A three-dimensional cutting dynamic parameter adjusting system and method based on deep learning

Abstract

The application discloses a three-dimensional cutting dynamic parameter adjusting system and method based on deep learning, relates to the application technical field of three-dimensional five-axis optical fiber laser cutting machines in the automobile manufacturing industry, and comprises a multi-modal sensor cluster, an edge computing unit, a closed-loop control module and an AI model library. The multi-modal sensor cluster collects cutting data. The edge computing unit receives multi-modal sensor data and outputs dynamic adjusting parameters. The closed-loop control module transmits the dynamic parameters to a laser generator and a five-axis motion controller through a high-speed communication bus, forming a 'perception-decision-execution' closed loop. The application solves the problems of artificial experience dependence and dynamic response lag by designing multi-modal sensor cooperative decision-making, real-time perception of cutting state and risk, reducing the defective rate and improving the response speed.

Description

TECHNICAL FIELD

[0001] The application relates to the application technology field of a three-dimensional five-axis optical fiber laser cutting machine in the automobile manufacturing industry, in particular to a three-dimensional cutting dynamic parameter adjustment system and method based on deep learning. BACKGROUND

[0002] With the development of the automobile industry towards light weight and individualization, the precision and efficiency requirements of laser cutting for body-in-white manufacturing are becoming higher and higher. The parameter adjustment of a traditional three-dimensional five-axis optical fiber laser cutting machine mainly depends on artificial experience, and it is difficult to adapt to the dynamic change requirements in the cutting process of a complex curved surface.

[0003] Patent CN110428506B discloses a dynamic three-dimensional geometric figure cutting implementation method based on parameters, which realizes the expansion of a cutting sequence, including parallel cutting and serial cutting, for user selection, is more persuasive, and meets the teaching needs.

[0004] The above patent uses multiple cutting surfaces to cut a three-dimensional figure, and dynamically controls the cutting rate through parameters to achieve a dynamic cutting effect, but in the automobile manufacturing industry, the cutting parameter adjustment mainly depends on artificial experience, and it is difficult to adapt to the dynamic change requirements of complex processing.

[0005] Therefore, the application provides a three-dimensional cutting dynamic parameter adjustment system and method based on deep learning, which realizes real-time sensing of a cutting state and risks, solves the problems of artificial experience dependence and dynamic response lag. SUMMARY

[0006] The application aims to provide a three-dimensional cutting dynamic parameter adjustment system and method based on deep learning, to solve the technical problem of being difficult to adapt to the dynamic change requirements in the cutting process of a complex curved surface due to dependence on artificial experience.

[0007] To achieve the above-mentioned purpose, the application provides the following technical scheme: a three-dimensional cutting dynamic parameter adjustment system based on deep learning, comprising a multi-modal sensor cluster, an edge computing unit, a closed-loop control module and an AI model library, the multi-modal sensor cluster collects cutting data, the edge computing unit receives multi-modal sensor data and outputs dynamic adjustment parameters, the closed-loop control module transmits the dynamic parameters to a laser generator and a five-axis motion controller through a high-speed communication bus, and forms a “perception-decision-execution” closed loop.

[0008] The edge computing unit is built-in with a deep learning model, and inputs include multi-modal sensor data, workpiece three-dimensional geometric parameters and material properties, and outputs dynamic adjustment parameters: laser power, cutting speed, nozzle height and focal point position.

[0009] The AI model library stores historical cutting data, abnormal scenarios and optimization strategies, supports online model updating and transfer learning.

[0010] Preferably, the multi-modal sensor cluster comprises:

[0011] A plasma detection sensor is used to monitor the laser scattering spectrum signal during cutting, and to determine missed cutting and non-penetration by spectral intensity threshold;

[0012] A capacitive sensing height adjustment module, composed of a ceramic nozzle, an amplifier board and a cable, dynamically calculates the nozzle height deviation by detecting the change in the capacitance value between the nozzle and the cutting workpiece;

[0013] A magnetic attraction anti-collision module, including upper and lower magnetic attraction contacts and an anti-collision detection circuit, triggers the device to stop and re-plan the path when the contacts are separated by more than 0.5mm due to external force.

