A method and system for intelligently optimizing cutting path and parameters in a laser cutting machine

Abstract

This application provides a method and system for intelligently optimizing the cutting path and parameters of a laser cutting machine, belonging to the field of laser cutting machine control technology. The method includes: acquiring task parameters and querying a process parameter experience model library to generate initial processing parameters; simultaneously employing an improved genetic-ant colony hybrid algorithm to solve for the theoretically optimal cutting path sequence; collecting multi-source feature information and inputting it into a pre-trained state recognition model to identify the current cutting state in real time; dynamically adjusting the laser power and cutting speed based on the current cutting state; and performing local replanning of the theoretically optimal cutting path sequence based on multi-source feature information, including geometric deviation compensation and dynamic collision avoidance, to generate corrected motion control commands; and using an incremental learning algorithm to update and optimize the dual-channel convolutional neural network and the process parameter experience model library. This application shortens processing time, improves cutting quality stability, reduces energy consumption and defect rate, and enhances the level of intelligence.

Description

Technical Field

[0001] This application belongs to the field of laser cutting machine control technology, specifically relating to a method and system for intelligently optimizing the cutting path and parameters of a laser cutting machine. Background Technology

[0002] Laser cutting, as a high-energy beam processing method, is widely used in metal manufacturing, automotive, aerospace, and precision instrumentation industries due to its significant advantages of high precision, high speed, flexibility, and non-contact processing. The performance of a laser cutting machine directly affects production efficiency and processing quality; therefore, improving cutting efficiency is a key issue that needs to be addressed.

[0003] However, despite the relative maturity of laser cutting technology, most existing laser cutting systems still suffer from the following problems in achieving efficient and high-quality processing: First, in terms of path planning, current laser cutting systems largely rely on simple geometric sorting or manual experience, lacking global optimization capabilities. This results in excessively long idle travel of the cutting head between different contours, leading to a high proportion of ineffective time. Second, in terms of process control, parameters are usually pre-set based on material type and thickness, failing to respond in real time to dynamic changes in working conditions such as material inhomogeneity, heat accumulation effects, and nozzle wear, easily causing quality defects such as incomplete cutting, overheating, or slag buildup. Finally, in terms of process monitoring, most existing laser cutting systems rely solely on height sensors to maintain a constant gap, lacking multi-dimensional, high-precision real-time perception and early warning capabilities for the core cutting state (such as molten pool stability and kerf quality), resulting in a passive, reactive approach of post-processing.

[0004] In summary, existing laser cutting technologies suffer from low efficiency, unstable quality, and excessive reliance on operator experience due to suboptimal path planning, poor parameter adaptation capabilities, and weak status monitoring. Therefore, there is an urgent need for an optimized method that integrates intelligent global path planning, multi-sensor fusion real-time perception, and dynamic adaptive control. Summary of the Invention

[0005] In a first aspect, embodiments of this application provide a method for intelligently optimizing the cutting path and parameters of a laser cutting machine, comprising the following steps: S1. Obtain the model information of the part to be processed, the material type and thickness of the sheet metal, and the target quality requirements, and generate task parameters; S2. Based on the task parameters, query the pre-built process parameter experience model library to generate initial processing parameters. At the same time, adopt an improved genetic-ant colony hybrid algorithm to optimize the cutting path globally with the goal of minimizing the total processing time, and output the theoretically optimal cutting path sequence. S3. Collect image, sound, and temperature field data during the cutting process of the cutting head, and perform preprocessing and feature extraction to generate multi-source feature information; S4. Input multi-source feature information into a pre-trained dual-channel convolutional neural network as a state recognition model to perform feature extraction, feature fusion based on attention mechanism and classification, and identify the current cutting state in real time; S5. Based on the current cutting state, the laser power and cutting speed are dynamically adjusted by a PID fuzzy controller. Based on multi-source feature information, the theoretically optimal cutting path sequence is locally reprogrammed with geometric deviation compensation and dynamic collision avoidance to generate corrected motion control commands.

[0006] Furthermore, the specific steps of step S1 are as follows: S11. Receive the three-dimensional model or two-dimensional layout diagram of the part to be processed input by the user; S12. Obtain the material type, thickness parameters, and target quality requirements of the sheet material; S13. Associate the 3D model or 2D layout diagram, material type, thickness parameters with the target quality requirements to generate structured task parameters.

[0007] Furthermore, the specific steps of step S2 are as follows: S21. Based on the task parameters, query the process parameter experience model library to generate initial processing parameters; S22. Based on the model information of the parts to be processed, the cutting path planning problem is modeled as a variant of the traveling salesman problem, wherein the starting point of each profile to be cut is defined as a node, the node visit sequence is obtained, and the movement cost between nodes is determined based on the idle travel time. S23. An improved genetic-ant colony hybrid algorithm is used to globally optimize the node access sequence, with the goal of minimizing the total empty travel cost, and outputs the theoretically optimal cutting path sequence.

