A laser cutting method and a laser cutting apparatus

Abstract

The application relates to the technical field of laser welding, in particular to a laser cutting method and a laser cutting device, which comprises the following steps: obtaining a two-dimensional surface feature map and a three-dimensional heat distribution map of cutting material, performing path planning on a cutting area according to an initial cutting parameter set, controlling a cutting mechanism to perform cutting processing according to a cutting path set, collecting a multi-dimensional sequence data stream in the cutting process, calculating a deviation value of a cutting-through probability and a preset threshold value, optimizing the initial cutting parameter set based on the deviation value, combining the cutting path set to perform secondary cutting processing, generating a second state vector, adjusting the cutting path set according to the similarity of the second state vector and a first state vector, and cutting all areas until the cutting is completed. The application can reduce the need for secondary polishing or rework, improve production efficiency, reduce the degree of manual intervention, greatly improve the automation and intelligent level of the cutting process, and significantly improve the laser cutting precision.

Description

Technical Field

[0001] This invention relates to the field of laser cutting technology, and more specifically, to a laser cutting method and laser cutting equipment. Background Technology

[0002] Laser cutting, with its advantages of high speed, non-contact operation, and high precision, has become a core manufacturing method in sheet metal processing, aerospace, and microelectronics packaging. Currently, the mainstream operating mode of equipment is typically as follows: the operator relies on material manuals or experience databases to set process parameters such as power, pulse width, feed rate, focal point position, and auxiliary gas pressure before operation; the control system generates a fixed cutting trajectory based on 2D CAD drawings and maintains this trajectory throughout the entire processing. However, for modern composite materials with large thickness gradients, complex structures, or heat sensitivity, this offline control method of preset parameters + static trajectory is gradually revealing its limitations: the quality of the cut is extremely sensitive to environmental disturbances. Even slight fluctuations in material batches, surface flatness, or external temperature can make it difficult to maintain the cutting depth and heat-affected zone within the ideal range, leading to a significant increase in secondary grinding or rework rates.

[0003] To address these shortcomings, the industry has successively introduced online monitoring methods such as slag illumination, plasma radiation, and temperature point measurement, and attempted to trigger shutdown alarms or reduce feed rates through simple threshold judgments. Some solutions also segment large contours to reduce heat accumulation. However, these improvements generally suffer from low information utilization and a loose control chain: various sensor signals only participate in adjustment locally or after the fact, making it difficult to link with core process parameters such as energy distribution and motion trajectory in real time; threshold settings rely on experience, and once the processing object changes, re-trial cutting and calibration are often required. As a result, the system still cannot quickly and stably obtain ideal cut quality under complex geometry or variable working conditions, and the processing cycle is subject to a high degree of manual intervention.

[0004] Therefore, a new laser cutting method is urgently needed to solve the above problems. Summary of the Invention

[0005] The main objective of this invention is to provide a laser cutting method and laser cutting equipment, aiming to overcome the technical problem that existing technologies cannot quickly and stably obtain ideal cut quality under complex geometries or variable working conditions.

[0006] To address the aforementioned problems, this invention proposes a laser cutting method, the method comprising:

[0007] Obtain a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut, and extract the high-dimensional feature vectors of the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain the initial cutting parameter set;

[0008] Based on the initial cutting parameter set, a path planning process is performed on the cutting area to obtain a cutting path set;

[0009] The cutting mechanism is controlled to perform cutting processing according to the cutting path set, and a multi-dimensional sequence data stream is collected during the cutting process to generate a first state vector;

[0010] The deviation between the cut probability and the preset threshold is calculated based on the first state vector. The initial cutting parameter set is then optimized based on the deviation to generate an optimized parameter set.

[0011] The optimized parameter set is combined with the cutting path set for secondary cutting to generate a second state vector. The cutting path set is then adjusted based on the similarity between the second state vector and the first state vector until all regions are cut.

[0012] Further, the step of obtaining a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut, and extracting high-dimensional feature vectors from the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain an initial cutting parameter set includes:

[0013] Collect surface texture, reflectivity information and three-dimensional thermal distribution data of the material surface to generate a raw multidimensional dataset containing two-dimensional surface feature maps and three-dimensional thermal distribution maps;

[0014] Extract high-dimensional feature vectors from the original multidimensional dataset to obtain a feature vector matrix, wherein the high-dimensional feature vectors include at least surface roughness, thermal conductivity gradient, material thickness and environmental variables;

[0015] The dimensionality reduction process is performed on the eigenvector matrix. During the dimensionality reduction process, orthogonality constraints are applied to the eigenvector matrix to generate a dimensionality-reduced eigenvector matrix.

[0016] Based on the reduced feature matrix, parameter modeling is performed, and training is conducted using the radial basis function kernel as the kernel function to generate an initial cutting parameter set.

[0017] Further, the step of performing path planning on the cutting region based on the initial cutting parameter set to obtain a cutting path set includes:

[0018] The initial cutting parameter set is processed by feature mapping to obtain the region feature vector set;

[0019] The segmented region is partitioned based on the regional feature vector set to obtain the sub-region parameter set;

[0020] The sub-region parameter set is initialized with paths, and global optimization is performed based on a genetic algorithm to adjust the path direction and order, thus obtaining an initial path set.

[0021] The cutting order is scheduled based on the initial path set to obtain the cutting path set.

[0022] Further, the step of controlling the cutting mechanism to perform cutting processing according to the cutting path set, and collecting multi-dimensional sequence data streams during the cutting process to generate a first state vector includes:

[0023] The cutting machine is controlled to perform cutting operations along the cutting path set, and the signal set generated during the cutting process is collected by multiple sensors to generate a raw multidimensional data stream.

[0024] The original multidimensional data stream is subjected to noise reduction and feature enhancement processing to obtain an optimized data stream;

[0025] High-dimensional features of the optimized data stream are extracted based on a deep learning feature extraction network to generate a feature vector set.

[0026] The feature vector set is used for state prediction, and the multiple prediction results are weighted and fused to obtain the first state vector.

[0027] Further, the step of calculating the deviation between the cutting probability and a preset threshold based on the first state vector, and optimizing the initial cutting parameter set based on the deviation to generate an optimized parameter set includes:

[0028] Based on a pre-trained deep neural network model, the first state vector is transformed into a probability distribution space, and the deviation between the current cut-through probability and the preset threshold is calculated.

[0029] Based on the deviation value, the initial cutting parameter set is initially adjusted to obtain the parameter adjustment vector;

[0030] The initial cutting parameter set is optimized using reinforcement learning based on the parameter adjustment vector to obtain an intermediate optimized parameter set.

[0031] The features of the intermediate optimization parameter set are weighted and fused with the spatiotemporal features of the first state vector to obtain a fused feature vector;

[0032] The intermediate optimization parameter set is subjected to a global search based on the fused feature vector to obtain the optimization parameter set;

[0033] The optimized parameter set and the cutting path set are adjusted together to obtain the adjusted cutting path set.