[0014] Preferably, the deep learning model adopts a spatio-temporal dual-stream network architecture, including:

[0015] A temporal stream network: based on LSTM units to process sensor time series data and predict cutting stability;

[0016] A spatial stream network: based on a 3D convolutional neural network to analyze workpiece geometric features and generate spatial cutting parameter suggestions;

[0017] A feature fusion layer: dynamically weights spatio-temporal features through an attention mechanism and outputs comprehensive parameter adjustment instructions;

[0018] A reinforcement learning optimization module: constructs a multi-objective reward function based on cutting efficiency, material utilization and defect rate, and optimizes model weights online through PPO algorithm.

[0019] Preferably, the magnetic attraction anti-collision module further comprises:

[0020] A collision prediction algorithm: based on five-axis motion trajectory and workpiece three-dimensional point cloud data, predicts the position offset of the cutting head within the next 5ms through Kalman filtering, and triggers path re-planning in advance if the offset exceeds the safety threshold;

[0021] A multi-level response mechanism:

[0022] First-level response: when there is a slight collision, only adjust the cutting speed to a safe level;

[0023] Second-level response: when there is a serious collision, stop and start A* algorithm to dynamically generate an obstacle-avoiding path;

[0024] Collision scenario learning module: uploads abnormal data to the AI model library to generate an anti-collision rule library for active avoidance during similar workpiece cutting.

[0025] Preferably, the adjustment logic of the capacitive sensing height adjustment module comprises:

[0026] Dynamic compensation algorithm: dynamically adjust the amplification multiple of the capacitive signal according to the surface roughness of the workpiece and the electrical conductivity of the material, to ensure that the height detection error is ≤0.03mm;

[0027] Adaptive filtering technology: use wavelet transform to filter out high-frequency noise and retain effective capacitive signals;

[0028] Double closed-loop control:

[0029] Inner loop: real-time adjustment of nozzle height based on PID controller;

[0030] Outer loop: predict the thermal deformation of the workpiece through a deep learning model and compensate for the height deviation in advance.

[0031] Preferably, the edge computing unit integrates the following model update mechanism:

[0032] Incremental learning module: automatically extract key data including cutting quality score and sensor anomaly mark after each cutting task is completed, update model parameters through online gradient descent method, and the update period is ≤3 minutes;

[0033] Model version management: keep historical model versions, and automatically roll back to the optimal version when new data causes performance (poor rate) to rise by >0.1%;

[0034] Federal learning interface: support encrypted sharing of local model parameters between multiple devices to build a globally optimized model.

[0035] Preferably, the parameter adjustment system further integrates a visual detection subsystem, which comprises:

[0036] High-speed industrial camera to capture infrared thermal imaging and visible light images of the cutting area;

[0037] Image fusion algorithm: superimpose thermal imaging data and visible light images, and segment the cutting defect area through a U-Net network;

[0038] Multi-sensor collaborative decision-making:

[0039] Trigger emergency stop when the magnetic anti-collision module detects collision risk and the visual detection subsystem simultaneously identifies path obstacles;

[0040] When only the visual detection subsystem detects local overheating, automatically reduce the laser power by 10%-20%.

[0041] Preferably, the deep learning model is embedded with an energy consumption optimization module, which specifically comprises:

[0042] Material thermal property analysis unit: calculate the minimum energy threshold based on material thermal conductivity and specific heat capacity to ensure the cutting depth meets the standard;

[0043] Laser pulse optimization strategy: change continuous laser to adaptive pulse mode, dynamically adjust pulse frequency and duty cycle by model to reduce energy consumption by ≥15%;

[0044] Waste recycling feedback: monitor waste quality in real time through weighing sensors, optimize cutting path to reduce scrap, and improve material utilization by ≥8%.

[0045] Preferably, the AI model library supports the following functions:

[0046] Knowledge distillation technology: extract expert experience and historical data into lightweight rule-based models;

[0047] Geometric feature matching engine: compare the curvature distribution of new workpieces and model library workpieces through point cloud registration algorithm, automatically recommend similar process parameters, and adapt within 2 minutes;

[0048] Cross-material transfer learning: if the material of the new workpiece is not in the library, map it to the existing parameter combination based on thermal physical property similarity.

