Cutting parameter inversion method, device and storage medium for high-power laser cutting

Abstract

This invention discloses a method, apparatus, and storage medium for inverting cutting parameters in high-power laser cutting, relating to the field of laser cutting technology. The method includes: acquiring a cross-sectional image of the workpiece to be cut using an optical image acquisition platform; preprocessing and locating the effective region of the cross-sectional image to obtain an original cross-sectional image containing the effective cutting cross-section; extracting the cutting cross-sectional texture curve from the original cross-sectional image to obtain a main texture image containing anisotropic intensity and tangential angle; superimposing the extracted cutting cross-sectional texture curve onto the original cross-sectional image to obtain a physical prior image containing the cutting cross-sectional texture curve; constructing a multi-task joint prediction model based on IMAF-ConvNeXt; and outputting prediction results of the laser cutting speed and defocusing amount during the actual cutting process based on the original cross-sectional image, main texture image, and physical prior image using the multi-task joint prediction model. Thus, accurate inversion of laser cutting speed and defocusing amount is achieved.

Description

Technical Field

[0001] This invention relates to the field of laser cutting technology, and in particular to a method for inverting cutting parameters in high-power laser cutting, a computer-readable storage medium, and a device for inverting cutting parameters in high-power laser cutting. Background Technology

[0002] With the deepening advancement of intelligent manufacturing, high-power fiber laser cutting has become a core processing technology in high-end equipment manufacturing fields such as aerospace and shipbuilding. In actual industrial production environments, unavoidable disturbances such as laser power attenuation, focal shift caused by thermal lensing, and optical lens contamination frequently alter the actual energy density acting on the material, leading to deviations from the preset process parameters of the CNC system. Traditional automation processes, relying on operator experience or preset parameters for quality control, have fundamental flaws. Therefore, predicting actual laser processing parameters based on the optical morphology of the cut surface is urgently needed for robust evaluation of processing conditions, fault diagnosis, and closed-loop process compensation.

[0003] In related technologies, although data-driven deep learning models have been widely used in industrial defect detection, existing end-to-end deep networks face a serious "black box" bottleneck when processing laser-cut cross-sectional images. Existing models typically treat cross-sectional images as purely two-dimensional visual data and are limited by the local receptive field of standard convolution. More seriously, due to the lack of explicit spatial guidance from the molten metal flow trajectory, existing models are highly susceptible to "attention drift," meaning they are misled by strong specular reflections on the metal surface or slag buildup at the bottom, resulting in poor prediction robustness and a complete loss of physical interpretability. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in related technologies. To this end, the first objective of this invention is to propose a method for inverting cutting parameters in high-power laser cutting, which utilizes explicit physical priors to guide a deep neural network, enabling accurate inversion of high-power laser cutting speed and defocusing amount.

[0005] A second objective of this invention is to provide a computer-readable storage medium.

[0006] The third objective of this invention is to provide a cutting parameter inversion device for high-power laser cutting.

[0007] To achieve the above objectives, the high-power laser cutting parameter inversion method proposed in the first aspect of the present invention includes: acquiring a cut section image of the workpiece to be cut through an optical image acquisition platform, and preprocessing and locating the effective region of the cut section image to obtain an original cross-sectional image containing the effective cut section; extracting the cut section texture curve from the original cross-sectional image to obtain a main texture image containing anisotropic intensity and tangential angle; superimposing the extracted cut section texture curve onto the original cross-sectional image to obtain a physical prior image containing the cut section texture curve; constructing a multi-task joint prediction model based on IMAF-ConvNeXt, and outputting the laser cutting speed and defocusing amount prediction results in the actual cutting process based on the original cross-sectional image, the main texture image, and the physical prior image through the multi-task joint prediction model.

[0008] According to an embodiment of the present invention, a method for inverting cutting parameters in high-power laser cutting involves acquiring a cross-sectional image of the workpiece to be cut using an optical image acquisition platform. The cross-sectional image is preprocessed and the effective region is located to obtain an original cross-sectional image containing the effective cutting cross-section. Furthermore, the original cross-sectional image is processed to extract the cutting cross-sectional texture curve, obtaining a main texture image containing anisotropic intensity and tangential angle. The extracted cutting cross-sectional texture curve is then superimposed onto the original cross-sectional image to obtain a physical prior image containing the cutting cross-sectional texture curve. Subsequently, a multi-task joint prediction model based on IMAF-ConvNeXt is constructed. Based on the original cross-sectional image, the main texture image, and the physical prior image, the multi-task joint prediction model outputs the predicted laser cutting speed and defocusing amount during the actual cutting process. Thus, by utilizing explicit physical prior to guide a deep neural network, accurate inversion of high-power laser cutting speed and defocusing amount is achieved.

[0009] In addition, the high-power laser cutting parameter inversion method according to the above embodiments of the present invention may also have the following additional technical features: According to an embodiment of the present invention, obtaining the original cross-sectional image containing the effective cutting section includes: converting the cutting section image into a grayscale image and performing edge-preserving and noise reduction processing using bilateral filtering; converting the grayscale image into a binary image using an adaptive threshold and performing morphological closing operation using elliptical structuring elements to fill the internal holes of the section area and connect the fracture contours; extracting all external contours through contour detection and filtering and retaining the effective section area according to the contour area; obtaining the minimum circumscribed rotation rectangle of the effective section area and performing reverse rotation correction on the original cross-sectional image according to the tilt angle of the workpiece to be cut, so that the longitudinal axis of the section is aligned with the vertical axis of the image; cropping the section area using a high-coverage strip cropping strategy, and drawing the boundary contour and enclosing rectangle of the effective cutting section on the original cross-sectional image to form the original cross-sectional image containing the effective cutting section.

[0010] According to an embodiment of the present invention, the step of acquiring a main texture image containing anisotropic intensity and tangential angle includes: performing image denoising enhancement on the original cross-sectional image; calculating the horizontal and vertical gradients of the original cross-sectional image using the Sobel operator, and constructing a local structure tensor matrix based on the horizontal and vertical gradients; performing eigenvalue decomposition on the local structure tensor matrix to obtain principal and secondary eigenvalues, and obtaining the anisotropic intensity and tangential angle based on the principal and secondary eigenvalues; mapping the tangential angle to a biangular domain, constructing two continuous weighted vector component fields by combining an effective cross-sectional mask, and reconstructing a globally continuous, phase-singular, pure tangential field from the two continuous weighted vector component fields using inverse trigonometric functions; and performing discrete Euler integral trajectory tracking on the pure tangential field and the anisotropic intensity to form the main texture image containing anisotropic intensity and tangential angle.

[0011] According to one embodiment of the present invention, the discrete Euler integral trajectory tracking of the pure tangential field and the anisotropic intensity includes: setting a starting point at the central axis of the transverse section, performing a line-by-line downward stepping using discrete Euler integral, and defining the first... The horizontal coordinate state update equation for the row is: ,in, Let be the projection operator constrained by the cross-section mask, and let the core lateral step offset be defined as... Among them, dynamic weight function ,in, and It is a constant. Let be the local tangent angle of the texture at position i. Let be the anisotropy intensity of the texture at the i-th row position.

[0012] According to an embodiment of the present invention, obtaining a physical prior image containing the cross-sectional texture curve includes: smoothing the extracted cross-sectional texture curve through triple joint regularization to generate a sub-pixel continuous texture curve; rendering the sub-pixel continuous texture curve and superimposing it onto the original cross-sectional image to form the physical prior image containing the cross-sectional texture curve.

