This model can learn and express the complex, non-linear impact of each feature on the final quality. For example, this model can be a multilayer perceptron (MLP) neural network. The specific operation of this model is as follows: it receives feature vectors composed of line width, edge sharpness, surface uniformity, etc., as input layer data; through non-linear transformations in two hidden layers, it performs deep combination and mapping of the features; finally, it outputs a three-dimensional, continuous numerical vector through the output layer. Each dimension of this vector corresponds to a quality score for a sub-item, such as geometric accuracy score, dynamic response score, and surface consistency score. This three-dimensional vector is the quality assessment result.

[0071] It should be noted that the specific structure of the model can be exemplarily set as follows: an input layer containing 3 neurons (corresponding to line width, edge sharpness, and surface uniformity, respectively), a first hidden layer containing 10 neurons, a second hidden layer containing 5 neurons, and an output layer containing 3 neurons that generates the aforementioned three sub-item quality scores. Each hidden layer uses a Corrected Linear Unit (ReLU) as its activation function.

[0072] It should be noted that the multilayer perceptron model was trained offline using supervised learning. The training process includes: establishing a large labeled dataset, where each data sample contains a set of feature vectors extracted using the methods described above, and a corresponding accurate quality score obtained through physical testing with a high-precision metrology instrument as the true label; dividing the dataset into training, validation, and test sets; then, on the training set, iteratively adjusting the network weights and biases using backpropagation and gradient descent optimizers to minimize the mean squared error between the model's predicted score and the true label; during training, the model's key hyperparameters are determined by selecting the combination that minimizes the error on the validation set using a grid search combined with cross-validation. It is worth noting that the grid search is performed within a predefined hyperparameter space, which can be exemplarily set as follows: the number of hidden layers is {1,2}, the number of neurons per layer is {8,16,32}, and the learning rate is {0.001,0.01,0.1}. Through this optimization process, the optimal hyperparameter combination is exemplarily determined to be: 2 hidden layers, 16 neurons per layer, and a learning rate of 0.01. Training is complete when the model's error on the validation set reaches the convergence criterion. It should be noted that the convergence criterion can be determined using an early stopping strategy. Specifically, after each epoch of training, the model's error is evaluated on the validation set; if the error on the validation set does not decrease for several consecutive epochs (e.g., 10 consecutive epochs), the system determines that the model has reached its optimal performance point and terminates training early to prevent overfitting.

[0073] In step S16, based on the quality assessment results, a closed-loop control signal for optimizing subsequent processing is generated, including: Based on the quality assessment results, parameter optimization is performed to obtain a parameter adjustment scheme; The parameter adjustment scheme is compared with the preset parameter safety range, and the adjustment values ​​that exceed the parameter safety range are corrected to obtain the corrected adjustment scheme. The closed-loop control signal is generated according to the revised adjustment scheme.

[0074] In one implementation, the parameter optimization based on the quality assessment results is performed using a pre-trained parameter mapping model. It should be noted that this model takes the quality assessment results as input, and its output is a set of suggested adjustments for multiple core control parameters. These parameters include the learning rate of the trajectory optimization algorithm in S12, the proportional gain coefficient for real-time correction operations in S13, and the risk weight coefficient for calculating the comprehensive deviation index in S14. This set of suggested adjustments constitutes the parameter adjustment scheme.

[0075] It should be noted that the parameter mapping model (e.g., a multilayer perceptron neural network) is trained offline through supervised learning. The training process includes: First, establishing a training dataset derived from thousands of historical production cycles, where each data sample contains a "quality result" (i.e., a multi-dimensional quality evaluation vector obtained after a complete processing cycle) and a corresponding "optimal policy update" vector. It is worth noting that the "optimal policy update" vector is determined through offline, large-scale simulation experiments. The determination process is as follows: For each sample with a poor "quality result" in the historical dataset, in a high-fidelity system simulation environment, starting with the system parameters of that sample, multiple rounds of exploratory simulation are conducted; in each round of simulation, a small, systematic perturbation is made to one or more core control parameters, and the improvement effect of this perturbation on the processing quality of the next round of simulation is evaluated; through this process, a parameter adjustment vector that can maximize the improvement of the quality score is found, and this vector is used as the true label of the "optimal policy update" corresponding to that "quality result" sample.