[0049] Preferably, the adjustment method comprises the following steps:

[0050] S1, data acquisition stage: real-time acquisition of plasma spectrum, capacitance value, magnetic attraction touch point state, workpiece geometric data and environment temperature through multi-modal sensor cluster, and acquisition of cutting area thermal imaging and topographic image through visual detection subsystem;

[0051] S2, model inference stage: input the collected data into the space-time double-flow network, output the dynamic parameter combination of laser power, cutting speed, nozzle height and focal point position, and generate pulse laser parameters through energy consumption optimization module;

[0052] S3, execution and feedback stage: closed-loop control module drives laser generator and five-axis motion mechanism to execute parameter adjustment, updates model weight through incremental learning module according to cutting quality and energy consumption data;

[0053] S4, abnormal processing stage: if collision risk or parameter overrun is detected, trigger multi-level response mechanism and record abnormal scene to AI model library;

[0054] S5: cloud collaborative optimization: upload local model parameters and abnormal data to the cloud after encryption, generate a global optimization model through federated learning, and distribute it to all devices;

[0055] S6, process chain traceability: generate a unique ID for each cutting task, associate parameter settings, sensor data and quality reports, and support full life cycle quality analysis.

[0056] Compared with the prior art, the present application has the following advantages:

[0057] 1. The present application solves the problems of artificial experience dependence and dynamic response lag by designing a multi-modal sensor cooperative decision-making system that can perceive cutting state and risk in real time, reduces the defect rate and improves response speed;

[0058] 2. The present application solves the problem of complex working condition parameter matching difficulty by designing a space-time double-flow network architecture to achieve dynamic parameter accurate prediction and adaptive adjustment, improves cutting efficiency and reduces energy consumption;

[0059] 3. The present application solves the problems of long new process development cycle and high debugging cost by designing a free process parameter library to realize the function of quickly adapting to new workpieces and materials, reduces the debugging time and reduces the material waste rate;

[0060] 4. The present application solves the problems of high energy consumption and low material utilization rate of traditional continuous laser by designing a green energy consumption optimization closed loop to realize energy saving and waste recovery cooperation, improves material utilization rate and reduces energy consumption. BRIEF DESCRIPTION OF DRAWINGS

[0061] Figure 1 is a parameter adjustment system framework schematic diagram of the present application;

[0062] Figure 2 is a parameter adjustment method schematic diagram of the present application. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0064] Please refer to Figure 1 and Figure 2 , the present application provides an embodiment: a three-dimensional cutting dynamic parameter adjustment system based on deep learning, including a multi-modal sensor cluster, an edge computing unit, a closed loop control module and an AI model library, the multi-modal sensor cluster collects cutting data, the edge computing unit receives multi-modal sensor data and outputs dynamic adjustment parameters, the closed loop control module transmits dynamic parameters to laser generator and five-axis motion controller through high-speed communication bus, forming a "perception-decision-execution" closed loop;

[0065] The edge computing unit is built-in with a deep learning model, which inputs multi-modal sensor data, workpiece three-dimensional geometric parameters and material attributes, and outputs dynamic adjustment parameters: laser power, cutting speed, nozzle height and focal point position.

[0066] The AI model library stores historical cutting data, abnormal scenarios and optimization strategies, supports online model updating and transfer learning.

[0067] The edge computing unit integrates the following model updating mechanisms:

[0068] Incremental learning module: after completing each cutting task, key data including cutting quality score and sensor anomaly label are automatically extracted, model parameters are updated through online gradient descent method, and the update cycle is ≤3 minutes;

[0069] Model version management: historical model versions are retained, and when new data causes performance (poor rate) to rise >0.1%, it is automatically rolled back to the optimal version;

[0070] Federal learning interface: supports encrypted sharing of local model parameters between multiple devices, and builds a global optimization model;