[0013] According to one embodiment of the present invention, the smoothing process of the extracted cross-sectional texture curve by triple joint regularization includes: initially suppressing local curvature abrupt changes in the one-dimensional discrete skeleton by using weighted cubic smoothing splines, and introducing a Whittaker smoothing model containing a second-order difference penalty term, and minimizing the energy functional. To achieve global smoothness, where, Let be the energy function to be minimized, x be the observation coordinate vector of the initial discrete texture skeleton, and z be the coordinate vector after smooth reconstruction. For data fidelity items, It is a second-order difference operator matrix. For curvature penalty term, For adaptive curvature penalty weights, and by employing a one-dimensional Gaussian filter... High-frequency artifact micro-polishing is performed, and the final coordinate update equation is: ,in, Let be the final smoothed coordinates of the i-th point after triple regularization. Let i be the coordinates of the i-th point after Whittaker smoothing. The standard deviation is One-dimensional Gaussian smoothing kernel, This is the convolution operator. This is a factor that preserves the original features.

[0014] According to an embodiment of the present invention, the step of outputting the laser cutting speed and defocusing amount prediction results in the actual cutting process based on the original cross-sectional image, the main texture image, and the physical prior image through the multi-task joint prediction model includes: performing shallow feature extraction and local feature refinement on the original cross-sectional image, the main texture image, and the physical prior image respectively through the multi-task joint prediction model; performing industrial multimodal alignment and fusion on the refined local features to obtain fused features, wherein the fused features include micro-roughness information of the original cross-sectional image, texture anisotropic physical field information, and explicit spatial guidance information of the texture curve; inputting the fused features into the ConvNeXt backbone network to output a high-dimensional feature map; performing feature refinement deep convolution on the high-dimensional feature map to obtain dimensionality reduction features; inputting the dimensionality reduction features into a global attention module to output a globally shared feature vector; and inputting the globally shared feature vector into two heterogeneous prediction branches to obtain the laser cutting speed and defocusing amount prediction results respectively.

[0015] According to an embodiment of the present invention, the method further includes: performing anti-slag truncation processing on the original cross-sectional image based on a local median statistical model, wherein the anti-slag truncation processing includes: calculating the absolute horizontal first-order difference of the discrete tracking trajectory; when a continuous horizontal coordinate mutation is detected to exceed the adaptive reference threshold, determining that the tracking trajectory has entered the poor slag discharge interference zone at the bottom of the specimen, and actively activating the truncation mechanism to remove the slag-covered portion.

[0016] To achieve the above objectives, a computer-readable storage medium is provided in the second aspect of the present invention, which stores a cutting parameter inversion program for high-power laser cutting. When the cutting parameter inversion program for high-power laser cutting is executed by a processor, it implements the cutting parameter inversion method for high-power laser cutting described in the embodiments of the present invention.

[0017] According to embodiments of the present invention, a computer-readable storage medium can accurately invert the cutting parameters of high-power laser cutting by executing a cutting parameter inversion program stored thereon, using explicit physical priors to guide a deep neural network.

[0018] To achieve the above objectives, the high-power laser cutting parameter inversion device proposed in the third aspect of the present invention includes: a first acquisition module, used to acquire a cross-sectional image of the workpiece to be cut through an optical image acquisition platform, and to preprocess and locate the effective area of ​​the cross-sectional image to acquire an original cross-sectional image containing the effective cross-section; a second acquisition module, used to extract the cross-sectional texture curve of the original cross-sectional image to acquire a main texture image containing anisotropic intensity and tangential angle; a third acquisition module, used to superimpose the extracted cross-sectional texture curve onto the original cross-sectional image to acquire a physical prior image containing the cross-sectional texture curve; and a parameter inversion module, used to construct a multi-task joint prediction model based on IMAF-ConvNeXt, and to output the laser cutting speed and defocusing amount prediction results in the actual cutting process based on the original cross-sectional image, the main texture image and the physical prior image through the multi-task joint prediction model.

[0019] According to an embodiment of the present invention, a high-power laser cutting parameter inversion device acquires a cross-sectional image of the workpiece to be cut through an optical image acquisition platform via a first acquisition module. The first acquisition module preprocesses and locates the effective region of the cross-sectional image to obtain an original cross-sectional image containing the effective cross-section. A second acquisition module extracts the cross-sectional texture curve from the original cross-sectional image to obtain a main texture image containing anisotropic intensity and tangential angle. A third acquisition module overlays the extracted cross-sectional texture curve onto the original cross-sectional image to obtain a physical prior image containing the cross-sectional texture curve. Then, a parameter inversion module constructs a multi-task joint prediction model based on IMAF-ConvNeXt. Based on the original cross-sectional image, the main texture image, and the physical prior image, the multi-task joint prediction model outputs the predicted laser cutting speed and defocusing amount during the actual cutting process. Thus, by utilizing explicit physical priors to guide a deep neural network, accurate inversion of high-power laser cutting speed and defocusing amount is achieved.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the cutting parameter inversion method for high-power laser cutting according to an embodiment of the present invention; Figure 2 This is a cross-sectional view of a 10mm thick stainless steel workpiece collected according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating a method for inverting cutting parameters in high-power laser cutting according to an embodiment of the present invention. Figure 4 This is a diagram showing the intelligent positioning effect of the cut section area acquired according to an embodiment of the present invention. Figure 5 This is a flowchart illustrating a method for inverting cutting parameters in high-power laser cutting according to another embodiment of the present invention; Figure 6 It is an anisotropic intensity map within a mask region acquired according to an embodiment of the present invention; Figure 7 This is a tangential guidance field matrix diagram within a mask region acquired according to an embodiment of the present invention; Figure 8 This is a flowchart illustrating a method for inverting cutting parameters in high-power laser cutting according to yet another embodiment of the present invention. Figure 9 This is a schematic diagram of the texture curves of the cut surface under different cutting qualities collected according to an embodiment of the present invention; Figure 10 This is a flowchart illustrating a high-power laser cutting parameter inversion method according to another embodiment of the present invention. Figure 11 This is a block diagram of a cutting parameter inversion device for high-power laser cutting according to an embodiment of the present invention. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following description, with reference to the accompanying drawings, describes a method for inverting cutting parameters for high-power laser cutting, a computer-readable storage medium, and a device for inverting cutting parameters for high-power laser cutting according to embodiments of the present invention.

[0024] It should be understood that in practical high-power laser cutting applications, uncontrollable external disturbances such as laser power attenuation and optical lens contamination often cause the actual processing state to deviate from the preset parameters, leading to a significant decrease or even deterioration in cutting quality. In particular, when faced with deteriorating cutting quality, traditional methods are difficult to quickly and accurately invert and diagnose the specific process parameters that have deviated, resulting in parameter adjustments that heavily rely on human experience and trial and error, which is inefficient and difficult to adapt to complex working conditions. To address this, this invention proposes a cutting parameter inversion method for high-power laser cutting. By using explicit physical priors to guide a deep neural network, the laser cutting speed and defocus parameters under the current processing state can be stably and accurately inverted and output.

[0025] Figure 1 This is a flowchart illustrating the cutting parameter inversion method for high-power laser cutting according to an embodiment of the present invention.

[0026] Specifically, in some embodiments of the present invention, such as Figure 1 As shown, the cutting parameter inversion method for high-power laser cutting includes: S101 acquires the cut section image of the workpiece to be cut through the optical image acquisition platform, and performs preprocessing and effective area positioning on the cut section image to obtain the original cross-sectional image containing the effective cut section.