[0034] Further, the step of combining the optimized parameter set with the cutting path set for secondary cutting processing to generate a second state vector, and adjusting the cutting path set according to the similarity between the second state vector and the first state vector until all regions are cut, includes:

[0035] The optimized parameter set is nonlinearly mapped to the cutting path set to generate an initial fusion parameter set;

[0036] The initial fusion parameter set is dynamically adjusted in real time to obtain a dynamic control command set;

[0037] The cutting mechanism is driven to perform secondary cutting according to the dynamic control instruction set, and the secondary cutting data stream during the secondary cutting process is collected to generate a second state vector.

[0038] The path adjustment trigger signal is obtained by performing a similarity analysis between the second state vector and the first state vector.

[0039] The cutting path set is locally optimized based on the path adjustment trigger signal to obtain an optimized path set.

[0040] By combining the optimized path set with the dynamic control instruction set, the cutting mechanism is controlled to perform iterative cutting of the cutting area to obtain a cutting result dataset.

[0041] Further, after the step of obtaining the segmentation result dataset, the following steps are included:

[0042] The data set of the segmentation results is subjected to quality scoring to obtain a comprehensive score;

[0043] Determine whether the overall score reaches a preset quality threshold;

[0044] If the overall score reaches a preset quality threshold, the cutting process is deemed qualified, and the cutting result dataset is output. If the overall score does not reach the preset quality threshold, the optimized path set and dynamic control instruction set are readjusted, and the iterative cutting process is repeated until the overall score reaches the preset quality threshold.

[0045] The present invention also proposes a laser cutting device, comprising:

[0046] The acquisition module is used to acquire a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut, and to extract the high-dimensional feature vectors of the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain an initial cutting parameter set.

[0047] The planning module is used to plan the path of the cutting area according to the initial cutting parameter set to obtain the cutting path set;

[0048] The acquisition module is used to control the cutting mechanism to perform cutting processing according to the cutting path set, and to acquire multi-dimensional sequence data streams during the cutting process to generate a first state vector;

[0049] The optimization module is used to calculate the deviation between the cutting probability and the preset threshold based on the first state vector, and optimize the initial cutting parameter set based on the deviation to generate an optimized parameter set;

[0050] The control module is used to combine the optimized parameter set with the cutting path set for secondary cutting processing to generate a second state vector, and adjust the cutting path set according to the similarity between the second state vector and the first state vector until all regions are cut.

[0051] The present invention also proposes a computer device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method.

[0052] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0053] Compared with the prior art, this application has the following beneficial effects:

[0054] This application proposes a laser cutting method and equipment that solves the problems of cutting quality fluctuations and inefficiency caused by environmental disturbances and changes in material properties in existing laser cutting technologies. Unlike existing methods that preset static process parameters and cutting paths, this application obtains the two-dimensional surface feature map and three-dimensional thermal distribution map of the material to be cut, extracts its high-dimensional feature vector, and generates an initial cutting parameter set. Based on this parameter set, path planning and cutting operations are performed, which can more accurately reflect the actual state of the material. During the cutting process, the acquired multi-dimensional sequence data stream is used to calculate the first state vector to further dynamically evaluate the cutting effect. Based on this evaluation, this method optimizes the initial cutting parameter set by calculating the deviation value between the cut penetration probability and the preset threshold, and then generates an optimized cutting parameter set for secondary cutting processing. During this process, the cutting path set is continuously adjusted according to the feedback information during the cutting process, so that the cutting accuracy and heat-affected zone are effectively controlled.

[0055] In summary, this closed-loop adaptive control mechanism can respond in real time to changes in materials and environmental disturbances, ensuring the stability of cutting quality under complex geometries or variable working conditions. Compared with existing methods, the technical solution of this application can reduce the need for secondary grinding or rework, improve production efficiency, reduce the degree of manual intervention, and significantly improve the automation and intelligence level of the cutting process, thus significantly enhancing the application potential of laser cutting technology in modern manufacturing with high precision and high quality requirements. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] The structures, proportions, sizes, etc., shown in the accompanying drawings are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the implementation conditions of this application. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size should still fall within the scope of the technical content disclosed in this application, provided that they do not affect the effects and purposes that this application can produce.

[0058] Figure 1 This is a schematic diagram of the steps of a laser cutting method in one embodiment of the present invention;

[0059] Figure 2 This is a schematic block diagram of a laser cutting device according to an embodiment of the present invention;

[0060] Figure 3 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention;

[0061] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is referred to as “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connection or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0064] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0065] Reference Figure 1 This invention provides a laser cutting method, comprising the following steps:

[0066] S1: Obtain the two-dimensional surface feature map and the three-dimensional thermal distribution map of the material to be cut, and extract the high-dimensional feature vectors of the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain the initial cutting parameter set;

[0067] In step S1, the thermal conductivity, thickness, surface reflectivity, and microstructural characteristics (such as grain size and porosity) of the material being cut are key parameters affecting the laser cutting effect. For example, an optical scanning device can capture a two-dimensional surface feature map of the material being cut, recording its surface texture, gloss, and possible minor defects, such as scratches or oxide layers. Simultaneously, infrared thermal imaging technology is used to generate a three-dimensional thermal distribution map of the material, reflecting its thermal response characteristics in different regions, such as heat accumulation in certain areas due to differences in material density. After acquiring the two-dimensional surface feature map and the three-dimensional thermal distribution map, high-dimensional feature vectors are extracted. Based on a CNN, through multi-layer convolution and pooling operations, spatial features such as edges and textures can be extracted from the two-dimensional surface feature map, while dynamic features such as thermal gradients and thermal diffusion rates can be extracted from the three-dimensional thermal distribution map. Taking an aluminum alloy plate as an example, assuming its surface feature map displays periodic micro-textures, a CNN can identify these texture patterns and convert them into a set of high-dimensional vectors with dimensions between 128 and 256, depending on the network structure and the complexity of the training dataset. Three-dimensional thermal distribution maps reveal heat concentration in certain areas. CNNs encode the thermal distribution characteristics of these areas as part of a high-dimensional vector. The generation of these high-dimensional feature vectors preserves the original information of the material, and redundant information is removed through dimensionality reduction techniques (such as Singular Value Decomposition, SVD), thereby improving computational efficiency. SVD projects high-dimensional data into a low-dimensional space by decomposing the feature matrix, while retaining key information; for example, reducing a 256-dimensional feature vector to 100 dimensions, ensuring manageable computational complexity and no loss of key features. An initial cutting parameter model is constructed based on the extracted high-dimensional feature vectors. The goal of this model is to generate a set of initial cutting parameters, including laser power, cutting speed, focal point position, and auxiliary gas pressure. Taking an aluminum alloy plate as an example, assuming its high-dimensional feature vectors show high surface reflectivity and thermal conductivity, the model suggests a higher laser power (e.g., 3000W) to overcome reflection loss, while selecting a faster cutting speed (e.g., 10m / min) to reduce the heat-affected zone. Furthermore, the focal position needs to be adjusted to 0.5 mm below the material surface to ensure that the laser energy is concentrated in the cutting area, and the pressure of the auxiliary gas (such as nitrogen) is set to 1.2 MPa to effectively remove molten material. These parameters can be generated using the Support Vector Regression (SVR) algorithm. SVR takes high-dimensional feature vectors as input and predicts the optimal parameter combination through regression analysis. The advantage of SVR is its ability to handle nonlinear relationships and to map feature vectors to a higher-dimensional space through kernel functions (such as radial basis functions), thereby improving prediction accuracy and forming the initial cutting parameter set P1.