[0008] Furthermore, the specific steps of step S23 are as follows: S231. Execute the genetic algorithm part, initialize the path population, and calculate the fitness of individuals. ,in, This represents the total empty trip time cost; S232. Iteratively optimize through roulette wheel selection, crossover and mutation operations, and retain elite individuals based on individual fitness until the preset first termination condition is met, and output a set of excellent path solutions; S233. Perform pheromone initialization. Use a set of good paths to perform non-uniform pheromone initialization on the algorithm path graph, and assign an enhanced initial pheromone concentration to the edges contained in the solution set. Assign a baseline concentration to the remaining edges. ; S234. Execute the ant colony algorithm part. On the initialized algorithm path graph, the ants construct paths according to the state transition probabilities. The formula for calculating the state transition probability is as follows:

[0009] in, Let be the pheromone concentration on edge (i,j) at time t, initially... The initial pheromone concentration set in step S233; For heuristic information, it is defined as the reciprocal of the distance between nodes i and j; and These are the importance parameters for pheromones and heuristic factors, respectively. Let k be the set of nodes that ant k can currently access; S235. During the ant colony algorithm iteration process, local pheromone updates and global pheromone updates are performed. When it is detected that the global optimal solution has not been updated after multiple iterations, the mutation operator of the genetic algorithm is introduced to mutate the current global optimal path according to a preset probability. S236. Repeat steps S234 to S235 until the second termination condition is met, and output the global historical optimal path as the theoretical optimal cutting path sequence.

[0010] Furthermore, the specific steps of step S3 are as follows: S31. Acquire image data of the cut area through a vision sensor; the vision sensor includes a high-speed CMOS camera, a narrow-band bandpass filter and a cooling protection structure, and performs ROI extraction, non-uniformity correction, filtering and noise reduction and contrast enhancement processing; S32. Acquire sound signals during the cutting process using acoustic sensors and convert them into a spectrogram; S33. Collect temperature field distribution data of the cutting area using a thermal imaging monitoring sensor; S34. The processed image data, acoustic spectrogram, and temperature field data are fused into multi-source feature information.

[0011] Furthermore, the specific steps of step S4 are as follows: S41. Construct a dual-channel convolutional neural network as a state recognition model, wherein the state recognition model includes a visual feature extraction channel, an acoustic feature extraction channel, a feature fusion layer based on an attention mechanism, and a fully connected classifier; S42. Construct a cutting process data set containing more than a set number of materials, thicknesses, and process parameters of different types as a training set, and label the cutting status. S43. Using the training set, with cross-entropy as the loss function, the Adam optimizer is used to supervise the training of the network parameters of the state recognition model until the model converges. S44. Input the multi-source feature information into the corresponding channels of the trained state recognition model respectively; S45. Extract features from the input image and temperature field data through the visual feature extraction channel to obtain a visual feature vector; S46. Extract features from the input spectrogram data through the acoustic feature extraction channel to obtain acoustic feature vectors; S47. Through the feature fusion layer, the visual feature vector and the acoustic feature vector are weighted and fused to obtain the fused feature vector; S48. Input the fused feature vector into the fully connected classifier and output the recognition result of the current cutting state.

[0012] Furthermore, the specific steps of step S5 are as follows: S51. Based on the current cutting status, the laser power and cutting speed are dynamically adjusted using a fuzzy PID controller; S52. Based on multi-source feature information, calculate the lateral deviation between the actual cut position and the theoretical optimal cutting path sequence in real time. and height deviation ; S53. If lateral deviation If the deviation exceeds the preset correction threshold, a compensation vector is calculated based on the current and historical deviations to perform geometric position compensation on subsequent path points in the theoretically optimal cutting path sequence. S54. If the height deviation If the preset safety threshold is exceeded, or if a collision risk is predicted based on temperature field data, a smooth obstacle avoidance path will be replanned within a local area in front of the current cutting point. S55. For the path point sequence processed by step S53 or S54, the S-curve acceleration / deceleration algorithm is used to smooth the trajectory and generate corrected motion control commands.

[0013] Furthermore, it also includes the following steps: S6. Store the final processing parameters, multi-source feature information samples, and final cutting status and quality results used in this processing as new case data into the database, and use incremental learning algorithms to update and optimize the dual-channel convolutional neural network and process parameter experience model library.

[0014] Furthermore, the new case data added in step S6 includes: The laser power, cutting speed, and assist gas pressure parameters ultimately used in this processing were: Multi-source feature information samples and their corresponding timestamps were collected during this processing. The final identified cutting status category and actual quality inspection results for this processing; In step S6, the incremental learning algorithm periodically uses newly added case data to fine-tune the dual-channel convolutional neural network, which serves as the state recognition model, and iteratively optimizes the parameter mapping rules in the process parameter empirical model library.

[0015] Secondly, embodiments of this application also provide a system for intelligently optimizing the cutting path and parameters of a laser cutting machine, comprising: The task preprocessing module obtains the model information of the part to be processed, the material type and thickness of the sheet metal, and the target quality requirements, and generates task parameters. The path planning module, based on task parameters, queries a pre-built empirical model library of process parameters to generate initial processing parameters. At the same time, it uses an improved genetic-ant colony hybrid algorithm to globally optimize the cutting path with the goal of minimizing the total processing time, and outputs the theoretically optimal cutting path sequence. The multi-sensor module collects image, sound, and temperature field data during the cutting process of the cutting head, and performs preprocessing and feature extraction to generate multi-source feature information; The state recognition module inputs multi-source feature information into a pre-trained dual-channel convolutional neural network as a state recognition model to perform feature extraction, feature fusion based on attention mechanism, and classification, and to identify the current cutting state in real time. The real-time control module dynamically adjusts the laser power and cutting speed according to the current cutting status through a PID fuzzy controller, and performs local replanning of the theoretically optimal cutting path sequence based on multi-source feature information, including geometric deviation compensation and dynamic collision avoidance, to generate corrected motion control commands. The knowledge base management module stores the final processing parameters, multi-source feature information samples, and final cutting status and quality results used in this processing as new case data into the database, and uses incremental learning algorithms to update and optimize the dual-channel convolutional neural network and process parameter experience model library.