[0027] It is understood that in this embodiment of the present invention, the optical image acquisition platform includes a workpiece to be cut, an isotropic diffuse illumination source, a high-resolution industrial vision sensor, and an image acquisition and control computer. Specifically, when acquiring images of the cut surface, the workpiece to be cut is fixed at the acquisition station, and the isotropic diffuse illumination source is turned on. This source provides uniform illumination to the cross-section through indirect scattered light, effectively suppressing strong specular reflection and local pixel saturation of the stainless steel cross-section, and enhancing the grayscale contrast between texture and background. Simultaneously, the high-resolution industrial vision sensor is mounted perpendicular to the cross-section, and the working distance is adjusted to ensure clear imaging of the cross-section on the focal plane. Furthermore, the image acquisition and control computer is connected to the isotropic diffuse illumination source and the high-resolution industrial vision sensor, uniformly controlling imaging parameters such as exposure time and gain, and receiving and storing the cut surface images. It should be noted that all imaging parameters are strictly locked once calibrated, thereby ensuring the consistency of the entire dataset.

[0028] For example, in some embodiments of the present invention, the laser cutting experiment used a Raycus-20kW fiber laser (wavelength 1080nm, maximum power 20kW) to cut a 10mm thick SUS304 stainless steel sheet. The auxiliary gas was high-pressure nitrogen (pressure 1.0MPa), the cutting speed was 5~16 m / min (step size 1 m / min), and the defocusing amount was +4~-9 mm (step size 1 mm). A total of 112 sets of effective process parameter combinations were designed, and each set of parameters was used to perform 3 repeated cutting experiments. Four cross-sectional areas were uniformly cut from each workpiece, and a total of 1344 high-definition cross-sectional images were acquired. Specifically, during the acquisition process, all optical and imaging parameters (working distance, focal plane, exposure time, sensor gain, etc.) were globally calibrated and strictly locked to ensure the measurement consistency of the entire dataset. The acquired cross-sectional images of the 10mm thick stainless steel workpiece are shown below. Figure 2 As shown.

[0029] S102, extract the cross-sectional texture curve of the original cross-sectional image to obtain the main texture image containing anisotropic intensity and tangential angle.

[0030] It is understood that, in this embodiment of the present invention, the cross-sectional texture curve refers to the physical trajectory left after the molten metal flows and solidifies rapidly. Accurately extracting its geometric skeleton is the key to converting the micro-morphology into a quantifiable physical prior. Here, the anisotropic intensity is used to characterize the clarity of the texture curve, and the tangential angle is used to characterize the texture tangent direction angle of the texture curve.

[0031] S103, the extracted cross-sectional texture curve is superimposed on the original cross-sectional image to obtain a physical prior image containing the cross-sectional texture curve.

[0032] It is understood that, in this embodiment of the present invention, a physical prior image containing the cut surface texture curve is obtained by rendering the cut surface texture curve and overlaying it onto the original cross-sectional image.

[0033] S104. Construct a multi-task joint prediction model based on IMAF-ConvNeXt, and output the prediction results of laser cutting speed and defocusing amount in the actual cutting process based on the original cross-sectional image, main texture image and physical prior image.

[0034] Understandably, unlike existing technologies that only input the cross-sectional image as two-dimensional visual data into the relevant prediction model, in this embodiment of the invention, a multi-task joint prediction model based on IMAF-ConvNeXt is constructed, and the original cross-sectional image, main texture image, and physical prior image are used as inputs to this multi-task joint prediction model. The extracted texture curve physical prior is explicitly injected into the network, guiding the model to focus on thermodynamic key regions. This overcomes the "attention drift" problem that traditional deep learning models are prone to in complex industrial contexts, enabling it to accurately output the prediction results of laser cutting speed and defocusing amount in the actual cutting process.

[0035] The following description, in conjunction with specific embodiments of the present invention, explains the acquisition process of the original cross-sectional image, the acquisition process of the main texture image, the acquisition process of the physical prior image, and the prediction process of the multi-task joint prediction model.

[0036] Furthermore, in some embodiments of the present invention, such as Figure 3 As shown, the original cross-sectional image containing the effective cut sections is obtained, including: S201 converts the cut section image into a grayscale image and uses bilateral filtering for edge-preserving and noise reduction processing.

[0037] It is understood that in this embodiment of the present invention, the acquired cross-sectional image is converted from a color image to a grayscale image, and bilateral filtering is used for edge-preserving and noise reduction processing, thereby smoothing out noise while preserving texture stripe edge information.

[0038] S202 uses an adaptive threshold to convert the grayscale image into a binary image, and uses elliptical structuring elements to perform morphological closing operations to fill the holes inside the cross-sectional area and connect the fracture contour.

[0039] It is understood that, in this embodiment of the invention, an adaptive threshold is used to convert the grayscale image into a binary image (e.g., neighborhood block size 51 pixels, constant offset 1), and morphological closing operations are performed using elliptical result elements (e.g., size 3x3 pixels) (e.g., iterated 15 times) to fill the holes inside the cross-sectional area and connect the fracture contour.

[0040] S203 extracts all external contours through contour detection and filters and retains effective cross-sectional areas based on contour area.

[0041] It is understood that in this embodiment of the present invention, all external contours are extracted by contour detection, and effective cross-sectional areas are selected and retained based on the contour area (for example, noise with an area less than 10% of the total image area is filtered out).

[0042] S204: Obtain the minimum circumscribed rectangle of the effective cross-sectional area, and perform reverse rotation correction on the original cross-sectional image according to the tilt angle of the workpiece to be cut, so that the longitudinal axis of the cross-section is aligned with the vertical axis of the image.

[0043] It is understood that in this embodiment of the present invention, the minimum circumscribed rectangle of the effective cross-sectional area is calculated, and after obtaining the tilt angle of the workpiece to be cut, the original cross-sectional image is reverse-rotated and corrected so that the longitudinal axis of the cross-section is strictly aligned with the vertical axis of the image.

[0044] S205 employs a high-coverage strip cutting strategy to crop the cross-sectional area, and draws the boundary contour and enclosing rectangle of the effective cut cross-section on the original cross-sectional image to form an original cross-sectional image containing the effective cut cross-section.

[0045] It is understood that in this embodiment of the present invention, after performing reverse rotation correction, the cross-sectional area is precisely cropped using a high-coverage strip cropping strategy (e.g., a row fill rate threshold of 99.5%). If the rotation filter has no effective row segments, it reverts to horizontal filtering (e.g., a threshold of 98%). Then, the boundary contours (e.g., green boundary contours) and enclosing rectangles (e.g., red precisely positioned enclosing rectangles) of the effective cut sections are drawn on the original cross-sectional image, thereby obtaining the following... Figure 4 The image shown contains the original cross-sectional image of the effective cut section.

[0046] Furthermore, in some embodiments of the present invention, such as Figure 5 As shown, the main texture image containing anisotropic intensity and tangential angle is obtained, including: S301, perform image denoising and enhancement on the original cross-sectional image.

[0047] It is understood that, in this embodiment of the present invention, the original cross-sectional image is sequentially subjected to median filtering (e.g., kernel size 5) to remove pulse salt-and-pepper noise, bilateral filtering (e.g., color standard deviation 25~40, spatial standard deviation) to preserve edges and denoise, vertical Gaussian filtering (e.g., kernel size 1×kv, kv=7~13) to directionally smooth and enhance stripe coherence, and finally supplemented with omnidirectional light Gaussian filtering to stabilize the underlying directional information, thereby achieving image denoising enhancement.

[0048] S302 uses the Sobel operator to calculate the horizontal and vertical gradients of the original cross-sectional image, and constructs a local structure tensor matrix based on the horizontal and vertical gradients.

[0049] It is understood that, in this embodiment of the present invention, the horizontal gradient of the image is calculated by employing the Sobel operator. and vertical gradient Construct the local structure tensor matrix Its mathematical expression is: ; in, The asymmetric elliptical Gaussian kernel differs from the traditional circular Gaussian kernel. The purpose of using an elliptical Gaussian kernel is to adapt to the arc-shaped characteristics of the cross-sectional fringes—the cross-sectional fringes are nearly horizontal in the top region and gradually bend in the middle and lower regions. The major axis of the elliptical Gaussian kernel is consistent with the direction of the fringes, so that the gradient information along the direction of the fringes will not be blurred during weighted smoothing, thus ensuring the accuracy of anisotropy measurement.