[0076] After establishing the training dataset, stratified random sampling was used to divide it into training, validation, and test sets to ensure that the distribution of samples of different quality levels in each dataset was consistent with the original dataset. Then, on the training set, the weights and biases of the network were iteratively adjusted using the backpropagation algorithm and the Adam optimizer to minimize the mean squared error between the model's predicted "policy update" vector and the "optimal policy update" vector in the dataset.

[0077] It should be noted that the specific network structure of the parameter mapping model can be set as follows: an input layer with 3 neurons (corresponding to the three dimensions of the quality evaluation vector), two hidden layers with 20 and 10 neurons respectively (both using the ReLU activation function), and an output layer with 3 neurons (corresponding to the adjustment amount of the three core control parameters).

[0078] During training, the key hyperparameters of the model are determined by selecting the combination that minimizes the error on the validation set using a grid search combined with cross-validation. It's worth noting that the grid search is performed within a predefined hyperparameter space, which can be exemplarily set as follows: number of hidden layers {1, 2, 3}, number of neurons per layer {10, 20, 30}, and learning rate {0.001, 0.005, 0.01}. Training is complete when the model's error on the validation set reaches the convergence criterion using an early stopping strategy (i.e., the validation error no longer decreases for 10 consecutive epochs).

[0079] In another implementation, to ensure system stability, the parameter adjustment scheme needs to undergo a safety verification. This verification involves: for each suggested adjustment value in the scheme, the system queries a preset parameter configuration file for the corresponding safe operating limits; if the suggested adjustment would cause the parameter value to exceed the safe range, the adjusted value is forcibly corrected to the closest boundary value within that range. All verified and corrected adjustment values ​​together constitute the corrected adjustment scheme.

[0080] It is worth noting that the preset process of the parameter configuration file, i.e., the process of determining the safe range of the parameters, is completed during the system design and debugging phase by combining theoretical analysis and experimental stress testing. The specific operations include: analyzing the physical limits of the hardware by consulting the technical specifications of core components such as the laser galvanometer and driver to obtain their theoretical limit values, such as maximum allowable operating current, highest response frequency, and maximum angular acceleration, and setting a preliminary, conservative parameter range accordingly; conducting extensive stress tests by writing a series of extreme condition test programs in a controlled experimental environment to systematically and progressively push the values ​​of each core control parameter towards and beyond the initially set range boundaries, while simultaneously using high-speed cameras and sensors to monitor the system's real-time response (such as vibration and temperature rise) and the final processing effect; and through this testing, accurately identifying the critical parameter points at which the system begins to exhibit unacceptable behaviors such as unstable oscillations, significant decreases in control accuracy, or deterioration in processing quality. These critical points collectively define the final operating range for stable and reliable system operation, which is then solidified as the safe range in the parameter configuration file.

[0081] For example, if the parameter mapping model suggests increasing the learning rate of an optimization algorithm from 0.01 to 0.5, and the preset safe range of the parameter determined by the above method is [0.001, 0.1], then the system will correct the suggested value of 0.5 to the upper limit of the safe range of 0.1.

[0082] In one implementation, generating the closed-loop control signal based on the modified adjustment scheme specifically involves formatting the modified adjustment scheme into a structured data instruction containing parameter identifiers and new parameter values. It should be noted that this formatting can employ a well-known, human-machine readable data exchange format in the computer field, such as JSON (JavaScript Object Notation). After formatting, the structured data instruction becomes a text object containing key-value pairs, where the key is a predefined parameter identifier and the value is the modified value of that parameter. This instruction is sent to the system's top-level control module to update the baseline parameter configuration loaded when the next processing task starts. This structured data instruction, i.e., the closed-loop control signal, completes the feedback optimization from the result of one processing step to the strategy for the next processing step, thereby closing the long-term learning outer loop of the entire method.