[0016] As can be seen from the above technical solutions, this application has the following advantages: The method and system for intelligently optimizing cutting paths and parameters in laser cutting machines provided in this application reduce idle travel time, improve cutting efficiency, shorten average processing time, and increase equipment utilization through globally optimized cutting path planning and dynamic adjustment of processing parameters. Based on a state recognition model using multi-sensor data fusion and a dual-channel convolutional neural network, it can monitor the cutting status in real time and dynamically adjust parameters, effectively reducing cutting defects, improving cutting quality stability, and lowering the defect rate. Furthermore, it possesses self-learning capabilities, continuously optimizing the model and process parameter library through incremental learning algorithms, reducing reliance on operator experience, and achieving intelligent production. This also reduces auxiliary time, lowers energy consumption, and reduces overall costs. Attached Figure Description

[0017] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying 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.

[0018] Figure 1 This is a flowchart illustrating the method for intelligently optimizing the cutting path and parameters of a laser cutting machine according to the present invention.

[0019] Figure 2 This is a schematic diagram of the system for intelligently optimizing the cutting path and parameters of the laser cutting machine according to the present invention. Detailed Implementation

[0020] The various embodiments of this disclosure will be described more fully in the following detailed steps of the method for intelligently optimizing the cutting path and parameters of a laser cutting machine. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0021] This embodiment provides a method for intelligently optimizing the cutting path and parameters of a laser cutting machine, which shortens processing time, improves the stability of cutting quality, reduces energy consumption and defect rate, and enhances the level of intelligence.

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 The diagram shows a flowchart of a method for intelligently optimizing the cutting path and parameters of a laser cutting machine in a specific embodiment. The method includes the following steps: S1. Obtain the model information of the part to be processed, the material type and thickness of the sheet metal, and the target quality requirements, and generate task parameters; It should be noted that by receiving the part model, material type, thickness and quality requirements input by the user, structured task parameters are generated, providing accurate basic data for path planning and parameter optimization, ensuring that the laser cutting system can be optimized according to actual needs, and improving versatility and reliability; S2. Based on the task parameters, query the pre-built process parameter experience model library to generate initial processing parameters. At the same time, adopt an improved genetic-ant colony hybrid algorithm to optimize the cutting path globally with the goal of minimizing the total processing time, and output the theoretically optimal cutting path sequence. It should be noted that this step effectively reduces idle travel time, improves cutting efficiency, and provides an optimized basic path for dynamic adjustment; S3. Collect image, sound, and temperature field data during the cutting process of the cutting head, and perform preprocessing and feature extraction to generate multi-source feature information; It should be noted that the fusion of multi-source data provides information support for state recognition, improves the accuracy and reliability of state recognition, and enables the laser cutting system to fully perceive various state changes during the cutting process. S4. Input multi-source feature information into a pre-trained dual-channel convolutional neural network as a state recognition model to perform feature extraction, feature fusion based on attention mechanism and classification, and identify the current cutting state in real time; It should be noted that the dual-channel convolutional neural network combined with the attention mechanism can effectively process multi-source data, improve the accuracy and real-time performance of state recognition, provide a basis for dynamically adjusting processing parameters, and ensure the stability and high quality of the cutting process; S5. Based on the current cutting status, the laser power and cutting speed are dynamically adjusted by the PID fuzzy controller, and based on multi-source feature information, the local replanning of the theoretically optimal cutting path sequence is performed for geometric deviation compensation and dynamic collision avoidance, generating corrected motion control commands. It should be noted that this step can respond to dynamic changes in the cutting process in real time, effectively reduce cutting defects, improve processing quality and equipment safety, and ensure the smooth progress of the cutting process.

[0024] This embodiment solves the problems of low path planning efficiency, inflexible parameter adjustment, and weak monitoring capabilities in the prior art by optimizing the global path, fusing multi-source information, identifying real-time status, and making dynamic adjustments, thus achieving efficient and high-quality laser cutting.