[0050] S303 performs eigenvalue decomposition on the local structure tensor matrix to obtain the principal and secondary eigenvalues, and obtains the anisotropy intensity and tangential angle based on the principal and secondary eigenvalues.

[0051] It is understood that, in this embodiment of the present invention, the eigenvalue decomposition of the local structure tensor matrix constructed in step S302 is performed, i.e., the characteristic equation is solved. To obtain the principal eigenvalues and secondary eigenvalues Furthermore, the anisotropy intensity is calculated based on these two eigenvalues. and tangential angle .

[0052] Anisotropy strength The calculation formula is: ; The value range of this indicator is [0, 1]. When When this occurs, it indicates the existence of a significant dominant gradient direction within the local neighborhood, meaning the texture fringes are clear and have good directional consistency, and when... When this occurs, it indicates that the gradient direction is dispersed or the texture is blurred in the local neighborhood, which may correspond to a smooth area or a slag-affected area on the cross-section.

[0053] Initial tangential angle The calculation formula is: ; in, The direction angle of the texture tangent is the direction of the molten metal flow at that pixel location.

[0054] Specifically, in the above embodiments of the present invention, such as Figure 6 As shown, the anisotropy intensity map displays the distribution of texture clarity within the cross-sectional area in the form of a heat map. Brighter colors indicate stronger anisotropy and clearer stripes. Meanwhile, as... Figure 7 As shown, the tangential guidance field matrix diagram displays the continuous direction of the texture tangent at each point within the cross-section using a color-coded method.

[0055] S304 maps the tangential angle to the biangular domain, constructs two continuous weighted vector component fields by combining an effective cross-section mask, and reconstructs a globally continuous, phase-free, singular pure tangential field from the two continuous weighted vector component fields using inverse trigonometric functions.

[0056] It should be understood that the tangential angle extracted in step S303 using the arctangent function Numerical calculations have limitations. For example, angle values ​​exhibit phase transitions at the boundary between 0 and π. Directly smoothing these angle values ​​with numerical smoothing results in severe directional distortion at the phase transition points, causing non-physical distortion of the texture curve at the corresponding locations. Therefore, in this embodiment of the invention, the tangential angle... Mapping to the biangular domain, i.e., computation and , making the angle The transitions at the 0 and π boundaries correspond to the same physical direction in the biangular domain, thus eliminating phase singularities. Furthermore, by combining the effective cross-sectional mask M, two continuous weighted vector component fields are constructed:

[0057] It should be noted that a mask M is introduced to ensure that only angle information within the effective cross-sectional area participates in subsequent calculations, and that random angle values ​​in the background and the area outside the kerf do not contaminate the vector field. Subsequently, Gaussian kernel G_s is used to perform mask normalization filtering on these two vector component fields. After filtering, the final global continuous phase-free singular pure tangential field is reconstructed from the smoothed vector components using an inverse trigonometric function. ; In this process, the angle values ​​of the two-angle domain are restored back to the original angle domain by multiplying by 1 / 2, and the resulting pure tangential field It is globally continuous and has no phase singularities, thus providing a high-quality guiding field for subsequent discrete Euler integral tracing.

[0058] S305 performs discrete Euler integral trajectory tracking on the pure tangential field and anisotropic intensity to form a main texture image containing anisotropic intensity and tangential angle.

[0059] It is understood that, in this embodiment of the present invention, obtaining a pure tangential field... After the anisotropic intensity field A, the pure tangential field is then analyzed. Discrete Euler integral trajectory tracking is performed on the anisotropic intensity field A to achieve one-dimensional extraction of the texture skeleton coordinate sequence.

[0060] Furthermore, in some embodiments of the present invention, the discrete Euler integral trajectory tracking of the pure tangential field and anisotropic intensity includes: setting a starting point at the central axis of the transverse section, performing a line-by-line downward stepping using discrete Euler integral, and defining the first... The horizontal coordinate state update equation for the row is: ,in, Let be the projection operator constrained by the cross-section mask, and let the core lateral step offset be defined as... Among them, dynamic weight function ,in, and It is a constant. Let be the local tangent angle of the texture at position i. Let be the anisotropy intensity of the texture at the i-th row position.

[0061] It is understood that, in this embodiment of the present invention, the starting point selection strategy for tracking is as follows: a starting constraint band with a width of approximately 20% of the cross-section is defined in the transverse central axis region of the cutting cross-section; the trajectory starting point is selected at the upper edge of the effective cross-section; and the macroscopic direction of the texture is tracked by progressively tracking downwards line by line using discrete Euler integrals. Horizontal coordinate The state update equation for the (i+1)th row is: ; Where i is the row index of the cross-sectional image from top to bottom; Let be the horizontal coordinate of the texture trajectory point in the i-th row; This represents the horizontal step offset from row i to row i+1; It is a projection operator constrained by the effective cross-sectional area mask M, used to constrain the calculated coordinates within the effective cutting section in real time and automatically remove out-of-bounds distortion points.

[0062] Core lateral step offset The calculation formula is: ; in, The local tangential angle of the texture at the i-th row position corresponds physically to the local direction of molten metal flow and is extracted from the reconstructed pure tangential angle field. The anisotropy intensity of the texture at the i-th row position is 0 to 1. The larger the value, the stronger the directionality of the stripes and the clearer the features. This is a dynamic weighting function used to adaptively adjust the step size based on texture sharpness; its analytical form is: ; in, The base step size coefficient, taken as an empirical value of 0.22, is used in cases with extremely weak textures. The minimum inertia required to maintain the longitudinal movement of the trajectory; The texture intensity gain coefficient is set to an empirical value of 0.78. It is used to amplify the horizontal deflection step size of the strong texture area, so that the trajectory strictly follows the direction of the real stripes.

[0063] Specifically, when the anisotropy intensity When this occurs, it indicates that the texture at that location is clear, the orientation is reliable, the weight is close to 1.0, the tracking algorithm strictly follows the guidance of the geometry guidance term, and when... When the texture at that location is blurred, the weight is reduced to 0.22, and the tracking algorithm automatically decays the lateral step, relying more on longitudinal inertia to maintain curve continuity. Thus, the dynamic weighting mechanism significantly enhances the algorithm's robustness under complex noise conditions.

[0064] Furthermore, in some embodiments of the present invention, such as Figure 8 As shown, the physical prior image containing the texture curve of the cut surface is obtained, including: S401 uses triple joint regularization to smooth the extracted cut surface texture curve, generating a sub-pixel continuous texture curve.

[0065] It is understandable that the initial discrete skeleton has pixel-level step-like artifacts due to the inherent limitations of discrete image grid sampling, which deviates from the actual molten metal flow trajectory. Therefore, in this embodiment of the invention, the extracted cut surface texture curve is smoothed by triple joint regularization to generate a sub-pixel continuous texture curve.

[0066] S402 renders and overlays the subpixel continuous texture curve onto the original cross-sectional image to form a physical prior image containing the texture curve of the cut section.

[0067] It is understood that in this embodiment of the invention, cubic spline interpolation with natural boundaries is used to densely resample discrete nodes to each integer row. The natural boundary condition requires that the second derivative of the interpolation curve be zero at both endpoints, ensuring that the interpolation curve extends as flatly as possible at the endpoints. Furthermore, through dense resampling, the curve achieves sub-pixel-level continuous mathematical approximation. Finally, the smoothed texture curve is rendered and superimposed onto the original cross-sectional image to form a physical prior image containing the texture curve of the cut cross-section (e.g., the texture curve of the cut cross-section of the aforementioned 10mm thick stainless steel material under four typical cutting qualities, as shown in the image). Figure 9 (As shown).