[0071] Further, data acquisition and input: sensor data: plasma spectrum intensity: real-time monitoring value 5200 lux; capacitance value: monitoring current nozzle height deviation +0.08mm; magnetic attraction contact state: closed; working parameters: three-dimensional geometric data: STL file analysis surface curvature; material properties: aluminum alloy 5052; deep learning model inference: spatio-temporal double-flow network processing: time series flow: analyze 10 consecutive frames of capacitance value data, predict the future 0.5s height deviation trend +0.12mm; spatial flow: identify high-curvature areas of the surface, and suggest reducing the cutting speed by 15% to avoid heat accumulation; feature fusion: weight spatio-temporal features through attention mechanism, generate parameter instructions: laser power from 500W to 530W, cutting speed from 4.0m / min to 3.4m / min, nozzle height from 2.0mm to 1.92mm; closed-loop execution and feedback: after the laser generator receives the instruction, it switches to 530W mode within 0.5ms, the cutting head moves along the optimized path, the speed is reduced to 3.4m / min, and the height of the nozzle is dynamically maintained at 1.92mm by the servo height adjustment device. After cutting is completed, the visual detection subsystem detects the cross-section roughness Ra=6.3um, which is determined to be qualified.

[0072] Incremental learning module: key data extraction: the cutting quality score in this task is 92, and there is no abnormal mark for the sensor; online gradient descent: the model updates the weights of the fully connected layer according to the score, and the update time is 2 minutes; model version management: version rollback logic: if the cutting failure rate is greater than 0.5% for 3 times in a row after updating, automatically switch to the historical optimal version; federal learning interface: multi-device cooperation: share local model parameters with 10 state devices in the factory, aggregate globally in the cloud to generate a global model, and distribute it to each device. The global model shortens the new workpiece adaptation time from 15 minutes to 7 minutes.

[0073] See Figure 1 and Figure 2 An embodiment provided by the application: a three-dimensional cutting dynamic parameter adjustment system based on deep learning, the multi-modal sensor cluster includes:

[0074] Plasma detection sensor, used to monitor the laser scattering spectrum signal in the cutting process, and judge the under-cutting and non-penetration through the spectrum intensity threshold;

[0075] Capacitive sensing height adjustment module, composed of ceramic nozzle, amplifier board and cable, dynamically calculates the nozzle height deviation by detecting the capacitance value change between the nozzle and the cutting workpiece;

[0076] Magnetic attraction anti-collision module, including upper and lower magnetic attraction contacts and anti-collision detection circuit, when the contact is separated by external force more than 0.5mm, triggering device shutdown and path re-planning instruction;

[0077] The magnetic attraction anti-collision module further includes:

[0078] Collision prediction algorithm: based on five-axis motion trajectory and workpiece three-dimensional point cloud data, through Kalman filtering to predict the position deviation of the cutting head within 5ms in the future, if the deviation exceeds the safety threshold, the path re-planning is triggered in advance;

[0079] Multi-level response mechanism:

[0080] Primary response: when there is a slight collision, only adjust the cutting speed to a safe level;

[0081] Secondary response: when there is a serious collision, stop and start A* algorithm to dynamically generate an obstacle avoidance path;

[0082] Collision scenario learning module: upload abnormal data to the AI model library to generate an anti-collision rule library for active avoidance during similar workpiece cutting;

[0083] The adjustment logic of the capacitive sensing height adjustment module includes:

[0084] Dynamic compensation algorithm: dynamically adjust the capacitance signal amplification multiple according to the workpiece surface roughness and material conductivity to ensure that the height detection error is less than or equal to 0.03mm;

[0085] Adaptive filtering technology: wavelet transform is used to filter out high-frequency noise and retain effective capacitance signals;

[0086] Double closed-loop control:

[0087] Inner loop: real-time adjustment of nozzle height based on PID controller;

[0088] Outer loop: predict the thermal deformation of the workpiece through a deep learning model, and compensate for the height deviation in advance;

[0089] Further, the plasma detection sensor monitors the laser scattering spectrum signal in real time during the cutting process to judge the cutting quality. During the cutting process, the laser interacts with the material to generate plasma. The plasma detection sensor collects the spectrum signal at a sampling frequency of 1 kHz. When the signal is detected to be lower than the threshold, it is determined as "not penetrated". If the signal fluctuates continuously beyond the limit, it is marked as "missed cutting". If an abnormality is found, the system immediately reduces the cutting speed to a safe value, and starts the process parameter optimization through the edge computing unit;