[0025] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process in this embodiment, another method for intelligently optimizing the cutting path and parameters of a laser cutting machine is provided. This method includes the following steps: S1. Obtain the model information of the part to be processed, the material type and thickness of the sheet metal, and the target quality requirements, and generate task parameters; the specific steps of step S1 are as follows: S11. Receive the three-dimensional model or two-dimensional layout diagram of the part to be processed input by the user; S12. Obtain the material type, thickness parameters, and target quality requirements of the sheet material; S13. Associate the 3D model or 2D layout diagram, material type, thickness parameters with the target quality requirements to generate structured task parameters; Specifically, in this embodiment, the part to be processed is a carbon steel bracket required for the production of automotive parts. The user inputs a three-dimensional model of the part, sets the material type of the sheet metal to Q235 carbon steel, the thickness parameter to 8mm, and the target quality requirements to be a surface roughness Ra≤1.6μm and a slag grade ≤1 (no obviously visible slag). Following the process of steps S11-S13, the three-dimensional model, the Q235 carbon steel material type, the 8mm thickness parameter and the set quality requirements are associated to generate structured task parameters containing part contour coordinate information, material thermophysical property parameters, thickness dimension data and quality index thresholds, providing accurate input for subsequent path planning and parameter configuration. S2. Based on the task parameters, query the pre-built empirical model library of process parameters to generate initial processing parameters. Simultaneously, employ an improved genetic-ant colony hybrid algorithm to globally optimize the cutting path with the goal of minimizing the total processing time, outputting the theoretically optimal cutting path sequence. The specific steps of step S2 are as follows: S21. Based on the task parameters, query the process parameter experience model library to generate initial processing parameters; S22. Based on the model information of the parts to be processed, the cutting path planning problem is modeled as a variant of the traveling salesman problem, wherein the starting point of each profile to be cut is defined as a node, the node visit sequence is obtained, and the movement cost between nodes is determined based on the idle travel time. S23. An improved genetic-ant colony hybrid algorithm is used to globally optimize the node visit sequence, aiming to minimize the total empty travel cost, and output the theoretically optimal cutting path sequence; the specific steps of step S23 are as follows: S231. Execute the genetic algorithm part, initialize the path population, and calculate the fitness of individuals. ,in, This represents the total empty trip time cost; S232. Iteratively optimize through roulette wheel selection, crossover and mutation operations, and retain elite individuals according to their fitness until the first preset termination condition is met (e.g., the preset number of genetic algorithm iterations is reached or the quality of the solution no longer significantly improves), and output a set of excellent path solutions; S233. Perform pheromone initialization. Use a set of good paths to perform non-uniform pheromone initialization on the algorithm path graph, and assign an enhanced initial pheromone concentration to the edges contained in the solution set. Assign a baseline concentration to the remaining edges. ; S234. Execute the ant colony algorithm part. On the initialized algorithm path graph, the ants construct paths according to the state transition probabilities. The formula for calculating the state transition probability is as follows:

[0026] in, Let be the pheromone concentration on edge (i,j) at time t, initially... The initial pheromone concentration set in step S233; For heuristic information, it is defined as the reciprocal of the distance between nodes i and j; and These are the importance parameters for pheromones and heuristic factors, respectively. Let k be the set of nodes that ant k can currently access; S235. During the ant colony algorithm iteration process, local pheromone updates and global pheromone updates are performed. When it is detected that the global optimal solution has not been updated after multiple iterations, the mutation operator of the genetic algorithm is introduced to mutate the current global optimal path according to a preset probability; for example, when it has not been updated for 10 consecutive iterations, mutation is performed with a probability of 0.2. After constructing the path, each ant immediately releases pheromones locally along the edges it traverses. The global pheromone update rule is as follows:

[0027] in, The pheromone evaporation coefficient, The pheromone increment is related to the optimal path in this iteration and the global historical optimal path; for example... It is a constant that is inversely proportional to the optimal path length; S236. Repeat steps S234 to S235 until the second termination condition is met (e.g., the maximum number of iterations of the ant colony algorithm is reached or a satisfactory solution is found), and output the global historical optimal path as the theoretical optimal cutting path sequence. Specifically, the process parameter experience model library stores historical successful process data for Q235 carbon steel under different thicknesses (2mm-20mm) and quality requirements. Based on the task parameters generated above, the model library is queried and interpolated to generate initial processing parameters: laser power 3200W, cutting speed 1.2m / min, auxiliary gas (oxygen) pressure 0.6MPa, and pulse frequency 500Hz. In the path planning stage, the 3D model of the part to be processed contains 8 contours to be cut (including external contours and internal hole structures). According to the modeling method in step S22, the optimal starting point of each contour is defined as a node, resulting in a total of 8 nodes. The movement cost between nodes is calculated based on the idle travel time. The idle travel time is determined by the Euclidean distance between nodes and the maximum acceleration (maximum acceleration of X / Y axis 5m / s²) and maximum speed (maximum speed of X / Y axis 80m / min) of each axis of the machine tool. When using the improved genetic-ant colony hybrid algorithm for global optimization, the specific implementation is as follows: The genetic algorithm initializes the path population size to 100. The individual fitness calculation uses the total idle travel time cost as the core indicator, which is the sum of the idle travel times between each node. Individuals with high fitness are selected using a roulette wheel selection method and placed into the mating pool. New offspring are generated using a partial mapping crossover operation, with crossover probability... Adaptive strategies are employed (when population diversity is low). =0.8; when population diversity is high, =0.6), the mutation operation uses exchange mutation, and the mutation probability is 0.6. Similarly, it is adaptive (when population diversity is low), =0.1; when population diversity is high, =0.05), during the iteration process, the 5 best elite individuals of each generation are retained, the first termination condition is preset to 100 iterations, and after the iteration ends, the top 10 excellent path solutions are output; During the pheromone initialization phase, edges contained in the optimal path solution set are assigned an enhanced initial pheromone concentration. =1.5, the remaining edges are assigned a baseline concentration. =0.5; The ant colony algorithm is set to use 50 ants and pheromone factors. =1.2, heuristic factor =2.0, Heuristic Information Defined as the reciprocal of the distance between nodes i and j; The global pheromone update rule is Where ρ=0.2 is the pheromone evaporation coefficient. ( (This is the total empty travel cost of the current globally optimal path); when the globally optimal solution has not been updated for 15 consecutive iterations, the mutation operator of the genetic algorithm is introduced with a probability of 0.2 to mutate the current globally optimal path; The second termination condition is preset to 200 iterations. After the iteration ends, the global historical optimal path is output. S3. Collect image, sound, and temperature field data during the cutting process of the cutting head, and perform preprocessing and feature extraction to generate multi-source feature information; the specific steps of step S3 are as follows: S31. Acquire image data of the cut area through a vision sensor; the vision sensor includes a high-speed CMOS camera, a narrow-band bandpass filter and a cooling protection structure, and performs ROI extraction, non-uniformity correction, filtering and noise reduction and contrast enhancement processing; S32. Acquire sound signals during the cutting process using acoustic sensors and convert them into a spectrogram; S33. Collect temperature field distribution data of the cutting area using a thermal imaging monitoring sensor; S34. The processed image data, acoustic spectrogram, and temperature field data are fused into multi-source feature information; For example, the vision sensor uses a high-speed industrial CMOS camera (parameters: 1200fps frame rate, 1280×1024 pixels resolution, global shutter, high dynamic range), paired with a narrow-band bandpass filter with a center wavelength of 777nm (matching the plasma characteristic spectrum when oxygen cuts carbon steel) and a water-cooled protective cover, installed at a 35° angle to the side and rear of the cutting head, with clean compressed air introduced to form positive pressure protection; during image preprocessing, a 200×200 pixel ROI area is first dynamically defined according to the real-time position of the cutting head, then non-uniformity correction is performed to compensate for the difference in sensor pixel response, median filtering (window size 3×3) is used to remove salt-and-pepper noise, and finally adaptive histogram equalization is used to enhance the contrast between the plasma area and the background; The acoustic sensor uses a high-sensitivity microphone to collect sound signals during the cutting process. The sampling frequency is 44.1kHz, and the sound signals are converted into a 128×128 pixel spectrogram through short-time Fourier transform. The thermal imaging monitoring sensor uses an infrared thermal imager with a temperature measurement range of -20℃ to 1500℃, a resolution of 320×240 pixels, and a sampling frequency of 50Hz to acquire real-time temperature field distribution data of the cutting area. The preprocessed image data (200×200 pixels), spectrogram (128×128 pixels), and temperature field data (320×240 pixels, pseudo-color mapped to RGB format) are aligned according to timestamps and fused into multi-source feature information; S4. Input multi-source feature information into a pre-trained dual-channel convolutional neural network as a state recognition model to perform feature extraction, attention-based feature fusion, and classification, and identify the current cutting state in real time; the specific steps of step S4 are as follows: S41. Construct a dual-channel convolutional neural network as a state recognition model, wherein the state recognition model includes a visual feature extraction channel, an acoustic feature extraction channel, a feature fusion layer based on an attention mechanism, and a fully connected classifier; For example, the visual feature extraction channel is implemented using a ResNet-18 backbone network to extract a 512-dimensional visual feature vector V; the acoustic feature extraction channel is implemented using a one-dimensional convolutional neural network to extract a 256-dimensional acoustic feature vector A. S42. Construct a cutting process data set containing more than a set number of materials, thicknesses, and process parameters of different types as a training set, and label the cutting status. S43. Using the training set, with cross-entropy as the loss function, the Adam optimizer is used to supervise the training of the network parameters of the state recognition model until the model converges. S44. Input the multi-source feature information into the corresponding channels of the trained state recognition model respectively; S45. Extract features from the input image and temperature field data through the visual feature extraction channel to obtain a visual feature vector; S46. Extract features from the input spectrogram data through the acoustic feature extraction channel to obtain acoustic feature vectors; S47. Through the feature fusion layer, the visual feature vector and the acoustic feature vector are weighted and fused to obtain the fused feature vector; Specifically, the visual feature vector V is concatenated with the acoustic feature vector A to obtain the feature vector before fusion. ; Calculate attention weights ,in, This is the weight matrix. It is the bias vector; Weighted fusion is performed based on attention weights to obtain the fused feature vector. ; S48. Input the fused feature vector into the fully connected classifier and output the recognition result of the current cutting state. Specifically, output the class probability of the current cutting state through the Softmax function. For example, the specific architecture of the dual-channel convolutional neural network is as follows: the visual feature extraction channel adopts a ResNet-18 backbone network (initialized by transfer learning, replacing the last fully connected layer); the acoustic feature extraction channel is a lightweight one-dimensional CNN (containing 4 convolutional layers and 4 max pooling layers); the feature fusion layer is a weighted fusion structure based on the attention mechanism; the fully connected classifier contains 3 fully connected layers (512-dimensional → 256-dimensional → 6-dimensional). The training set includes six materials such as Q235 carbon steel, stainless steel, and aluminum alloy, with a thickness range of 2mm-20mm. It covers cutting process data with different laser power, cutting speed and other process parameters, totaling more than 100,000 samples. The labeled cutting status includes six categories: normal cutting, piercing, about to penetrate, overheated, not cut through, and other abnormalities. When training the model using the training set, cross-entropy was used as the loss function, and the Adam optimizer was employed (learning rate 0.001, decay coefficients β1=0.9, β2=0.999). The model converged after 200 iterations of training. The multi-source feature information generated in step S3 is input into the corresponding channels, wherein the visual feature extraction channel is input with the fused image of image data and temperature field mapping data (224×224 pixels), and the acoustic feature extraction channel is input with the spectrogram data. The visual feature extraction channel outputs a 512-dimensional visual feature vector through convolution and pooling operations of ResNet-18. This vector contains key information such as plasma morphology and temperature distribution. The acoustic feature extraction channel outputs a 256-dimensional acoustic feature vector through convolution and pooling operations of a one-dimensional CNN. This vector contains information such as the frequency distribution and energy changes of the sound during the cutting process. During the feature fusion process, the 512-dimensional visual feature vector V and the 256-dimensional acoustic feature vector A are first concatenated to form a 768-dimensional feature vector before fusion. Attention weights are calculated using a single-layer neural network (weight matrix W, dimension 768×2; bias vector b, dimension 2×1) and the Softmax function. In this embodiment, when identifying the "overheated" state, =0.82, =0.18; When identifying the "not cut through" state, =0.75, =0.25; then through Obtain the fused feature vector; The fused feature vectors are input into a fully connected classifier, and the output state probability distribution is obtained through the Softmax activation function. The single inference time is 42ms, which meets the real-time requirements. In this embodiment, when the cutting process has been going on for 12 seconds, the model outputs a probability distribution of [0.03, 0.01, 0.02, 0.92, 0.01, 0.01], and determines the current cutting state as "overburning". S5. Based on the current cutting state, dynamically adjust the laser power and cutting speed using a PID fuzzy controller, and based on multi-source feature information, perform local replanning for geometric deviation compensation and dynamic collision avoidance on the theoretically optimal cutting path sequence to generate corrected motion control commands; the specific steps of step S5 are as follows: S51. Based on the current cutting status, the laser power and cutting speed are dynamically adjusted using a fuzzy PID controller; S52. Based on multi-source feature information, calculate the lateral deviation between the actual cut position and the theoretical optimal cutting path sequence in real time. and height deviation ; S53. If lateral deviation If the deviation exceeds the preset correction threshold, a compensation vector is calculated based on the current and historical deviations to perform geometric position compensation on subsequent path points in the theoretically optimal cutting path sequence. S54. If the height deviation If the preset safety threshold is exceeded, or if a collision risk is predicted based on temperature field data, a smooth obstacle avoidance path will be replanned within a local area in front of the current cutting point. S55. For the path point sequence processed by step S53 or S54, the S-curve acceleration and deceleration algorithm is used to smooth the trajectory and generate the corrected motion control command. For example, based on the "overheating" state identification result, the following dynamic optimization and real-time control operations are performed: The fuzzy PID controller uses the "overburning" state deviation (with the normal cutting state as the benchmark and the deviation value quantized as 0.8) as input to dynamically adjust the process parameters: the laser power is reduced from 3200W to 2900W, and the cutting speed is increased from 1.2m / min to 1.35m / min. The parameter adjustment range is controlled within ±10% to ensure system stability. Based on visual sensor data, the actual cut centerline is identified and compared with the theoretical optimal cutting path sequence. The current lateral deviation ΔX = 0.3mm, ΔY = 0.2mm, and height deviation ΔZ = 0.1mm are calculated. The preset correction threshold is ±0.25mm, and the safety threshold is ±0.3mm. The current lateral deviation ΔX exceeds the correction threshold. A Kalman filter is used to calculate the compensation vector based on the deviation data of the most recent 5 sampling periods (sampling period 50ms). =(-0.28mm,-0.18mm), perform geometric position compensation on the subsequent 100 path points in the theoretical path to obtain the corrected path point coordinates; Based on thermal imaging data, there is no risk of collision within the range of the next 20 path points, and no dynamic obstacle avoidance planning is required. An S-curve acceleration / deceleration algorithm is used to smooth the trajectory of the compensated path point sequence. The maximum acceleration is set to 5 m / s² and the jerk is set to 100 m / s³. Continuous and smooth motion control commands are generated to drive the machine tool to perform the cutting operation.