[0068] Furthermore, in some embodiments of the present invention, the extracted cut surface texture curve is smoothed by triple joint regularization, including: initially suppressing local curvature abrupt changes in the one-dimensional discrete skeleton by using weighted cubic smoothing splines, and introducing a Whittaker smoothing model containing a second-order difference penalty term, by minimizing the energy functional. To achieve global smoothness, where, Let be the energy function to be minimized, x be the observation coordinate vector of the initial discrete texture skeleton, and z be the coordinate vector after smooth reconstruction. For data fidelity items, It is a second-order difference operator matrix. For curvature penalty term, For adaptive curvature penalty weights, and by employing a one-dimensional Gaussian filter... High-frequency artifact micro-polishing is performed, and the final coordinate update equation is: ,in, Let be the final smoothed coordinates of the i-th point after triple regularization. Let i be the coordinates of the i-th point after Whittaker smoothing. The standard deviation is One-dimensional Gaussian smoothing kernel, This is the convolution operator. This is a factor that preserves the original features.

[0069] It is understood that in this embodiment of the present invention, the first step of the triple joint regularization uses weighted cubic smoothing splines to preprocess the initial discrete texture skeleton, thereby initially suppressing local curvature abrupt changes and pixel-level staircase artifacts.

[0070] The second layer of the triple joint regularization introduces a Whittaker smoothing model containing a second-order difference penalty term, achieving global smoothness by minimizing the energy functional, which is: ; in, Let be the energy function to be minimized; x is the observation coordinate vector of the initial discrete texture skeleton; z is the coordinate vector after smooth reconstruction. This is a data fidelity measure that assesses the degree of deviation between the smoothed curve and the original observation points, ensuring that the smoothed curve does not deviate from the overall shape of the original texture. It is a second-order difference operator matrix used to calculate the curvature change of the curve; This is a curvature penalty term, which measures the overall degree of fluctuation of the curve; An adaptive curvature penalty weight is used to balance data fidelity and global smoothness.

[0071] The physical meaning of this energy functional is as follows: the first term measures the degree of fit between the reconstructed curve and the original observation data; the second term measures the overall curvature of the reconstructed curve, which is adjusted... The value of can achieve the optimal balance between fitting accuracy and smoothness.

[0072] A closed-form analytical solution is obtained by solving this convex optimization problem: ; Where I is the identity matrix, with dimensions consistent with the length of the coordinate vector; The second-order difference operator matrix The transpose of .

[0073] The advantage of this closed-form solution is that the globally optimal solution can be obtained with only one matrix inversion operation.

[0074] The third step of the triple joint regularization uses a one-dimensional Gaussian filter to perform microscopic polishing on the globally smoothed curve, eliminating high-frequency micro-fluctuations left over from numerical calculations. The final coordinate update equation is: ; in, The final smoothed coordinates of the i-th point after triple regularization; Let i be the coordinates of the i-th point after Whittaker smoothing; The standard deviation is A one-dimensional Gaussian smoothing kernel; This is the convolution operator; This is the original feature preservation factor, with a value between 0 and 1, used to balance the effects of macroscopic contour preservation and microscopic noise filtering.

[0075] Specifically, in the above embodiments of the present invention, the processing flow for obtaining a physical prior image containing the texture curve of the cut surface is as follows: cross-section region localization → image filtering and enhancement → structural tensor anisotropy analysis → discrete Euler integral trajectory tracking → cascaded regularization smoothing → texture curve output. Among them, the texture stripes of the cut surface are the physical trajectories left after the molten metal flows and solidifies rapidly. Accurately extracting its geometric skeleton is the key to transforming the micro-morphology into a quantifiable physical prior.

[0076] Furthermore, in some embodiments of the present invention, such as Figure 10 As shown, the multi-task joint prediction model, based on the original cross-sectional image, main texture image, and physical prior image, outputs the predicted results of laser cutting speed and defocusing amount during the actual cutting process, including: S501 uses a multi-task joint prediction model to perform shallow feature extraction and local feature refinement on the original cross-sectional image, main texture image, and physical prior image, respectively.

[0077] It is understood that, in this embodiment of the present invention, the acquired original cross-sectional image is denoted as I_ori. This image retains low-level visual information such as cross-sectional micro-roughness, color distribution, and local contrast, and the extracted main texture image (i.e., anisotropic intensity) is used. With tangential angle The combined encoding diagram, specifically the encoding method is to use... Mapped to brightness, The image mapped to hue is denoted as I_tex, which explicitly encodes the anisotropic characteristics of the molten metal flow, and the physical prior image is denoted as I_cur, which provides explicit global spatial guidance.

[0078] Optionally, the image resolution of the three modalities is uniformly adjusted to 224×224 pixels, and a random cropping strategy is adopted after scaling while maintaining the aspect ratio to avoid non-physical deformation of the texture bending angle caused by forced square scaling. At the same time, adaptive histogram equalization is used to enhance the contrast of the cutting stripes.

[0079] Furthermore, by feeding the images of the three modalities into three independent branches, each branch performs a 1×1 convolution with a stride of 1 to map the input from 3 channels to 48 channels of feature maps. After batch normalization and ReLU activation, the feature maps of each branch are fed into a local spatial-channel collaborative attention module (lightweight CBAM) for feature refinement. Since the interference sources in the cut section image (metal mirror reflection, bottom slag occlusion) have local spatial concentration, the attention module adopts a cascaded architecture of "channel priority, space secondary" to identify and suppress these regions in the spatial dimension.

[0080] The computation process of the Local Spatial-Channel Collaborative Attention Module (Lightweight CBAM) is as follows: For the input features... Global adaptive average pooling is performed to compress the spatial information of each channel into a scalar, which is then passed through a 1×1 bottleneck convolution and a sigmoid activation to generate a channel attention map. : ; in, and 1×1 convolutions for dimensionality reduction and dimensionality enhancement, respectively. For ReLU, It is Sigmoid.

[0081] The local features after channel refinement are denoted as ( (This indicates element-wise multiplication).

[0082] The spatial attention submodule performs mean pooling and max pooling on F' along the channel dimension and concatenates them, then generates a spatial attention map through a 7×7 large kernel convolution. : ; Final single-branch output It effectively suppresses local noise such as metal reflection and slag buildup.

[0083] S502 performs industrial multimodal alignment and fusion on the refined local features to obtain fused features. The fused features include the micro-roughness information of the original cross-sectional image, the texture anisotropic physical field information, and the explicit spatial guidance information of the texture curve.

[0084] Understandably, the local features refined from the three branches are multiplied by the modal weights learned through adaptive learning. Rear edge channel splicing.

[0085] Optionally, the three modality weights are scalar parameters that are automatically learned during network training. The initial value of each is set to 1.0, and the network automatically adjusts the contribution ratio of each modality as training progresses.

[0086] After splicing, the output is reduced to a standard 3-channel output via a 1×1 fusion convolutional layer: ; This fusion feature At the same time, it preserves the micro-roughness information, texture anisotropic physical field information, and explicit spatial guidance information of the texture curve of the original image.

[0087] S503 inputs the fused features into the ConvNeXt backbone network, enabling the ConvNeXt backbone network to output a high-dimensional feature map.

[0088] It is understood that, in this embodiment of the present invention, the fusion feature is... The data is fed into the ConvNeXt backbone network, which systematically modernizes the standard ResNet architecture (e.g., introducing 7×7 deep separable convolutions to expand the receptive field, adopting an inverted bottleneck structure, using GELU activation function instead of ReLU, and replacing Batch Normalization with LayerNormalization). This achieves performance comparable to the Swin Transformer without introducing a self-attention mechanism. Furthermore, the backbone network performs multi-level abstraction on the fused features (containing four stages, each stage containing several ConvNeXt modules), progressively reducing spatial resolution and increasing channel dimension, ultimately outputting a high-dimensional feature map. .