[0068] S2: Based on the initial cutting parameter set, perform path planning on the cutting area to obtain a cutting path set;

[0069] In step S2, the boundaries of the cutting area are extracted using image processing techniques. Based on the Canny edge detection algorithm, the boundary contours of the cutting area are identified through gradient calculation and thresholding, generating a boundary point set B. For example, suppose a rectangular stainless steel plate needs to be cut, with a complex geometric pattern designed on its surface, containing multiple curves and straight line segments. The Canny edge detection algorithm maps the two-dimensional surface features of the stainless steel plate, extracting the boundary points of the pattern to form a boundary point set B containing hundreds of coordinate points. These points record the geometric shape of the pattern. Based on the boundary point set B, path planning generates one or more continuous cutting paths, which can be achieved using an improved A* algorithm. The A* algorithm finds the optimal path from the starting point to the ending point while satisfying constraints. In laser cutting scenarios, the heuristic function of the A* algorithm comprehensively considers path length, corner constraints, and minimization of the heat-affected zone (HAZ). Path length directly affects cutting efficiency; shorter paths reduce processing time. Corner constraints ensure the stability of the cutting mechanism during path execution, preventing equipment vibration or reduced cutting quality caused by sharp turns. Minimizing the heat-affected zone is related to the thermal conductivity and heat distribution information in the initial cutting parameter set P1. For example, for stainless steel plates with high thermal conductivity, heat easily dissipates; the A* algorithm prioritizes shorter paths to reduce heat accumulation while avoiding overly dense paths that could lead to localized overheating. The heuristic function is designed using a weighted combination of Euclidean distance and heat diffusion coefficient, with weight optimization relying on gradient descent. Path planning is based on a region segmentation strategy, decomposing the entire cutting area into multiple sub-regions, each corresponding to an independent path segment. Region segmentation is based on the K-means clustering algorithm, dynamically determining cluster centers and the number of clusters by analyzing material characteristics such as thickness and thermal conductivity. For example, for a stainless steel plate with a thickness gradually increasing from 2 mm to 5 mm, the cutting area is divided into 3 to 10 sub-regions based on thickness distribution and thermal conductivity differences, each corresponding to an independent path segment. Cluster centers are determined by material property data in the initial cutting parameter set P1. For example, thicker regions may require higher laser power and slower cutting speeds, and are therefore divided into separate sub-regions. The output cutting path set T1 contains the path coordinates of each sub-region and the corresponding initial cutting parameters P1.

[0070] S3: Control the cutting mechanism to perform cutting processing according to the cutting path set, and collect multi-dimensional sequence data stream during the cutting process to generate a first state vector;

[0071] In step S3, during actual operation, the laser cutting machine starts according to these cutting path sets. For example, assuming the object to be cut is a 5mm thick 304 stainless steel plate, the laser cutting machine drives the laser head to move along a predetermined path via a servo motor. Simultaneously, the laser beam is focused on the material surface, achieving melting or vaporization of the material through high energy density, thereby completing the cut. At the same time, a multi-dimensional sequence data stream is acquired in real time during the cutting process, using integrated multi-modal sensors to capture multiple dimensions of information about the cutting state. Specifically, a laser reflection sensor monitors the reflection intensity of the laser beam on the material surface, which can indirectly reflect the cutting state of the material; for example, excessively high reflection intensity may indicate that the laser has not completely cut through. A thermal imager captures the temperature distribution in the cutting area; temperature data can indicate whether the material is overheated or has thermal stress concentration. A vibration sensor records the vibration frequency of the equipment or material during the cutting process, used to detect possible mechanical instability or cutting defects. These sensors work together to generate a multi-dimensional time-series data stream containing various parameters such as laser reflection intensity, regional temperature distribution, and vibration frequency, constituting the multi-dimensional sequence data stream D1. The data stream D1 is fed into a Long Short-Term Memory (LSTM) network model to predict the cutting state. For example, during the cutting process, the LSTM can predict the current degree of penetration and kerf width based on the laser reflection intensity and temperature distribution in the previous few seconds. The model's training dataset comes from historical cutting records, which may include cutting data under different materials (such as steel, aluminum, and copper), different thicknesses, and different cutting parameters. During training, the LSTM optimizes the weights through backpropagation, keeping the prediction error within ±3%, and outputs a real-time state vector S1, which includes indicators such as penetration probability and kerf quality score.

[0072] S4: Calculate the deviation between the cutting probability and the preset threshold based on the first state vector, and optimize the initial cutting parameter set based on the deviation to generate an optimized parameter set;

[0073] In step S4, the deviation between the penetration probability and a preset threshold is calculated based on the first state vector S1. The penetration probability refers to the likelihood that the laser can completely penetrate the material under the current parameters, and is predicted by a machine learning model (such as a support vector machine or neural network) on the first state vector. The preset threshold can be set according to process requirements. In implementation, the penetration probability can be calculated based on a model trained on historical cutting data, with the first state vector S1 as input and the penetration probability P as output. When the deviation ΔP exceeds a certain range (e.g., 5%), the parameter optimization process is triggered. The optimization process adjusts the initial cutting parameter set P1 to better adapt to the current cutting conditions. The optimization algorithm can be based on a Bayesian optimization algorithm, which constructs a surrogate model of the objective function (which can be a Gaussian process) and iteratively searches for the optimal solution in the parameter space. During optimization, the first layer (initial parameter model) and the third layer (real-time state prediction) are linked. Specifically, the first layer feature vector, namely the high-dimensional feature vectors of the two-dimensional surface feature map and the three-dimensional thermal distribution map extracted in step S1, extracts the main features through principal component analysis (PCA) to reduce dimensionality and retain key information. These feature vectors are weighted and fused with the first state vector S1. The weight coefficients of the fusion are calculated using the covariance matrix to generate a comprehensive feature vector. The fused data is used to update the parameter set, generating an optimized parameter set P2. The optimized parameter set P2 replaces some parameters of the initial parameter set P1, and the output is provided for the secondary cutting process in the subsequent step S5.

[0074] S5: Combine the optimized parameter set with the cutting path set for secondary cutting processing to generate a second state vector. Adjust the cutting path set according to the similarity between the second state vector and the first state vector until all regions are cut.