[0028] During the subsequent cutting process, when the vision sensor detected a lateral deviation of ΔX=0.15mm and ΔY=0.1mm (both within the correction threshold), the thermal imager data showed that the highest temperature in the cutting area dropped from 850℃ to 720℃, and the state recognition model output probability distribution was [0.94,0.01,0.02,0.01,0.01,0.01], it was determined that the current state had recovered to "normal cutting". The fuzzy PID controller adjusted the laser power back to 3100W and the cutting speed back to 1.25m / min. S6. Store the final processing parameters, multi-source feature information samples, and final cutting status and quality results used in this processing as new case data into the database, and use incremental learning algorithms to update and optimize the dual-channel convolutional neural network and process parameter experience model library. The new case data added in step S6 includes: The laser power, cutting speed, and assist gas pressure parameters ultimately used in this processing were: Multi-source feature information samples and their corresponding timestamps were collected during this processing. The final identified cutting status category and actual quality inspection results for this processing; In step S6, the incremental learning algorithm periodically uses newly added case data to fine-tune the dual-channel convolutional neural network, which serves as the state recognition model, and iteratively optimizes the parameter mapping rules in the process parameter empirical model library. Specifically, the newly added case data processed this time includes: Final processing parameters: laser power (3200W→2900W→3100W), cutting speed (1.2m / min→1.35m / min→1.25m / min), auxiliary gas pressure 0.6MPa, pulse frequency 500Hz, and timestamps corresponding to each parameter; Multi-source feature information samples: including 1200 sets of preprocessed image data, spectrogram data and temperature field data, each sample with a corresponding timestamp; Cutting status category and actual quality inspection results: "Overheating" was identified once during the cutting process, and "normal cutting" accounted for 98%. The final quality inspection results were: surface roughness Ra=1.4μm, slag level 0, which met the target quality requirements. The incremental learning algorithm employs a periodic update strategy, with an update cycle of 10 processing tasks. After merging the newly added case data with historical data, the dual-channel convolutional neural network was fine-tuned and trained (learning rate 0.0001, 50 iterations). Simultaneously, the parameter mapping rules for 8mm thick Q235 carbon steel in the process parameter experience model library were iteratively optimized. The updated model library recommends an initial laser power of 3150W and a cutting speed of 1.22m / min under the same processing conditions, improving processing quality stability.