[0089] S504 performs feature refinement deep convolution on high-dimensional feature maps to obtain dimensionality-reduced features.

[0090] It is understandable that high-dimensional feature maps By sequentially performing 3×3 grouped depthwise convolutions with a stride of 1 and padding of 1, and then 1×1 pointwise convolutions, the number of channels is strictly halved (from C' to C' / 2), thus achieving feature refinement depthwise convolution and obtaining dimensionality-reduced features. The grouped depthwise convolution divides the input channels into several groups for independent convolution operations, significantly reducing the number of parameters. The feature independence between groups helps capture diverse feature patterns. Additionally, 1×1 point convolution is used for cross-channel information fusion, and GELU activation introduces nonlinearity. Thus, the large-kernel grouped depthwise convolution further expands the global receptive field, capturing the macroscopic morphological trends of the cross-section and suppressing local optical artifacts.

[0091] S505 inputs the dimensionality-reduced features into the global attention module and outputs a globally shared feature vector.

[0092] It is understood that, in this embodiment of the present invention, the dimensionality reduction feature is... The data is fed into the global attention module, which is fundamentally different from the aforementioned local spatial-channel collaborative attention module (lightweight CBAM): the local spatial-channel collaborative attention module (lightweight CBAM) works at the network front end to process low-level features, and its main goal is to filter out optical noise; this module works in a deep, high-dimensional abstract feature space, and the goal of the attention mechanism is to accurately locate the core semantic clusters that determine the segmentation quality in a highly compressed semantic space.

[0093] First, calculate the global channel attention weights along the channel dimension: ; Obtain channel refinement characteristics .

[0094] Subsequently Perform mean pooling and max pooling along the channel dimension and concatenate them. Then, generate a spatial attention mask through a 7×7 large-kernel spatial convolution. ; The final output is a globally shared feature. This precisely guides the model to focus on the core semantic regions that determine the quality of the cuts before the final global average pooling.

[0095] S506 inputs the globally shared feature vector into two heterogeneous prediction branches to obtain the prediction results of laser cutting speed and defocusing amount, respectively.

[0096] It is understood that, in this embodiment of the present invention, the global shared feature vector... The feature vector is reduced to a global one-dimensional shared feature vector through global adaptive average pooling. It is then passed through a [C' / 2,256] mapping layer, a Dropout layer with a dropout rate of 0.3, and ReLU activation in sequence, and then fed into two heterogeneous prediction branches in parallel to obtain the prediction results of laser cutting speed and defocusing amount, respectively.

[0097] Optionally, the cutting speed classification branch uses a fully connected layer [256, N_speed] to output the logits of the speed category (N_speed=12, corresponding to 12 levels from 5 to 16 m / min); the defocus amount classification branch uses a fully connected layer [256, N_defocus] to output the logits of the defocus amount category (N_defocus=14, corresponding to 14 levels from +4 to -9 mm). During forward inference, the model applies the Argmax operator to the classification logits to extract the class label corresponding to the maximum posterior probability.

[0098] Meanwhile, Focal Loss is used as the classification loss function, which adds a modulation factor to the standard cross-entropy loss. ,in This represents the model's predicted probability for the correct class. When a sample is correctly classified and the probability is high, the modulation factor is close to 0, contributing little to the loss; when a sample is misclassified, the modulation factor is close to 1, contributing significantly to the loss, causing the model to focus more on difficult-to-classify samples. The overall objective function is a weighted sum of the classification losses for velocity and defocus: ; in, =8.25, =1.95. The AdamW optimizer (initial learning rate 2×10⁻⁴, weight decay 0.01), cosine annealing learning rate scheduling, training for 100 epochs, batch size 4.

[0099] Specifically, in the above embodiments of the present invention, the prediction process of the multi-task joint prediction model is as follows: multimodal input construction → shallow feature extraction and attention refinement → industrial multimodal alignment and fusion → ConvNeXt backbone network deep feature extraction → feature refinement deep convolution → global deep semantic focusing attention → multi-task joint prediction → process parameter output. Thus, the extracted texture curve physical prior is explicitly injected into the network, guiding the model to focus on thermodynamically critical regions and overcoming the "attention drift" problem that traditional deep learning models are prone to in complex industrial contexts.

[0100] Furthermore, in some embodiments of the present invention, the method further includes: performing anti-slag truncation processing on the original cross-sectional image based on a local median statistical model. The anti-slag truncation processing includes: calculating the absolute horizontal first-order difference of the discrete tracking trajectory; when a continuous horizontal coordinate mutation is detected to exceed the adaptive reference threshold, it is determined that the tracking trajectory has entered the poor slag discharge interference zone at the bottom of the specimen, and the truncation mechanism is actively activated to remove the slag-covered part.

[0101] It should be understood that during the laser cutting of thick plates, slag often forms at the bottom due to poor slag removal, and the texture of the slag-covered area is completely different from the physical wire-drawing trend of the main cross-section texture. Therefore, in this embodiment of the invention, anti-slag-covering truncation processing based on a local median statistical model is also performed on the original cross-sectional image: firstly, the absolute value sequence of the first-order difference of the trajectory is calculated. Data from the 65% stable region in the upper half of the trajectory is extracted as a statistical benchmark. The median (med) of this region is calculated, and then the cutoff threshold is dynamically calculated. The system scans backward from the tail of the curve trajectory. Once the absolute value of the first-order difference of the horizontal coordinates of multiple consecutive rows exceeds the adaptive threshold, the system determines that the trajectory has entered the bottom slag distortion blind zone. It then actively triggers the adaptive tail truncation mechanism, truncating the tail of the curve at the row that first exceeds the threshold. In this way, erroneous data contaminated by slag deformation is completely removed, achieving active annealing and precise truncation of the high-deformation noise zone at the bottom.

[0102] In summary, the high-power laser cutting parameter inversion method according to embodiments of the present invention acquires a cross-sectional image of the workpiece to be cut using an optical image acquisition platform, preprocesses and locates the effective region of the cross-sectional image to obtain an original cross-sectional image containing the effective cross-section, extracts the cross-sectional texture curve from the original cross-sectional image to obtain a main texture image containing anisotropic intensity and tangential angle, and superimposes the extracted cross-sectional texture curve onto the original cross-sectional image to obtain a physical prior image containing the cross-sectional texture curve. Then, a multi-task joint prediction model based on IMAF-ConvNeXt is constructed, and the multi-task joint prediction model outputs the predicted results of laser cutting speed and defocusing amount during the actual cutting process based on the original cross-sectional image, main texture image, and physical prior image. Thus, by using explicit physical prior to guide a deep neural network, accurate inversion of high-power laser cutting speed and defocusing amount is achieved.

[0103] Based on the cutting parameter inversion method for high-power laser cutting according to the foregoing embodiments of the present invention, the present invention also proposes a computer-readable storage medium storing a cutting parameter inversion program for high-power laser cutting. When the cutting parameter inversion program for high-power laser cutting is executed by a processor, it implements the cutting parameter inversion method for high-power laser cutting according to the foregoing embodiments of the present invention.

[0104] It should be understood that the specific implementation of the computer-readable storage medium in the embodiments of the present invention can be found in the specific implementation of the cutting parameter inversion method for high-power laser cutting in the foregoing embodiments of the present invention. To reduce redundancy, it will not be repeated here.

[0105] In summary, according to the computer-readable storage medium of the present invention, by executing the cutting parameter inversion program of high-power laser cutting stored thereon, it is possible to use explicit physical priors to guide a deep neural network to achieve accurate inversion of high-power laser cutting speed and defocusing amount.