[0075] In step S5, the optimized parameter set generated in step S4 is combined with the cutting path set obtained in step S2 to perform secondary cutting on the cutting area. Simultaneously, a second state vector is generated and compared with the first state vector to dynamically adjust the cutting path and ensure cutting quality, ultimately completing the cutting of all areas. Specifically, during the secondary cutting based on the optimized parameter set and cutting path set, the parameter values ​​in the optimized parameter set are applied to the control system of the laser cutting equipment. During the cutting process, the equipment executes the cutting operation one by one according to the path segments in the cutting path set. Similar to step S2, multi-modal sensors collect cutting state data in real time to generate a second state vector. This state vector contains a multi-dimensional sequence data stream during the cutting process, such as kerf width, surface roughness, and temperature distribution in the heat-affected zone. This data is collected in real time by sensors; for example, an infrared thermal imager is used to monitor heat distribution, an optical sensor is used to measure the kerf width, and a laser interferometer is used to detect surface roughness. After generating the second state vector, it is compared and analyzed with the first state vector, and the cosine similarity calculation method is used to evaluate the difference between the two. The cosine similarity is calculated as cos(θ) = (S1·S2) / (||S1||·||S2||), where S1 and S2 are the first and second state vectors, respectively. If the cosine similarity is lower than a preset threshold (e.g., 0.9), it indicates a deviation between the secondary cutting effect and expectations, which may manifest as uneven kerf width or excessive surface roughness. In this case, local adjustments to the cutting path set are necessary. Path adjustment can minimize the heat-affected zone by re-optimizing path segments. The weight of a path segment can be defined as the weighted sum of the area of ​​the heat-affected zone or the cutting time, prioritizing the path that minimizes thermal damage to the material. For example, if the cutting path of a certain sub-region results in an excessively large heat-affected zone, a new path bypassing the high-temperature region will be recalculated. The adjusted path may increase the cutting time slightly, but it can significantly improve the cutting quality. The cutting state of the next sub-region is predicted using an Autoregressive Integral Moving Average (ARIMA) model. The parameters p (autoregressive order), d (difference order), and q (moving average order) of the ARIMA model are determined using the Akaike Information Criterion (AIC) to ensure the accuracy of the model's prediction of the cutting state. The prediction results can be used to adjust the cutting parameters in advance. For example, if it is predicted that the kerf width in the next sub-region may exceed the limit, the laser power can be appropriately reduced or the auxiliary gas pressure increased. The adjusted path and parameters are used for secondary cutting to generate a new second state vector. Once the requirements are met, cutting of the next sub-region continues until the entire cutting is completed.

[0076] In one embodiment, the step of obtaining a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut, and extracting high-dimensional feature vectors from the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain an initial cutting parameter set includes:

[0077] Collect surface texture, reflectivity information and three-dimensional thermal distribution data of the material surface to generate a raw multidimensional dataset containing two-dimensional surface feature maps and three-dimensional thermal distribution maps;

[0078] Extract high-dimensional feature vectors from the original multidimensional dataset to obtain a feature vector matrix, wherein the high-dimensional feature vectors include at least surface roughness, thermal conductivity gradient, material thickness and environmental variables;

[0079] The dimensionality reduction process is performed on the eigenvector matrix. During the dimensionality reduction process, orthogonality constraints are applied to the eigenvector matrix to generate a dimensionality-reduced eigenvector matrix.

[0080] Based on the reduced feature matrix, parameter modeling is performed, and training is conducted using the radial basis function kernel as the kernel function to generate an initial cutting parameter set.

[0081] In the above embodiments, surface texture, reflectivity information, and three-dimensional thermal distribution data of the material are acquired using an optical scanner and an infrared thermal imager, respectively. These acquired data are initially integrated to form a multidimensional dataset containing a two-dimensional surface feature map and a three-dimensional thermal distribution map. The original multidimensional dataset is preprocessed: Gaussian filtering is applied to the two-dimensional surface feature map to remove noise, and wavelet transform is applied to the three-dimensional thermal distribution map for denoising. Both feature maps are then normalized to the [0, 1] interval. Specifically, the Gaussian filtering uses a parameter with a standard deviation of 0.5 to smooth the two-dimensional surface feature map, effectively removing random noise that may be introduced during optical scanning while retaining effective surface texture information. The wavelet transform uses the Daubechies basis function to denoise the three-dimensional thermal distribution map. This method better preserves the local features of the thermal distribution data and avoids the loss of details that may occur with traditional filtering methods. In the feature extraction stage, a residual neural network (ResNet) is used to perform deep feature analysis on the preprocessed two-dimensional surface feature map and three-dimensional thermal distribution map, extracting high-dimensional feature vectors containing surface roughness, thermal conductivity gradient, material thickness, and environmental variables. In its implementation, the ResNet network employs a deep structure of at least 50 layers, combined with the ReLU activation function. Its residual connection mechanism effectively mitigates the vanishing gradient problem in deep networks, thereby improving the accuracy of feature extraction. The input 2D surface feature map and 3D thermal distribution map are fed into the ResNet network, undergoing multiple convolutional and pooling operations to extract high-dimensional feature vectors. These feature vectors not only contain physical properties such as surface roughness and thermal conductivity gradients but also incorporate external factors such as material thickness and environmental variables (e.g., temperature, humidity), fully characterizing the multidimensional properties of the cut material. To ensure the representativeness of the extracted features, the ResNet model is pre-trained on a large amount of labeled data and fine-tuned on specific material datasets to adapt to the characteristics of different materials. The feature vector matrix is ​​decomposed using Singular Value Decomposition (SVD), retaining the feature information corresponding to the first 90% of the singular values ​​to generate a dimensionality-reduced feature matrix. In the SVD decomposition process, orthogonality constraints are applied to ensure that the reduced feature matrix retains the geometric structure of the data. Specifically, the SVD algorithm in numerical computing libraries (such as NumPy or MATLAB) can be used to ensure the efficiency and stability of the computation. Support Vector Regression (SVR) is employed, using a radial basis function (RBF) kernel to train the prediction model and generate an initial set of cutting parameters including laser power, cutting speed, focal position, and auxiliary gas pressure. The SVR algorithm maps the reduced feature matrix to a high-dimensional space, searching for an optimal regression hyperplane to minimize the mean square error while satisfying physical constraints (such as laser power range limitations). The choice of the RBF kernel function effectively captures the nonlinear relationships between features, thereby improving the model's prediction accuracy.During training, cross-validation is used to determine the hyperparameters of the SVR model (such as the regularization parameter C and the bandwidth parameter γ of the RBF kernel) to ensure the model's generalization ability. The parameter set is validated and optimized using a finite element simulation model. Specifically, the finite element simulation model is used to simulate the thermal stress distribution and cut quality during the cutting process, and a physical model based on the material response calculates indices such as cut depth and cut smoothness. Based on the simulation results, the Adam optimization algorithm is used to iteratively adjust the initial cutting parameter set. The optimization objective is to minimize the cost function based on cut depth and material response, thereby generating the initial cutting parameter set.

[0082] In one embodiment, the step of performing path planning on the cutting region based on the initial cutting parameter set to obtain a cutting path set includes:

[0083] The initial cutting parameter set is processed by feature mapping to obtain the region feature vector set;

[0084] The segmented region is partitioned based on the regional feature vector set to obtain the sub-region parameter set;

[0085] The sub-region parameter set is initialized with paths, and global optimization is performed based on a genetic algorithm to adjust the path direction and order, thus obtaining an initial path set.

[0086] The cutting order is scheduled based on the initial path set to obtain the cutting path set.