[0029] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0030] like Figure 2 As shown, the following are embodiments of a system for intelligently optimizing the cutting path and parameters of a laser cutting machine provided in this disclosure. This system and the methods for intelligently optimizing the cutting path and parameters of a laser cutting machine in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the system for intelligently optimizing the cutting path and parameters of a laser cutting machine, please refer to the embodiments of the methods for intelligently optimizing the cutting path and parameters of a laser cutting machine described above.

[0031] The system includes: The task preprocessing module obtains the model information of the part to be processed, the material type and thickness of the sheet metal, and the target quality requirements, and generates task parameters. The path planning module, based on task parameters, queries a pre-built empirical model library of process parameters to generate initial processing parameters. At the same time, it uses an improved genetic-ant colony hybrid algorithm to globally optimize the cutting path with the goal of minimizing the total processing time, and outputs the theoretically optimal cutting path sequence. The multi-sensor module collects image, sound, and temperature field data during the cutting process of the cutting head, and performs preprocessing and feature extraction to generate multi-source feature information; The state recognition module inputs multi-source feature information into a pre-trained dual-channel convolutional neural network as a state recognition model to perform feature extraction, feature fusion based on attention mechanism, and classification, and to identify the current cutting state in real time. The real-time control module dynamically adjusts the laser power and cutting speed according to the current cutting status through a PID fuzzy controller. Based on multi-source feature information, it performs local replanning of the theoretically optimal cutting path sequence for geometric deviation compensation and dynamic collision avoidance, and generates corrected motion control commands.

[0032] This embodiment achieves improved cutting efficiency, optimized processing quality, reduced production costs, and intelligent production through the interactive collaboration of the task preprocessing module, path planning module, multi-sensor module, status recognition module, and real-time control module.

[0033] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent optimization of cutting path and parameters for a laser cutting machine, characterized in that, Comprising the following steps: S1. Obtain the model information of the part to be processed, the material type and thickness of the plate, and the target quality requirement, and generate task parameters; S2. Based on the task parameters, query the pre-constructed process parameter experience model library to generate initial processing parameters, and at the same time, adopt an improved genetic-ant colony hybrid algorithm to globally optimize the cutting path with the shortest total processing time as the target, and output the theoretically optimal cutting path sequence; S3. Collect image, sound and temperature field data during the cutting process of the cutting head, and perform preprocessing and feature extraction to generate multi-source feature information; S4. Input the multi-source feature information into the pre-trained dual-channel convolutional neural network as a state recognition model, perform feature extraction, feature fusion based on attention mechanism, and classification, and identify the current cutting state in real time; S5. According to the current cutting state, dynamically adjust the laser power and cutting speed through the PID fuzzy controller, and based on the multi-source feature information, perform local re-planning of geometric deviation compensation and dynamic collision avoidance on the theoretically optimal cutting path sequence, and generate corrected motion control instructions.

2. The method of claim 1, wherein, Step S1 has the following specific steps: S11. Receive the three-dimensional model or two-dimensional layout drawing of the part to be processed input by the user; S12. Obtain the material type, thickness parameters and target quality requirements of the plate; S13. Associate the three-dimensional model or two-dimensional layout drawing, material type, thickness parameters and target quality requirements to generate structured task parameters.

3. The method of claim 1, wherein, Step S2 has the following specific steps: S21. Based on the task parameters, query the process parameter experience model library to generate initial processing parameters; S22. Based on the model information of the part to be processed, model the cutting path planning problem as a variant of the traveling salesman problem, where the starting point of each cutting contour is defined as a node to obtain a node access sequence, and the movement cost between nodes is determined based on the empty travel time; S23. Adopt an improved genetic-ant colony hybrid algorithm to globally optimize the node access sequence with the minimum total empty travel cost as the target, and output the theoretically optimal cutting path sequence.