[0106] Figure 11 This is a block diagram of a cutting parameter inversion device 100 for high-power laser cutting according to an embodiment of the present invention.

[0107] Specifically, in some embodiments of the present invention, such as Figure 11 As shown, the cutting parameter inversion device 100 for high-power laser cutting includes: a first acquisition module 10, a second acquisition module 20, a third acquisition module 30, and a parameter inversion module 40.

[0108] The first acquisition module 10 is used to acquire the cut section image of the workpiece to be cut through the optical image acquisition platform, and to preprocess and locate the effective area of ​​the cut section image to obtain the original cross-sectional image containing the effective cut section; the second acquisition module 20 is used to extract the cut section texture curve from the original cross-sectional image to obtain the main texture image containing the anisotropic intensity and tangential angle; the third acquisition module 30 is used to superimpose the extracted cut section texture curve onto the original cross-sectional image to obtain the physical prior image containing the cut section texture curve; the parameter inversion module 40 is used to construct a multi-task joint prediction model based on IMAF-ConvNeXt, and to output the prediction results of the laser cutting speed and defocusing amount in the actual cutting process based on the original cross-sectional image, the main texture image and the physical prior image through the multi-task joint prediction model.

[0109] Further, in some embodiments of the present invention, the first acquisition module 10 is used to: convert the cut section image into a grayscale image and perform edge-preserving and noise reduction processing using bilateral filtering; convert the grayscale image into a binary image using an adaptive threshold and perform morphological closing operation using elliptical structuring elements to fill the internal holes of the section area and connect the fracture contour; extract all external contours through contour detection and filter and retain the effective section area according to the contour area; obtain the minimum circumscribed rotation rectangle of the effective section area and perform reverse rotation correction on the original section image according to the tilt angle of the workpiece to be cut, so that the longitudinal axis of the section is aligned with the vertical axis of the image; crop the section area using a high-coverage strip cropping strategy, and draw the boundary contour and enclosing rectangle of the effective cut section on the original section image to form an original section image containing the effective cut section.

[0110] Furthermore, in some embodiments of the present invention, the second acquisition module 20 is further configured to: perform image denoising enhancement on the original cross-sectional image; calculate the horizontal and vertical gradients of the original cross-sectional image using the Sobel operator, and construct a local structure tensor matrix based on the horizontal and vertical gradients; perform eigenvalue decomposition on the local structure tensor matrix to obtain principal and secondary eigenvalues, and obtain anisotropic intensity and tangential angle based on the principal and secondary eigenvalues; map the tangential angle to the biangular domain, construct two continuous weighted vector component fields in combination with an effective cross-sectional mask, and reconstruct a globally continuous, phase-free, and singular pure tangential field from the two continuous weighted vector component fields using inverse trigonometric functions; perform discrete Euler integral trajectory tracking on the pure tangential field and anisotropic intensity to form a main texture image containing anisotropic intensity and tangential angle.

[0111] Furthermore, in some embodiments of the present invention, the second acquisition module 20 is further configured to set a starting point at the central axis of the transverse section, perform a row-by-row downward stepping using discrete Euler integrals, and define the first... The horizontal coordinate state update equation for the row is: ,in, Let be the projection operator constrained by the cross-section mask, and let the core lateral step offset be defined as... Among them, dynamic weight function ,in, and It is a constant. Let be the local tangent angle of the texture at position i. Let be the anisotropy intensity of the texture at the i-th row position.

[0112] Furthermore, in some embodiments of the present invention, the third acquisition module 30 is also used to: smooth the extracted cross-sectional texture curve through triple joint regularization to generate a sub-pixel continuous texture curve; render the sub-pixel continuous texture curve and superimpose it onto the original cross-sectional image to form a physical prior image containing the cross-sectional texture curve.

[0113] Furthermore, in some embodiments of the present invention, the third acquisition module 30 is also used to: initially suppress local curvature abrupt changes in the one-dimensional discrete skeleton by employing weighted cubic smoothing splines, and introduce a Whittaker smoothing model containing a second-order difference penalty term, by minimizing the energy functional To achieve global smoothness, where, Let be the energy function to be minimized, x be the observed coordinate vector of the initial discrete texture skeleton, and z be the coordinate vector after smooth reconstruction. For data fidelity items, It is a second-order difference operator matrix. For curvature penalty term, For adaptive curvature penalty weights, and by employing a one-dimensional Gaussian filter... High-frequency artifact micro-polishing is performed, and the final coordinate update equation is: ,in, Let be the final smoothed coordinates of the i-th point after triple regularization. Let i be the coordinates of the i-th point after Whittaker smoothing. The standard deviation is One-dimensional Gaussian smoothing kernel, This is the convolution operator. This is a factor that preserves the original features.

[0114] Furthermore, in some embodiments of the present invention, the parameter inversion module 40 is also used to: perform shallow feature extraction and local feature refinement on the original cross-sectional image, main texture image, and physical prior image respectively through a multi-task joint prediction model; perform industrial multimodal alignment and fusion on the refined local features to obtain fused features, which include micro-roughness information of the original cross-sectional image, texture anisotropic physical field information, and explicit spatial guidance information of the texture curve; input the fused features into the ConvNeXt backbone network to make the ConvNeXt backbone network output a high-dimensional feature map; perform feature refinement deep convolution on the high-dimensional feature map to obtain dimensionality reduction features; input the dimensionality reduction features into the global attention module to output a globally shared feature vector; and input the globally shared feature vector into two heterogeneous prediction branches to obtain the laser cutting speed and defocus prediction results respectively.

[0115] Furthermore, in some embodiments of the present invention, the third acquisition module 30 is also used to: perform anti-slag truncation processing on the original cross-sectional image based on a local median statistical model. The anti-slag truncation processing includes: calculating the absolute horizontal first-order difference of the discrete tracking trajectory; when a continuous horizontal coordinate mutation is detected to exceed the adaptive reference threshold, it is determined that the tracking trajectory has entered the poor slag discharge interference area at the bottom of the specimen, and the truncation mechanism is actively activated to remove the slag-covered part.

[0116] It should be understood that the specific implementation of the cutting parameter inversion device 100 for high-power laser cutting in the embodiments of the present invention can be found in the specific implementation of the cutting parameter inversion method for high-power laser cutting in the foregoing embodiments of the present invention. To reduce redundancy, it will not be described again here.

[0117] In summary, the high-power laser cutting parameter inversion device according to an embodiment of the present invention acquires a cross-sectional image of the workpiece to be cut through an optical image acquisition platform via a first acquisition module, and preprocesses and locates the effective area of ​​the cross-sectional image to obtain an original cross-sectional image containing the effective cross-section. A second acquisition module extracts the cross-sectional texture curve from the original cross-sectional image to obtain a main texture image containing anisotropic intensity and tangential angle. A third acquisition module overlays the extracted cross-sectional texture curve onto the original cross-sectional image to obtain a physical prior image containing the cross-sectional texture curve. Then, a parameter inversion module constructs a multi-task joint prediction model based on IMAF-ConvNeXt, and the multi-task joint prediction model outputs the predicted laser cutting speed and defocusing amount during the actual cutting process based on the original cross-sectional image, the main texture image, and the physical prior image. Thus, by using explicit physical prior to guide a deep neural network, accurate inversion of high-power laser cutting speed and defocusing amount is achieved.