[0087] In the above embodiments, the initial cutting parameter set is processed by feature mapping, transforming it into a feature vector set that comprehensively describes the characteristics of the cutting region. A multi-scale feature extraction algorithm is used to process the two-dimensional surface feature map and three-dimensional thermal distribution map of the cutting region, and a high-dimensional feature vector is extracted using a convolutional neural network (CNN). The CNN, through multi-layer convolution and pooling operations, can capture the local geometric features and thermal distribution patterns of the region, such as edge curvature, surface unevenness, and thermal gradient distribution. To reduce the dimensionality of the feature vectors while retaining key information, principal component analysis (PCA) is used to reduce the dimensionality of the high-dimensional features, allowing the feature vectors to incorporate material property parameters (such as density and thermal conductivity) while preserving geometric and thermal distribution characteristics, generating a regional feature vector set. Next, the cutting region is partitioned according to the regional feature vector set, dividing the complex cutting region into several sub-regions for personalized path planning based on the characteristics of different regions. Based on the regional feature vector set, an adaptive clustering algorithm (such as DBSCAN density clustering) is used to partition the cutting region. Density clustering combines geometric information (such as boundary curvature and region area) and thermal distribution information (such as thermal gradient and thermally affected zone range) from feature vectors. It dynamically adjusts the number of clusters (between 4 and 12, depending on region complexity and material properties) to divide the cut region into multiple sub-regions. Each sub-region has clear boundary coordinates and feature labels, such as its geometric center, average thermal conductivity, and material thickness. After partitioning, path initialization is performed on the sub-region parameter set. The boundary coordinates and feature labels of each sub-region are analyzed, and by introducing dynamic weighting factors, the optimal path from the starting point to the ending point is calculated, taking into account factors such as the sub-region's thermal conductivity, material thickness, and boundary curvature. This generates a relatively reasonable initial path set. The initial path set contains the path coordinates of each sub-region. Based on the initial path set and the sub-region parameter set, a genetic algorithm is used for global path optimization. The genetic algorithm uses path smoothness and cutting efficiency as fitness functions, generating new path candidates through crossover and mutation operations. The crossover operation reorganizes different path segments, while the mutation operation fine-tunes the path direction and order. Simultaneously, it adjusts the path's geometric properties by combining cutting parameters (such as laser power and cutting speed) from the sub-region parameter set. Furthermore, a collision detection mechanism is introduced to avoid path intersections or overlaps by examining the spatial relationships between paths. Iterative processing ensures path convergence, generating an optimized path set containing the coordinates and corresponding parameters of the optimized paths. The cutting sequence is then scheduled based on this optimized path set. A time-series scheduling algorithm is used, comprehensively considering the connection distance between paths, thermal accumulation effects, and cutting equipment motion constraints, and the cutting sequence is optimized using dynamic programming. By constructing a state transition equation, the total cost of different cutting sequences (including time cost and thermal deformation risk) is calculated, prioritizing the cutting of heat-sensitive areas to reduce material deformation caused by thermal accumulation.Meanwhile, in order to reduce the acceleration changes during the movement of the equipment, the transition sections between paths are smoothed. For example, the path connection points are optimized by using Bézier curves or spline interpolation methods. The generated cutting path set contains the optimized path coordinates of each sub-region, the corresponding cutting parameter configuration, and the global cutting order.

[0088] In one embodiment, the step of controlling the cutting mechanism to perform cutting processing according to the cutting path set, and acquiring multi-dimensional sequence data streams during the cutting process to generate a first state vector includes:

[0089] The cutting machine is controlled to perform cutting operations along the cutting path set, and the signal set generated during the cutting process is collected by multiple sensors to generate a raw multidimensional data stream.

[0090] The original multidimensional data stream is subjected to noise reduction and feature enhancement processing to obtain an optimized data stream;

[0091] High-dimensional features of the optimized data stream are extracted based on a deep learning feature extraction network to generate a feature vector set.

[0092] The feature vector set is used for state prediction, and the multiple prediction results are weighted and fused to obtain the first state vector.

[0093] In the above embodiment, a laser cutting machine is activated to perform cutting operations along the cutting path set, and multidimensional data streams are acquired through configured sensors. During this process, a laser reflection sensor monitors the reflection intensity of the laser beam on the material surface in real time; this data reflects changes in the material surface state. A thermal imager captures the temperature distribution of the cutting area, generating two-dimensional temperature field data for analyzing the extent of the heat-affected zone. A vibration sensor records the vibration frequency between the material and the cutting mechanism, reflecting the mechanical stability during the cutting process. An acoustic sensor collects acoustic signals generated during the cutting process; these signals may contain characteristic information of cutting defects (such as cracks or pores). By integrating this data from different sensors, a raw multidimensional sequence data stream containing reflection intensity, temperature distribution, vibration frequency, and acoustic characteristics is formed. An adaptive Kalman filter dynamically adjusts the filter gain to eliminate noise interference based on the material's thermal response and vibration characteristics. Wavelet transform technology is used for feature enhancement; by performing time-frequency analysis on the data stream, key features are highlighted, such as abrupt changes in reflection intensity, peaks in temperature gradients, or abnormal changes in vibration frequency. After noise reduction and feature enhancement processing, an optimized data stream is generated, which has a higher signal-to-noise ratio in both time and space dimensions. The optimized data stream is input into a deep learning feature extraction network to extract high-dimensional features and generate a feature vector set. The network structure can employ a three-layer convolutional neural network (CNN) and two fully connected layers. The CNN extracts local spatiotemporal features from the optimized data stream using 64 convolutional kernels (with a stride of 2), such as capturing local gradient changes in temperature distribution or periodic patterns in vibration frequency. Convolutional operations effectively extract spatial correlations and temporal dependencies in the data stream. The fully connected layers map these local features to a high-dimensional space, generating a feature vector set containing reflection intensity distribution, temperature gradient, vibration patterns, and sound wave characteristics. These feature vectors not only retain the key information of the original data but also enhance the data's expressive power through high-dimensional representation. Based on the feature vector set, a pre-trained Long Short-Term Memory (LSTM) network is used for state prediction. The LSTM network consists of three layers, each with 128 hidden units. Through training on historical cutting data, the network learns to map feature vectors to specific cutting state indicators, such as cut penetration, kerf width, and surface quality. The prediction results include the penetration probability (reflecting whether the material is completely cut through), the estimated kerf width (reflecting cutting accuracy), and the surface roughness score (reflecting cutting quality). The initial state vector is weighted and fused to generate the final first state vector. The weighted fusion algorithm comprehensively considers the influence of material type, cutting speed, and laser power on each state index, dynamically adjusting the weight coefficients to generate a first state vector that includes a comprehensive penetration probability, kerf quality score, and cutting stability score, comprehensively reflecting the state of the cutting process.

[0094] In one embodiment, the step of calculating the deviation between the cutting probability and a preset threshold based on the first state vector, and optimizing the initial cutting parameter set based on the deviation to generate an optimized parameter set includes:

[0095] Based on a pre-trained deep neural network model, the first state vector is transformed into a probability distribution space, and the deviation between the current cut-through probability and the preset threshold is calculated.

[0096] Based on the deviation value, the initial cutting parameter set is initially adjusted to obtain the parameter adjustment vector;

[0097] The initial cutting parameter set is optimized using reinforcement learning based on the parameter adjustment vector to obtain an intermediate optimized parameter set.