4. The method of claim 3, wherein, Step S23 has the following specific steps: S231. Perform genetic algorithm part, initialize path population, calculate individual fitness wherein, is the total empty travel time cost; S232. Optimize through roulette selection, crossover and mutation operations, and reserve elite individuals according to individual fitness until the first preset termination condition is met, and output a set of good path solutions; S233. Perform pheromone initialization, use the good path solution set to perform non-uniform pheromone initialization on the algorithm path graph, assign enhanced initial pheromone concentration to the edges contained in the solution set , assign a reference concentration to the remaining edges ; S234. Perform the ant colony algorithm part, and ants construct paths according to the state transition probability on the initialized algorithm path graph; The calculation formula of the state transition probability is as follows: wherein, is the pheromone concentration on edge (i, j) at time t, initially is the initial pheromone concentration set in step S233; is the heuristic information defined as the inverse of the distance between nodes i and j; and are the importance parameters of pheromone and heuristic factor, respectively; is the set of nodes currently accessible by ant k; S235. During the iteration process of the ant colony algorithm, perform local pheromone update and global pheromone update, and when it is detected that the global optimal solution has not been updated for multiple iterations, introduce the mutation operator of the genetic algorithm to mutate the current global optimal path according to a preset probability; S236. Repeat steps S234 to S235 until the second termination condition is met, and output the global historical optimal path as the theoretically optimal cutting path sequence.

5. The method of claim 1, wherein, Step S3 has the following specific steps: S31. Acquire image data of the cut area through a vision sensor; the vision sensor includes a high-speed CMOS camera, a narrow-band bandpass filter and a cooling protection structure, and performs ROI extraction, non-uniformity correction, filtering and noise reduction and contrast enhancement processing; S32. Acquire sound signals during the cutting process using acoustic sensors and convert them into a spectrogram; S33. Collect temperature field distribution data of the cutting area using a thermal imaging monitoring sensor; S34. The processed image data, acoustic spectrogram, and temperature field data are fused into multi-source feature information.

6. The method of claim 1, wherein, The specific steps of step S4 are as follows: S41. Construct a dual-channel convolutional neural network as a state recognition model, wherein the state recognition model includes a visual feature extraction channel, an acoustic feature extraction channel, a feature fusion layer based on an attention mechanism, and a fully connected classifier; S42. Construct a cutting process data set containing more than a set number of materials, thicknesses, and process parameters of different types as a training set, and label the cutting status. S43. Using the training set, with cross-entropy as the loss function, the Adam optimizer is used to supervise the training of the network parameters of the state recognition model until the model converges. S44. Input the multi-source feature information into the corresponding channels of the trained state recognition model respectively; S45. Extract features from the input image and temperature field data through the visual feature extraction channel to obtain a visual feature vector; S46. Extract features from the input spectrogram data through the acoustic feature extraction channel to obtain acoustic feature vectors; S47. Through the feature fusion layer, the visual feature vector and the acoustic feature vector are weighted and fused to obtain the fused feature vector; S48. Input the fused feature vector into the fully connected classifier and output the recognition result of the current cutting state.

7. The method of claim 1, wherein, The specific steps of step S5 are as follows: S51. Based on the current cutting status, the laser power and cutting speed are dynamically adjusted using a fuzzy PID controller; S52. Based on the multi-source feature information, the lateral deviation between the actual cutting seam position and the theoretical optimal cutting path sequence is calculated in real time and height deviation ; S53. If the lateral deviation exceeds a preset correction threshold, a compensation vector is calculated based on the current and historical deviations, and the subsequent path points in the theoretical optimal cutting path sequence are compensated in geometric position. S54. If the height deviation If the height deviation exceeds a preset safety threshold, or if a collision risk is predicted based on the temperature field data, a smooth obstacle avoidance path is re-planned within a local range in front of the current cutting point; S55. For the path point sequence processed by step S53 or S54, the S-curve acceleration / deceleration algorithm is used to smooth the trajectory and generate corrected motion control commands.

8. The method of claim 1, wherein, It also includes the following steps: S6. Store the final processing parameters, multi-source feature information samples, and final cutting status and quality results used in this processing as new case data into the database, and use incremental learning algorithms to update and optimize the dual-channel convolutional neural network and process parameter experience model library.

9. The method of claim 8, wherein, The new case data added in step S6 includes: The laser power, cutting speed, and assist gas pressure parameters ultimately used in this processing were: Multi-source feature information samples and their corresponding timestamps were collected during this processing. The final identified cutting status category and actual quality inspection results for this processing; In step S6, the incremental learning algorithm periodically uses newly added case data to fine-tune the dual-channel convolutional neural network, which serves as the state recognition model, and iteratively optimizes the parameter mapping rules in the process parameter empirical model library.

10. A system for intelligent optimization of cutting path and parameters for a laser cutting machine, characterized in that, include: The task preprocessing module obtains the model information of the part to be processed, the material type and thickness of the sheet metal, and the target quality requirements, and generates task parameters. The path planning module, based on task parameters, queries a pre-built empirical model library of process parameters to generate initial processing parameters. At the same time, it uses an improved genetic-ant colony hybrid algorithm to globally optimize the cutting path with the goal of minimizing the total processing time, and outputs the theoretically optimal cutting path sequence. The multi-sensor module collects image, sound, and temperature field data during the cutting process of the cutting head, and performs preprocessing and feature extraction to generate multi-source feature information; The state recognition module inputs multi-source feature information into a pre-trained dual-channel convolutional neural network as a state recognition model to perform feature extraction, feature fusion based on attention mechanism, and classification, and to identify the current cutting state in real time. The real-time control module dynamically adjusts the laser power and cutting speed according to the current cutting status through a PID fuzzy controller. Based on multi-source feature information, it performs local replanning of the theoretically optimal cutting path sequence for geometric deviation compensation and dynamic collision avoidance, and generates corrected motion control commands.

Images