[0090] The capacitance type induction height adjustment module dynamically adjusts the nozzle height to avoid scratching or distance deviation caused by workpiece deformation or surface unevenness. The ceramic nozzle forms a capacitor with the workpiece surface, and the amplifier board detects the capacitance value change with a resolution of 0.01 pF. If the workpiece surface roughness Ra value is high, the amplification factor is automatically increased to 50 times to enhance the signal sensitivity. For non-conductive materials such as composite materials, a low-frequency excitation signal of 1 kHz is used to reduce interference. The PID control adjusts the nozzle height in real time, with a response time ≤5 ms, ensuring that the height error ≤0.03 mm. The deep learning model analyzes the thermal expansion coefficient of the workpiece to predict the height deviation within the next 2 seconds and compensate in advance;

[0091] The magnetic attraction anti-collision module detects the collision risk and triggers the avoidance strategy to ensure the safety of the equipment. Based on the five-axis motion trajectory and the workpiece three-dimensional point cloud, the Kalman filter predicts the position deviation of the cutting head within the next 5 ms. If the predicted deviation exceeds ±0.3 mm, the path optimization is triggered in advance. First response: when the contact force is ≤0.5 mm, the cutting speed is reduced to a safe value, and the laser power is simultaneously reduced by 20%. Second response: when the contact separation is >0.5 mm, the machine is stopped and an obstacle-avoiding path is generated within 10 ms through the A* algorithm. Abnormal data such as collision position and external force direction are transmitted to the AI model library to build the anti-collision rule library.

[0092] Please refer to Figure 1 and Figure 2 An embodiment provided by the present application: a three-dimensional cutting dynamic parameter adjustment system based on deep learning, the deep learning model adopts a space-time dual-flow network architecture, including:

[0093] Temporal flow network: process sensor time series data based on LSTM unit, predict cutting stability;

[0094] Spatial flow network: analyze workpiece geometric features based on 3D convolutional neural network, generate spatial cutting parameter suggestions;

[0095] Feature fusion layer: dynamically weight spatio-temporal features through attention mechanism, output comprehensive parameter adjustment instructions;

[0096] Reinforcement learning optimization module: construct multi-objective reward function with cutting efficiency, material utilization rate and defect rate, online optimize model weights through PPO algorithm;

[0097] Further, the temporal flow network processes sensor time series data such as capacitance values and plasma signals to predict cutting process stability, the spatial flow network analyzes three-dimensional geometric features of the workpiece such as surface curvature and thickness distribution to generate spatial cutting parameter suggestions, the feature fusion layer dynamically weights spatio-temporal features to output comprehensive adjustment instructions, and the reinforcement learning module online optimizes model weights to achieve multi-objective balance of efficiency, material utilization rate and defect rate;

[0098] Temporal flow network input: time series sensor signals and time window, normalize capacitance values to 0-1 range, remove transient noise, process through LSTM unit, input 3 dimensions including capacitance, plasma and contact state, hidden layer: 64 neurons, remember cutting stability trend, output cutting stability score, score < 80 triggers parameter correction;

[0099] Temporal flow network input: workpiece three-dimensional geometric data: STL file parsed into voxel grid; material properties: thermal conductivity, reflectivity and melting point; processing logic: 3D convolution operation: first layer: 3x3x3 convolution kernel extracts local geometric features such as curvature change and thickness mutation, second layer: dilated convolution expands receptive field, identifies large-scale structural features such as reinforcing rib position; feature mapping: outputs high curvature area suggestions to reduce cutting speed to avoid heat accumulation, outputs thin-walled area suggestions to reduce laser power to prevent burn-through; parameter suggestion generation: spatial cutting speed distribution map and focal point position offset to compensate for surface focal length deviation;

[0100] Feature fusion layer input: temporal flow features and spatial flow features, fusion logic: feature alignment: upsample temporal features to the same time resolution as spatial features; attention weight calculation: generate attention scores through fully connected layer, dynamically allocate weights; weighted output: comprehensive parameter instruction = 0.7x spatial suggestion + 0.3x temporal suggestion;

[0101] The reinforcement learning optimization module optimizes the target: cutting efficiency: cutting length per unit time, material utilization rate and defective rate; training process: state space: sensor data + workpiece geometric characteristics; action space: continuous adjustment of laser power, cutting speed and nozzle height; reward function design: basic reward = efficiency x 0.4 + material utilization rate x 0.3 + (1-defective rate) x 0.3; penalty term: collision triggers a deduction of 50 points, parameter over-limit deduction of 30 points; PPO policy update: 1000 cutting trajectory data are collected online, and the policy network is updated through importance sampling.