[0098] The features of the intermediate optimization parameter set are weighted and fused with the spatiotemporal features of the first state vector to obtain a fused feature vector;

[0099] The intermediate optimization parameter set is subjected to a global search based on the fused feature vector to obtain the optimization parameter set;

[0100] The optimized parameter set and the cutting path set are adjusted together to obtain the adjusted cutting path set.

[0101] In the above embodiment, the deviation between the cut-through probability and a preset threshold is calculated based on the first state vector. A pre-trained deep neural network model is used to perform feature mapping on the first state vector and transform it into a probability distribution space, thereby quantifying the gap between the current cutting state and the ideal cut-through effect. A convolutional neural network (CNN) is used to extract spatiotemporal features from the first state vector. Through multiple convolutional and pooling operations, the CNN can effectively capture local and global features in the state vector, forming a high-dimensional feature tensor. This feature tensor is fed into a fully connected layer, where an activation function (such as ReLU or Sigmoid) maps the feature tensor to a scalar value, i.e., the current cut-through probability. The preset threshold is determined based on the optimal cut-through standard from historical cutting data; for example, by analyzing a large number of successful cutting cases, an ideal cut-through probability value (e.g., 0.95) is determined. The deviation is calculated as the difference between the current cut-through probability and the preset threshold. The absolute value of the deviation directly reflects the gap between the current cutting effect and the expectation. Based on the calculated deviation, the Gradient Boosting Decision Tree (GBDT) algorithm is used to analyze the correlation between the deviation and the initial cutting parameter set. Gradient boosting decision trees effectively capture the nonlinear relationship between deviation values ​​and various parameters by constructing multiple decision trees and progressively optimizing the loss function. In the algorithm implementation, the deviation value is set as the target variable, and the parameters in the initial cutting parameter set (such as laser power and cutting speed) are considered as input features. Through training, the gradient boosting decision tree model can calculate the weight of each parameter's influence on the deviation value. Based on these weights, the adjustment direction and magnitude of each parameter are further determined. For example, if the deviation value is negative, it indicates insufficient cutting penetration, which may require increasing laser power or decreasing the cutting speed, generating a parameter adjustment vector where each component corresponds to the adjustment amount of a parameter. After obtaining the parameter adjustment vector, the initial cutting parameter set is optimized using reinforcement learning. Based on a Markov Decision Process (MDP) reinforcement learning algorithm, a better parameter combination is found through dynamic optimization. The reinforcement learning algorithm uses the parameter adjustment vector as the action space, the first state vector and historical cutting data as the state space, and a weighted sum of cutting quality score and cutting efficiency as the reward function. The parameter adjustment magnitude is progressively optimized through multiple iterations. Specifically, in each iteration, an action (i.e., a parameter adjustment vector) is selected based on the current state (i.e., the first state vector and historical data), and a reward value is calculated through a simulated or actual cutting process. The reward value reflects the effect of the current parameter adjustment; for example, higher cutting quality and efficiency will result in a higher reward value. By updating the policy network (e.g., using a deep Q-network or policy gradient method), the action selection is progressively optimized, generating an intermediate set of optimized parameters.The first state vector undergoes similarity processing to extract its high-dimensional spatiotemporal features, such as the temporal variation trend of the cutting depth or the spatial distribution characteristics of the heat-affected zone. By calculating the covariance matrix of the two sets of features, the weighting coefficients for weighted fusion can be determined; these coefficients reflect the correlation between the parameter features and the real-time state features. The fusion process generates a fused feature vector through weighted summation. This vector integrates the cutting parameters and real-time state information, forming a more comprehensive feature representation. The fused feature vector contains the direction and magnitude of parameter adjustments, incorporating dynamic state information during the cutting process. Based on the fused feature vector, a global search is performed on the intermediate optimization parameter set, employing a genetic algorithm to search for the optimal solution in the parameter space. The genetic algorithm uses the fused feature vector as input to the fitness function, with the parameter space being the range of the initial cutting parameter set. It performs operations including population initialization, crossover, mutation, and selection, where each individual in the population is a variant of an intermediate optimization parameter set. The fitness function comprehensively considers kerf quality, cutting efficiency, and cut-through probability, for example, by calculating the fitness value through weighted summation. The genetic algorithm iteratively optimizes individuals in the population, exploring potential optimal solutions in the parameter space through crossover and mutation operations, while retaining individuals with higher fitness through selection. Its global search capability effectively overcomes the limitations of local optimization, generating an optimized parameter set containing optimized values ​​for parameters such as laser power, cutting speed, and focal point position. Based on this optimized parameter set, the cutting path set is adjusted in a coordinated manner, specifically through a three-layer structure. The first layer analyzes the incomplete cutting features in the second state vector using a gradient boosting decision tree, such as the location and extent of incomplete cutting areas, to determine the path adjustment direction. The second layer optimizes the path adjustment magnitude through reinforcement learning, based on historical path data and the current task objective, such as minimizing cutting time or maximizing path smoothness. The third layer uses the genetic algorithm to globally optimize the adjusted path, ensuring its global optimality. During execution, the three layers work collaboratively through information feedback: the first layer outputs the path adjustment direction to the second layer, the second layer optimizes the magnitude and feeds it back to the third layer, and the third layer verifies the adjusted path set back to the first layer, iterating until the path set converges. The generated set of adjusted cutting paths works in conjunction with the optimized parameter set to ensure the accuracy and efficiency of the cutting process. For example, the optimized laser power and cutting speed are matched with the adjusted path curvature and angular velocity to achieve high-quality through-cutting results in cutting complex geometries.

[0102] In one embodiment, the step of combining the optimized parameter set with the cutting path set for secondary cutting to generate a second state vector, and adjusting the cutting path set according to the similarity between the second state vector and the first state vector until all regions are cut, includes:

[0103] The optimized parameter set is nonlinearly mapped to the cutting path set to generate an initial fusion parameter set;

[0104] The initial fusion parameter set is dynamically adjusted in real time to obtain a dynamic control command set;

[0105] The cutting mechanism is driven to perform secondary cutting according to the dynamic control instruction set, and the secondary cutting data stream during the secondary cutting process is collected to generate a second state vector.

[0106] The path adjustment trigger signal is obtained by performing a similarity analysis between the second state vector and the first state vector.

[0107] The cutting path set is locally optimized based on the path adjustment trigger signal to obtain an optimized path set.

[0108] By combining the optimized path set with the dynamic control instruction set, the cutting mechanism is controlled to perform iterative cutting of the cutting area to obtain a cutting result dataset.