[0118] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0119] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0120] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0121] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0122] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0123] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0124] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0125] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for inverting cutting parameters in high-power laser cutting, characterized in that, The method includes: The cutting section image of the workpiece to be cut is acquired through an optical image acquisition platform, and the cutting section image is preprocessed and the effective area is located to obtain the original cross-sectional image containing the effective cutting section. The original cross-sectional image is subjected to cross-sectional texture curve extraction to obtain a main texture image containing anisotropic intensity and tangential angle; The extracted cross-sectional texture curves are superimposed onto the original cross-sectional image to obtain a physical prior image containing the cross-sectional texture curves. A multi-task joint prediction model based on IMAF-ConvNeXt is constructed, and the prediction results of laser cutting speed and defocusing amount in the actual cutting process are output by the multi-task joint prediction model based on the original cross-sectional image, the main texture image and the physical prior image.

2. The cutting parameter inversion method for high-power laser cutting according to claim 1, characterized in that, The process of obtaining the original cross-sectional image containing the effective cut surface includes: The cut section image is converted into a grayscale image, and edge-preserving and noise-reducing processing is performed using bilateral filtering; An adaptive threshold is used to convert the grayscale image into a binary image, and morphological closing operations are performed using elliptical structuring elements to fill the internal holes of the cross-sectional area and connect the fracture contour. All external contours are extracted through contour detection, and valid cross-sectional areas are retained based on contour area. Obtain the minimum circumscribed rectangle of the effective cross-sectional area, and perform reverse rotation correction on the original cross-sectional image according to the tilt angle of the workpiece to be cut, so that the longitudinal axis of the cross-section is aligned with the vertical axis of the image. A high-coverage strip cropping strategy is used to crop the cross-sectional area, and the boundary contour and enclosing rectangle of the effective cut cross-section are drawn on the original cross-sectional image to form the original cross-sectional image containing the effective cut cross-section.

3. The cutting parameter inversion method for high-power laser cutting according to claim 1, characterized in that, The acquisition of the main texture image containing anisotropic intensity and tangential angle includes: The original cross-sectional image is then subjected to image denoising and enhancement. The Sobel operator is used to calculate the horizontal and vertical gradients of the original cross-sectional image, and a local structure tensor matrix is ​​constructed based on the horizontal and vertical gradients. The local structure tensor matrix is ​​decomposed into eigenvalues ​​to obtain principal eigenvalues ​​and secondary eigenvalues. Based on the principal eigenvalues ​​and secondary eigenvalues, the anisotropy intensity and the tangential angle are obtained. The tangential angle is mapped to the biangular domain, and two continuous weighted vector component fields are constructed by combining the effective cross-section mask. A globally continuous, phase-free, and singular pure tangential field is reconstructed from the two continuous weighted vector component fields using inverse trigonometric functions. Discrete Euler integral trajectory tracking is performed on the pure tangential field and the anisotropic intensity to form the main texture image containing the anisotropic intensity and the tangential angle.

4. The cutting parameter inversion method for high-power laser cutting according to claim 3, characterized in that, The discrete Euler integral trajectory tracking of the pure tangential field and the anisotropic intensity includes: A starting point is set at the center axis of the transverse section, and discrete Euler integrals are used to proceed downwards row by row, defining the first... The horizontal coordinate state update equation for the row is: ,in, Let be the projection operator constrained by the cross-section mask, and let the core lateral step offset be defined as... Among them, dynamic weight function ,in, and It is a constant. Let be the local tangent angle of the texture at position i. Let be the anisotropy intensity of the texture at the i-th row position.

5. The cutting parameter inversion method for high-power laser cutting according to claim 1, characterized in that, The process of obtaining a physical prior image containing the texture curve of the cut surface includes: The extracted cut surface texture curve is smoothed by triple joint regularization to generate a sub-pixel continuous texture curve. The subpixel continuous texture curve is rendered and superimposed onto the original cross-sectional image to form the physical prior image containing the cut-section texture curve.

6. The cutting parameter inversion method for high-power laser cutting according to claim 5, characterized in that, The smoothing process of the extracted cut surface texture curve through triple joint regularization includes: initially suppressing local curvature abrupt changes in the one-dimensional discrete skeleton by using weighted cubic smoothing splines, introducing a Whittaker smoothing model containing a second-order difference penalty term, and minimizing the energy functional. To achieve global smoothness, where, Let be the energy function to be minimized, x be the observed coordinate vector of the initial discrete texture skeleton, and z be the coordinate vector after smooth reconstruction. For data fidelity items, It is a second-order difference operator matrix. For curvature penalty term, For adaptive curvature penalty weights, and by employing a one-dimensional Gaussian filter... High-frequency artifact micro-polishing is performed, and the final coordinate update equation is: ,in, Let be the final smoothed coordinates of the i-th point after triple regularization. Let i be the coordinates of the i-th point after Whittaker smoothing. The standard deviation is One-dimensional Gaussian smoothing kernel, This is the convolution operator. This is a factor that preserves the original features.

7. The cutting parameter inversion method for high-power laser cutting according to claim 1, characterized in that, The step of using the multi-task joint prediction model to output the predicted results of laser cutting speed and defocusing amount in the actual cutting process based on the original cross-sectional image, the main texture image, and the physical prior image includes: The multi-task joint prediction model is used to perform shallow feature extraction and local feature refinement on the original cross-sectional image, the main texture image, and the physical prior image, respectively. The refined local features are aligned and fused using industrial multimodal methods to obtain fused features. The fused features include micro-roughness information of the original cross-sectional image, texture anisotropic physical field information, and explicit spatial guidance information of the texture curve. The fused features are input into the ConvNeXt backbone network, which outputs a high-dimensional feature map. Perform feature refinement deep convolution on the high-dimensional feature map to obtain dimensionality-reduced features; The dimensionality-reduced features are input into the global attention module, which outputs a globally shared feature vector. The globally shared feature vector is input into the cutting speed classification branch and the defocusing amount classification branch to obtain the laser cutting speed and the defocusing amount prediction results, respectively. The cutting speed classification branch uses a fully connected layer [256, N_speed] to output the logits of the speed category, where N_speed=12, corresponding to 12 levels from 5 to 16 m / min. The defocusing amount classification branch uses a fully connected layer [256, N_defocus] to output the logits of the defocusing amount category, where N_defocus=14, corresponding to 14 levels from +4 to -9 mm.

8. The cutting parameter inversion method for high-power laser cutting according to claim 3, characterized in that, The method further includes: performing anti-slag truncation processing on the original cross-sectional image based on a local median statistical model. The anti-slag truncation processing includes: calculating the trajectory horizontal first-order difference absolute value sequence of the tracking trajectory obtained by tracking the discrete Euler integral trajectory; when a continuous horizontal coordinate mutation exceeding the adaptive benchmark threshold is detected according to the trajectory horizontal first-order difference absolute value sequence of the tracking trajectory, it is determined that the tracking trajectory has entered the poor slag discharge interference zone at the bottom of the specimen, and the truncation mechanism is actively activated to remove the slag-attached part.

9. A computer-readable storage medium, characterized in that, It stores a cutting parameter inversion program for high-power laser cutting, which, when executed by a processor, implements the cutting parameter inversion method for high-power laser cutting as described in any one of claims 1-8.

10. A cutting parameter inversion device for high-power laser cutting, characterized in that, The device includes: The first acquisition module is used to acquire the cut cross-section image of the workpiece to be cut through the optical image acquisition platform, and to preprocess and locate the effective area of ​​the cut cross-section image to obtain the original cross-sectional image containing the effective cut cross-section. The second acquisition module is used to extract the cross-sectional texture curve of the original cross-sectional image to obtain a main texture image containing anisotropic intensity and tangential angle; The third acquisition module is used to superimpose the extracted cross-sectional texture curve onto the original cross-sectional image to obtain a physical prior image containing the cross-sectional texture curve. The parameter inversion module is used to construct a multi-task joint prediction model based on IMAF-ConvNeXt, and output the prediction results of laser cutting speed and defocusing amount in the actual cutting process based on the original cross-sectional image, the main texture image and the physical prior image.

Images