[0083] It should be noted that the training data for the machine learning models (such as fuzzy rule bases, radial basis function networks, and multilayer perceptrons) comes from historical processing records, including galvanometer scanning data, environmental variables, bias data, and quality assessment results; the data scale is recommended to be no less than 10,000 sets of samples, covering typical processing tasks (such as straight line filling and curve engraving) and changes in working conditions (such as temperature fluctuations and vibration interference); data preprocessing includes normalization, denoising, and labeling to ensure the model's generalization ability.

[0084] It is worth noting that the system should include an anomaly detection mechanism, such as switching to a conservative control mode (e.g., fixed parameter backup) and triggering an alarm when sensor data exceeds limits or model prediction confidence is low; and regularly calibrating sensors and cameras to ensure data reliability.

[0085] Those skilled in the art should understand that the method can be implemented on an industrial control computer or embedded system. Recommended configurations include a multi-core processor (clock speed ≥ 2.0 GHz), memory ≥ 8 GB, and a GPU acceleration unit (for machine learning model inference). For scenarios with high real-time requirements, it is recommended to use an FPGA or DSP to process image data and control signals to ensure that the sampling frequency and response delay meet the processing requirements. Specific hardware selection can be adjusted according to the galvanometer speed, image resolution, and model complexity.

[0086] In summary, this invention discloses a laser galvanometer parameter adjustment method based on machine learning. By constructing a three-stage, dual-closed-loop intelligent control architecture that integrates offline simulation optimization, online real-time correction, and post-processing long-term learning, this invention solves the problem of low processing accuracy and efficiency caused by rigid control strategies and inability to dynamically adapt to changes throughout the processing process in existing technologies, and achieves highly intelligent and adaptive control of the laser processing process.

[0087] Reference Figure 2 The second embodiment of the present invention provides a laser galvanometer parameter adjustment device based on machine learning, comprising: The trajectory generation module is used to acquire real-time environmental variables and galvanometer scanning data, process them, and generate the initial motion trajectory. The path optimization module is used to simulate the processing process based on the initial motion trajectory and iteratively optimize to obtain a stable processing path; The deviation monitoring module is used to perform processing operations according to the stable processing path, obtain the actual processing deviation during the processing, and generate deviation correction increments. The parameter determination module is used to integrate the deviation correction increment with the real-time environmental variables to determine a dynamic parameter set for adjusting the processing operation; The quality assessment module is used to acquire images of the processing results of the processing operations driven by the dynamic parameter set, and to evaluate them to obtain quality assessment results. The signal generation module is used to generate a closed-loop control signal for optimizing subsequent processing based on the quality assessment results.

[0088] It should be noted that the machine learning-based laser galvanometer parameter adjustment device provided in this embodiment of the invention is used to execute all the process steps of the machine learning-based laser galvanometer parameter adjustment method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0089] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a machine learning-based laser galvanometer parameter adjustment program. When the processor executes the computer program, it implements the steps in the various machine learning-based laser galvanometer parameter adjustment method embodiments described above, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the trajectory generation module.

[0090] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0091] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0092] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0093] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0094] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0095] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0096] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for adjusting laser galvanometer parameters based on machine learning, characterized in that, include: Acquire real-time environmental variables and galvanometer scanning data, process them, and generate the initial motion trajectory; The processing procedure is simulated based on the initial motion trajectory, and a stable processing path is obtained through iterative optimization. The processing operation is performed according to the stable processing path, and the actual processing deviation during the processing is obtained to generate a deviation correction increment; By combining the deviation correction increment with the real-time environmental variables, a dynamic parameter set for adjusting the processing operation is determined; Images of the processing results driven by the dynamic parameter set are acquired and evaluated to obtain quality assessment results; Based on the quality assessment results, a closed-loop control signal is generated to optimize subsequent processing.