[0109] In the above embodiments, a multi-layer parameter mapping network is constructed to achieve nonlinear mapping. This mapping network can perform high-dimensional spatial correlation analysis between the physical parameters in the optimized parameter set and the geometric features in the cutting path set. For example, the laser power may need to be dynamically adjusted according to the corner angle or path length to avoid overheating or insufficient cutting. The mapping network integrates these parameters into an initial fusion parameter set through methods such as deep learning or nonlinear regression. Based on the initial fusion parameter set, real-time dynamic adjustments are performed to generate a dynamic control command set. Real-time data is collected using multimodal sensors, such as temperature sensors to monitor the heat distribution in the cutting area, vibration sensors to detect the stability of the equipment, and light intensity sensors to evaluate the laser output effect. This multi-dimensional data is processed through time series analysis algorithms, such as using Kalman filtering or recurrent neural networks to smooth and predict the data, thereby identifying the changing trends of environmental variables. Based on these analysis results, the parameters in the initial fusion parameter set are adjusted in real time, such as reducing the laser power when the material surface temperature is detected to be too high, or slowing down the cutting speed at corners to improve accuracy, generating a dynamic control command set containing specific control signals for the servo motor, laser, and auxiliary gas system. During the cutting operation, a dynamic control instruction set drives the cutting mechanism to perform secondary cutting, while a multimodal sensor system comprehensively monitors the cutting area. These sensors can collect multidimensional data streams in real time, including kerf width, surface roughness, heat-affected zone range, and cutting depth. These data streams are mapped to a high-dimensional vector space through feature extraction algorithms, such as principal component analysis or deep convolutional networks, generating a second state vector. A similarity analysis is performed between the second and first state vectors using an improved cosine similarity algorithm to calculate a similarity score. Specifically, the cosine similarity algorithm evaluates the similarity of the angle between the two vectors and compares it with a preset similarity threshold. If the similarity score is lower than the threshold, it indicates that the current cutting state deviates from expectations, and a path adjustment trigger signal is generated. The signal not only identifies the path segment that needs adjustment but also contains information about the degree of deviation, such as the extent of the kerf width exceeding the standard or the abnormal range of the heat-affected zone. Based on the path adjustment trigger signal, the cutting path set is locally optimized to generate an optimized path set, and the path segments are replanned. The coordinates and order of path segments can be recalculated by combining the constraints of minimizing the heat-affected zone and the requirements of path smoothness. For example, when a path segment is found to have degraded cutting quality due to heat accumulation, a smoother path is prioritized to reduce the heat impact while maintaining cutting efficiency. The optimized path set can effectively reduce quality problems caused by path deviations or heat accumulation, while ensuring that the overall cutting efficiency is not significantly affected. This process iteratively approximates the optimal path configuration. The optimized path set and dynamic control instruction set are used to iteratively cut the cutting area until the final cutting result dataset is generated.The cutting mechanism operates precisely based on the coordinates of the optimized path set and the signals of the dynamic control command set, while continuously monitoring the cutting status through multimodal sensors. This monitoring data is used not only for real-time adjustments but also to predict the state of subsequent cutting areas using an autoregressive prediction model. For example, based on the current kerf width and heat-affected zone data, the model can predict the cutting state of the next area, thus updating the path and parameters in advance. The final cutting result dataset includes key indicators such as kerf width, surface roughness, and cutting depth, which comprehensively reflect the cutting quality. Through continuous iterative optimization of the path and commands, high precision and high quality are ensured throughout the entire cutting process until all areas are processed.

[0110] In one embodiment, after the step of obtaining the segmented result dataset, the following is included:

[0111] The data set of the segmentation results is subjected to quality scoring to obtain a comprehensive score;

[0112] Determine whether the overall score reaches a preset quality threshold;

[0113] If the overall score reaches a preset quality threshold, the cutting process is deemed qualified, and the cutting result dataset is output. If the overall score does not reach the preset quality threshold, the optimized path set and dynamic control instruction set are readjusted, and the iterative cutting process is repeated until the overall score reaches the preset quality threshold.

[0114] In the above embodiments, a quantitative relationship model between various indicators and cutting quality is established, such as by using regression analysis or machine learning models, to score the cutting result dataset and obtain a comprehensive score. A preset quality threshold is determined based on the optimal quality level of historical cutting data. By analyzing a large number of successful cutting cases, an ideal comprehensive score value is determined as a standard. If the comprehensive score reaches or exceeds the preset quality threshold, the cutting process is deemed qualified, and the cutting result dataset is output for subsequent use. If the comprehensive score is lower than the preset quality threshold, it indicates insufficient cutting quality, requiring readjustment of the optimized path set and dynamic control instruction set. The readjustment process is based on quality score feedback, analyzing specific problems in the cutting process, such as excessive kerf width or excessive heat-affected zone, and adjusting the path and parameters accordingly. By iteratively re-processing the cutting, the optimal cutting state is gradually approached until the comprehensive score reaches the preset quality threshold, ensuring that the final output cutting result dataset meets high-quality requirements. Specifically, the readjustment of the optimized path set and dynamic control instruction set can be divided into two steps. First, based on the inadequacy of the comprehensive score, it is analyzed which feature sub-scores performed poorly. For example, if the edge smoothness score is low, it may be necessary to adjust the curvature of the cutting path or the dynamic changes in laser power; if the heat-affected zone is too large, it may be necessary to reduce the laser power or optimize the cooling parameters. For instance, the reinforcement learning optimization of the initial cutting parameter set based on the deviation value mentioned in the above embodiment can be achieved using a similar reinforcement learning algorithm. This involves constructing a reward function (with the comprehensive score as the optimization objective) to iteratively adjust the optimized path set and the dynamic control instruction set. The reinforcement learning model will explore different combinations of paths and control parameters based on the feedback of the current comprehensive score, gradually approaching the optimal solution. Alternatively, a similarity analysis can be performed between the second state vector and the first state vector to obtain a path adjustment trigger signal for this embodiment. By comparing the difference between the current cutting state and the target state, a specific adjustment signal is generated. After adjustment, the updated optimized path set and dynamic control instruction set are re-input into the cutting mechanism to execute a new round of iterative cutting processing. This process is similar to the secondary cutting process, i.e., a new initial fusion parameter set is generated through nonlinear mapping, and the cutting mechanism is driven to cut according to the dynamic control instruction set. Each iterative cutting generates a new cutting result dataset and repeats the above quality scoring and threshold judgment process until the comprehensive score reaches the preset quality threshold. To improve efficiency, an adaptive step size mechanism can be introduced to dynamically adjust the optimization magnitude based on the similarity of the comprehensive scores, thus avoiding excessive iterations.

[0115] Reference Figure 2 A laser cutting device, comprising:

[0116] The acquisition module 100 is used to acquire a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut, and to extract the high-dimensional feature vectors of the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain an initial cutting parameter set.

[0117] Planning module 200 is used to plan the path of the cutting area according to the initial cutting parameter set to obtain a cutting path set;

[0118] The acquisition module 300 is used to control the cutting mechanism to perform cutting processing according to the cutting path set, and to acquire multi-dimensional sequence data streams during the cutting process to generate a first state vector.

[0119] The optimization module 400 is used to calculate the deviation value between the cutting probability and the preset threshold based on the first state vector, and optimize the initial cutting parameter set based on the deviation value to generate an optimized parameter set;

[0120] The control module 500 is used to combine the optimized parameter set with the cutting path set for secondary cutting processing to generate a second state vector, and adjust the cutting path set according to the similarity between the second state vector and the first state vector until all regions are cut.

[0121] Reference Figure 3 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data such as a laser cutting method database. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a laser cutting method.