[0102] Referring to Figure 1 and Figure 2 , the present application provides an embodiment: a three-dimensional cutting dynamic parameter adjustment system based on deep learning, the parameter adjustment system further integrates a visual detection subsystem, comprising:

[0103] A high-speed industrial camera captures infrared thermal imaging and visible light images of the cutting area;

[0104] Image fusion algorithm: superimpose thermal imaging data and visible light images, and segment the cutting defect area through a U-Net network;

[0105] Multi-sensor collaborative decision-making:

[0106] When the magnetic attraction anti-collision module detects collision risk and the visual detection subsystem synchronously identifies path obstacles, emergency stop is triggered;

[0107] When only the visual detection subsystem detects local overheating, the laser power is automatically reduced by 10%-20%;

[0108] Further, during the cutting process, the visible light camera and the infrared thermal imaging camera synchronously capture images, the image data is transmitted to the image processing unit through an optical fiber, the processing result is uploaded to the edge computing unit, and participates in multi-sensor decision-making; image alignment and registration: calibration: align the spatial coordinate systems of the visible light and infrared cameras through the chessboard calibration method, and the registration error is ≤0.1 pixels; real-time synchronization: based on the cutting head motion trajectory, dynamically compensate the image offset;

[0109] Image fusion and U-Net segmentation: fusion algorithm: visible light image: extract texture details such as cutting seam edge sharpness; infrared image: map temperature field (high temperature area > material melting point + 50℃ marked as red); fusion output: superimpose 4-channel images of RGB and thermal map; U-Net network segmentation: input: fused image; output: defect area mask; network structure: 5-layer encoder-decoder, skip connection retains details, and the training set contains more than 100,000 material cutting images;

[0110] Defect classification and feedback: Classification rule: burr: edge protrusion height > 0.1mm, trigger nozzle height adjustment; slag: temperature residue > material melting point + 100℃, suggest to improve auxiliary gas pressure; incomplete penetration: abnormal temperature gradient, trigger laser power compensation;

[0111] Multi-sensor cooperative decision mechanism: collision risk double confirmation emergency stop: trigger condition: magnetic attraction anti-collision module detects point separation > 0.5mm, visual detection subsystem identifies path obstacles; decision flow: binary calculation unit receives double sensor alarm information; start emergency stop protocol: laser power is zero, five-axis motion is paused; trigger A* algorithm to re-plan path, avoid obstacle area; record obstacle position and image to AI model library, generate "welding slag pre-scanning" rule;

[0112] Local overheating power adaptive adjustment: trigger condition: only visual detection subsystem detects local overheating (temperature > material melting point + 150℃), no magnetic attraction collision signal; adjustment strategy: if the overheating area is located in the center of the cutting seam: reduce power by 10%; if the overheating area spreads to the edge: reduce power by 20% and increase cutting speed by 15%.

[0113] Please refer to Figure 1 and Figure 2 , the present application provides an embodiment: a three-dimensional cutting dynamic parameter adjustment system based on deep learning, the deep learning model is embedded with an energy consumption optimization module, specifically comprising:

[0114] Material thermal property analysis unit: according to material thermal conductivity and specific heat capacity, calculate the minimum energy threshold to ensure that the cutting depth meets the standard;

[0115] Laser pulse optimization strategy: change continuous laser to adaptive pulse mode, pulse frequency and duty cycle are dynamically adjusted by the model, so that energy consumption is reduced by ≥15%;

[0116] Waste recycling feedback: real-time monitoring of waste quality through weighing sensor, optimizing cutting path to reduce scrap, material utilization rate is increased by ≥8%;

[0117] Further, material database matching: input material type, call pre-stored thermal conductivity, specific heat capacity and melting point, if the material is not recorded in the database, estimate the thermal properties by real-time measurement of laser reflectivity; Energy threshold calculation formula: Where, ρ is the material density, Cp is the material specific heat capacity, Tmelt is the material melting point temperature, Tambient is the ambient temperature, Lfusion is the material melting latent heat, η is the laser absorption rate, α is the spot diameter correction coefficient; If the cutting depth is insufficient, increase the energy by 10% steps to meet the standard;