2. The laser galvanometer parameter adjustment method based on machine learning according to claim 1, characterized in that, The process of acquiring real-time environmental variables and galvanometer scanning data, processing them, and generating an initial motion trajectory includes: The real-time environmental variables are acquired, and the galvanometer scanning data is collected to form the original data sequence; The original data sequence is filtered and then fused with the real-time environmental variables to obtain a smooth data sequence. Logical control processing is performed on the smoothed data sequence to generate a continuous adjustment sequence; Key transition points are extracted from the continuous adjustment sequence, and a smooth curve path is generated based on the key transition points. Interpolation filling is then performed to determine the initial motion trajectory.

3. The laser galvanometer parameter adjustment method based on machine learning according to claim 1, characterized in that, The step of simulating the processing based on the initial motion trajectory and iteratively optimizing to obtain a stable processing path includes: The processing is simulated based on the initial motion trajectory, surface quality indicators are calculated, and predicted indicator values ​​are obtained. If the predicted index value is lower than the preset standard threshold, the initial motion trajectory will be marked as an unqualified path; For the unqualified paths, a trajectory parameter set is obtained and optimized to obtain an updated trajectory parameter set; A new motion trajectory is generated based on the updated trajectory parameter set, and a stability judgment is performed to determine the stable processing path.

4. The laser galvanometer parameter adjustment method based on machine learning according to claim 1, characterized in that, The step of performing processing operations according to the stable processing path, obtaining the actual processing deviation during the processing, and generating deviation correction increments includes: Based on the stable machining path, an initial control signal is generated and sent to drive the galvanometer to perform the machining operation; During the processing operation, real-time processing deviation data is collected, processed, and the deviation value is determined. If the deviation value exceeds a preset deviation threshold, a correction operation is performed to obtain the deviation correction increment.

5. The laser galvanometer parameter adjustment method based on machine learning according to claim 1, characterized in that, The process of integrating the deviation correction increment with the real-time environmental variables to determine the dynamic parameter set for adjusting the processing operation includes: The deviation correction increment is fused with the real-time environmental variables to generate an initial deviation environment dataset. The initial deviation environment dataset is filtered to obtain the filtered deviation environment dataset. Calculate the comprehensive deviation index from the filtered deviation environment dataset; If the comprehensive deviation index exceeds the preset correction threshold, parameter optimization calculation is performed to determine the dynamic parameter set.

6. The laser galvanometer parameter adjustment method based on machine learning according to claim 1, characterized in that, The process involves acquiring images of the processing results from the processing operations driven by the dynamic parameter set, evaluating them, and obtaining quality evaluation results, including: The processing results of the processing operation driven by the dynamic parameter set are image acquired, and the acquired image data is cleaned to obtain a cleaned image dataset. Quality features are extracted from the cleaned image dataset to obtain a feature vector; Based on the feature vector, a quality assessment calculation is performed to determine the quality assessment result.

7. The laser galvanometer parameter adjustment method based on machine learning according to claim 1, characterized in that, The step of generating a closed-loop control signal for optimizing subsequent processing based on the quality assessment results includes: Based on the quality assessment results, parameter optimization is performed to obtain a parameter adjustment scheme; The parameter adjustment scheme is compared with the preset parameter safety range, and the adjustment values ​​that exceed the parameter safety range are corrected to obtain the corrected adjustment scheme. The closed-loop control signal is generated according to the revised adjustment scheme.

8. A laser galvanometer parameter adjustment device based on machine learning, characterized in that, include: The trajectory generation module is used to acquire real-time environmental variables and galvanometer scanning data, process them, and generate the initial motion trajectory. The path optimization module is used to simulate the processing process based on the initial motion trajectory and iteratively optimize to obtain a stable processing path; The deviation monitoring module is used to perform processing operations according to the stable processing path, obtain the actual processing deviation during the processing, and generate deviation correction increments. The parameter determination module is used to integrate the deviation correction increment with the real-time environmental variables to determine a dynamic parameter set for adjusting the processing operation; The quality assessment module is used to acquire images of the processing results of the processing operations driven by the dynamic parameter set, and to evaluate them to obtain quality assessment results. The signal generation module is used to generate a closed-loop control signal for optimizing subsequent processing based on the quality assessment results.

Images