[0122] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements a laser cutting method, including the following steps: acquiring a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut; extracting high-dimensional feature vectors from the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain an initial cutting parameter set; performing path planning on the cutting area according to the initial cutting parameter set to obtain a cutting path set; controlling a cutting mechanism to perform cutting processing according to the cutting path set, and collecting multi-dimensional sequence data streams during the cutting process to generate a first state vector; calculating the deviation value between the cut-through probability and a preset threshold based on the first state vector; optimizing the initial cutting parameter set based on the deviation value to generate an optimized parameter set; combining the optimized parameter set with the cutting path set for secondary cutting processing to generate a second state vector; adjusting the cutting path set according to the similarity between the second state vector and the first state vector until all areas are cut.

[0123] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0124] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A laser cutting method, characterized by, include: Obtain a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut, and extract the high-dimensional feature vectors of the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain the initial cutting parameter set; Based on the initial cutting parameter set, a path planning process is performed on the cutting area to obtain a cutting path set; The cutting mechanism is controlled to perform cutting processing according to the cutting path set, and a multi-dimensional sequence data stream is collected during the cutting process to generate a first state vector; The deviation between the cut probability and the preset threshold is calculated based on the first state vector. The initial cutting parameter set is then optimized based on the deviation to generate an optimized parameter set. The optimized parameter set is combined with the cutting path set for secondary cutting to generate a second state vector. The cutting path set is then adjusted based on the similarity between the second state vector and the first state vector until all regions are cut. The step of calculating the deviation between the cutting probability and a preset threshold based on the first state vector, and optimizing the initial cutting parameter set based on the deviation to generate an optimized parameter set includes: Based on a pre-trained deep neural network model, the first state vector is transformed into a probability distribution space, and the deviation between the current cut-through probability and the preset threshold is calculated. Based on the deviation value, the initial cutting parameter set is initially adjusted to obtain the parameter adjustment vector; The initial cutting parameter set is optimized using reinforcement learning based on the parameter adjustment vector to obtain an intermediate optimized parameter set. The features of the intermediate optimization parameter set are weighted and fused with the spatiotemporal features of the first state vector to obtain a fused feature vector; The intermediate optimization parameter set is subjected to a global search based on the fused feature vector to obtain the optimization parameter set; The optimized parameter set and the cutting path set are adjusted together to obtain the adjusted cutting path set.

2. The laser cutting method according to claim 1, characterized in that, The steps of obtaining a two-dimensional surface feature map and a three-dimensional thermal distribution map of the material to be cut, and extracting high-dimensional feature vectors from the two-dimensional surface feature map and the three-dimensional thermal distribution map to obtain an initial cutting parameter set include: Collect surface texture, reflectivity information and three-dimensional thermal distribution data of the material surface to generate a raw multidimensional dataset containing two-dimensional surface feature maps and three-dimensional thermal distribution maps; Extract high-dimensional feature vectors from the original multidimensional dataset to obtain a feature vector matrix, wherein the high-dimensional feature vectors include at least surface roughness, thermal conductivity gradient, material thickness and environmental variables; The dimensionality reduction process is performed on the eigenvector matrix. During the dimensionality reduction process, orthogonality constraints are applied to the eigenvector matrix to generate a dimensionality-reduced eigenvector matrix. Based on the reduced feature matrix, parameter modeling is performed, and training is conducted using the radial basis function kernel as the kernel function to generate an initial cutting parameter set.

3. The laser cutting method of claim 1, wherein, The step of performing path planning on the cutting region based on the initial cutting parameter set to obtain a cutting path set includes: The initial cutting parameter set is processed by feature mapping to obtain the region feature vector set; The segmented region is partitioned based on the regional feature vector set to obtain the sub-region parameter set; The sub-region parameter set is subjected to path initialization processing, and global optimization is performed based on a genetic algorithm to adjust path direction and sequence, to obtain an initial path set; According to the initial path set, scheduling processing is performed on the cutting sequence to obtain a cutting path set.

4. The laser cutting method of claim 1, wherein, The step of controlling the cutting mechanism to perform cutting processing according to the cutting path set and collecting a multi-dimensional sequence data stream in the cutting process to generate a first state vector comprises: The cutting mechanism is controlled to perform a cutting operation along the cutting path set, and a signal set generated in the cutting process is collected through a plurality of sensors to generate an original multi-dimensional data stream; The original multi-dimensional data stream is subjected to noise reduction and feature enhancement processing to obtain an optimized data stream; High-dimensional features of the optimized data stream are extracted based on a deep learning feature extraction network to generate a feature vector set; The feature vector set is subjected to state prediction, and a plurality of prediction results are subjected to weighted fusion processing to obtain a first state vector.

5. The laser cutting method of claim 1, wherein, The step of combining the optimized parameter set and the cutting path set to perform secondary cutting processing to generate a second state vector, and adjusting the cutting path set according to the similarity between the second state vector and the first state vector until all regions are cut comprises: The optimized parameter set and the cutting path set are subjected to nonlinear mapping to generate an initial fusion parameter set; The initial fusion parameter set is subjected to real-time dynamic adjustment to obtain a dynamic control instruction set; The cutting mechanism is driven according to the dynamic control instruction set to perform secondary cutting processing, and secondary cutting data stream in the secondary cutting process is collected to generate a second state vector; The second state vector and the first state vector are subjected to similarity analysis to obtain a path adjustment trigger signal; The cutting path set is subjected to local optimization processing according to the path adjustment trigger signal to obtain an optimized path set; The optimized path set and the dynamic control instruction set are combined to control the cutting mechanism to perform iterative cutting processing on the cutting region to obtain a cutting result data set.

6. The laser cutting method of claim 5, wherein, After the step of obtaining the cutting result data set, the following steps are included: The cutting result data set is subjected to quality scoring processing to obtain a comprehensive score; It is judged whether the comprehensive score reaches a preset quality threshold; If the comprehensive score reaches the preset quality threshold, it is determined that the cutting processing is qualified, and the cutting result data set is output; if the comprehensive score does not reach the preset quality threshold, the optimized path set and the dynamic control instruction set are subjected to re-adjustment processing, and iterative cutting processing is performed again until the comprehensive score reaches the preset quality threshold.

7. A laser cutting apparatus applied to the laser cutting method according to any one of claims 1 to 6, characterized by, It comprises: An acquisition module is configured to acquire a two-dimensional surface feature map and a three-dimensional thermal distribution map of a cutting material, extract high-dimensional feature vectors of the two-dimensional surface feature map and the three-dimensional thermal distribution map, and obtain an initial cutting parameter set; A planning module is configured to perform path planning on a cutting region according to the initial cutting parameter set to obtain a cutting path set; An acquisition module is configured to control a cutting mechanism to perform cutting processing according to the cutting path set and collect a multi-dimensional sequence data stream in the cutting process to generate a first state vector; An optimization module is configured to calculate a deviation value of the breakthrough probability from a preset threshold value according to the first state vector, optimize the initial cutting parameter set based on the deviation value, and generate an optimized parameter set; A control module is configured to combine the optimized parameter set with the cutting path set for secondary cutting processing, generate a second state vector, and adjust the cutting path set according to a similarity between the second state vector and the first state vector until all regions are cut.

8. A computer device, comprising: The computer program is stored in the memory and executable in the processor, and the processor executes the computer program to implement the method in any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executable in the processor to implement the steps of the method in any one of claims 1 to 6.

Images