[0118] Laser pulse optimization strategy: pulse parameter dynamic adjustment: frequency adjustment: thin material <3mm uses high frequency 5kHz, reduces heat accumulation, thick material >6mm uses low frequency 2kHz, enhances single pulse energy; duty cycle adjustment: duty cycle of high curvature area is reduced to 30% to avoid local overheating; linear cutting duty cycle is increased to 70% to speed up the progress; real-time energy consumption monitoring: power calculates the actual energy consumption and compares it with the continuous laser benchmark value; adaptive learning: the reinforcement learning module optimizes the pulse parameter combination according to historical data, and the reward function weights the energy consumption and cutting quality;

[0119] Waste recycling feedback and path optimization: real-time waste monitoring: a weighing sensor is installed in the waste collection tank to monitor the waste quality in real time; path optimization algorithm: nested algorithm: based on the geometric arrangement of the workpiece, the plate utilization rate is optimized from 82% to 90%; adaptive compensation: if the edge angle material ratio is detected to be out of limit, such as >12%, trigger path re-planning: reduce idle movement speed to 120%, increase common edge cutting to share cutting seam and reduce waste; waste data is uploaded to the AI model library to generate a "high utilization rate path template".

[0120] Please refer to Figure 1 and Figure 2 , the present application provides an embodiment: a three-dimensional cutting dynamic parameter adjustment system based on deep learning, the AI model library supports the following functions:

[0121] Knowledge distillation technology: extract expert experience and historical data into a lightweight rule model;

[0122] Geometric feature matching engine: compare the curvature distribution of new workpieces and model library workpieces through point cloud registration algorithm, automatically recommend similar process parameters, adaptation time ≤2 minutes;

[0123] Cross-material transfer learning: if the material of the new workpiece is not in the library, it is mapped to the existing parameter combination according to the thermal physical property similarity;

[0124] Further, data input: expert experience: manual parameter adjustment table such as laser power and speed combination for different curvatures, historical data: 100,000 cutting records containing sensor data, parameter settings and quality scores; key rule mining: extract high-frequency parameter combinations through decision tree algorithm, for example: when the curvature >0.1 / mm, the power is increased by 10%; model compression: distill the original deep learning model 1.2GB into a rule model 50MB, retaining 95% accuracy; rule model embedded in edge computing unit, inference speed shortened from 15ms to 3ms;

[0125] New workpiece scanning: three-dimensional laser scanning generates point cloud resolution 0.1mm, feature extraction: calculate curvature distribution Gaussian curvature, average curvature, thickness gradient, edge sharpness; ICP registration algorithm: align the new workpiece point cloud with the 500 workpieces in the model library, calculate the curvature difference mean square error <0.05, Top-3 recommendation: output the 3 groups of process parameters with the highest similarity such as power, speed, focal point position, if the recommended parameters do not meet the cutting quality standard such as burr height >0.1mm, trigger the online learning module for dynamic optimization;

[0126] New material parameters: thermal conductivity 160W / m·K, specific heat capacity 0.9J / g·K, reflectivity 80%, Euclidean distance calculation: compare with the thermal physical properties of the materials in the library, find the closest material such as aluminum alloy 6061, thermal conductivity 167W / m·K, error 4.3%, weight adjustment: according to the thermal conductivity difference, correct the power in proportion such as +5%, output mapping parameters: laser power = aluminum alloy 6061 reference value x 1.05, cutting speed = reference value x 0.95.

[0127] Working principle: the plasma sensor monitors the laser scattering spectrum to judge the cutting depth and quality, the capacitance type height adjustment module detects the nozzle-workpiece distance in real time, the magnetic attraction anti-collision module predicts the collision risk and triggers the multi-level response mechanism, the sensor time sequence data and the workpiece geometric features are weighted and fused through the attention mechanism to generate dynamic parameter suggestions;

[0128] The time sequence flow predicts the cutting stability trend, the spatial flow analyzes the workpiece geometric complexity, and a spatial parameter distribution map is generated, taking cutting efficiency, material utilization rate and defect rate as a multi-objective reward function, the model weights are optimized online through PPO algorithm to realize parameter self-exploration;

[0129] The dynamic parameters are transmitted to the laser generator and five-axis controller through the edge computing unit to adjust the power, speed and focal point position in real time, and the cutting quality data is fed back to the AI model library to trigger incremental learning and federated learning, and continuously improve the model generalization ability.

[0130] It is apparent to those skilled in the art that the present application is not limited to the details of the foregoing exemplary embodiments, but can be implemented in other concrete forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be considered in all respects as illustrative and not restrictive, and the scope of the present application should be defined by the appended claims rather than the above description, and it is intended to encompass all changes falling within the meaning and scope of the equivalent elements of the claims. Any reference signs in the claims should not be considered as limiting the claims involved.

Claims

1. A deep learning-based 3D cutting dynamic parameter adjustment system, comprising a multimodal sensor cluster, an edge computing unit, a closed-loop control module, and an AI model library, characterized in that: The multimodal sensor cluster collects cutting data, the edge computing unit receives the multimodal sensor data and outputs dynamic adjustment parameters, and the closed-loop control module transmits the dynamic parameters to the laser generator and the five-axis motion controller through a high-speed communication bus, forming a "perception-decision-execution" closed loop. The edge computing unit has a built-in deep learning model. The inputs include multimodal sensor data, workpiece three-dimensional geometric parameters and material properties, and the outputs dynamically adjustable parameters: laser power, cutting speed, nozzle height and focal position. The AI ​​model library stores historical segmentation data, abnormal scenarios, and optimization strategies, and supports online model updates and transfer learning. The multimodal sensor cluster includes: Plasma detection sensors are used to monitor laser scattering spectral signals during the cutting process and determine missed cuts and incomplete penetration based on spectral intensity thresholds. The capacitive sensing height adjustment module consists of a ceramic cutting nozzle, an amplifier board, and cables. It dynamically calculates the nozzle height deviation by detecting the change in capacitance between the cutting nozzle and the workpiece. The magnetic anti-collision module includes upper and lower magnetic contacts and an anti-collision detection circuit. When an external force causes the contacts to separate by more than 0.5mm, it triggers a device shutdown and path replanning command. The deep learning model employs a spatiotemporal dual-stream network architecture, including: Temporal Streaming Network: Based on LSTM units, it processes sensor temporal data to predict cutting stability; Spatial Flow Network, based on 3D convolutional neural network, analyzes workpiece geometric features and generates spatial cutting parameter suggestions; Feature fusion layer: Dynamically weights spatiotemporal features through an attention mechanism and outputs comprehensive parameter adjustment instructions; The reinforcement learning optimization module constructs a multi-objective reward function based on cutting efficiency, material utilization rate, and defect rate, and optimizes the model weights online through the PPO algorithm. The adjustment logic of the capacitive sensing height adjustment module includes: Dynamic compensation algorithm: Based on the surface roughness of the workpiece and the conductivity of the material, the amplification factor of the capacitance signal is dynamically adjusted to ensure that the height detection error is ≤0.03mm; Adaptive filtering technology: Wavelet transform is used to filter out high-frequency noise while retaining effective capacitance signals; Dual closed-loop control: Inner loop: Nozzle height is adjusted in real time based on a PID controller; Outer ring: Predicts the thermal deformation of the workpiece using a deep learning model to compensate for height deviation in advance; The edge computing unit integrates the following model update mechanism: Incremental learning module: After each cutting task is completed, key data including cutting quality scores and sensor anomaly markers are automatically extracted, and model parameters are updated using online gradient descent, with an update cycle of ≤3 minutes; Model version management: Retain historical model versions, and automatically roll back to the optimal version when new data causes a performance increase (i.e., defect rate) greater than 0.1%. Federated learning interface: Supports encrypted sharing of local model parameters among multiple devices to build a globally optimized model; The parameter adjustment system further integrates a visual inspection subsystem, including: High-speed industrial cameras capture infrared thermal and visible light images of the cut area; Image fusion algorithm: Overlays thermal imaging data with visible light images, and segments and cuts defect areas using a U-Net network; Multi-sensor collaborative decision-making: When the magnetic anti-collision module detects a collision risk and the visual detection subsystem simultaneously identifies obstacles in the path, an emergency stop is triggered. When the visual inspection subsystem detects localized overheating, it automatically reduces the laser power by 10%-20%. The deep learning model embeds an energy consumption optimization module, specifically including: Material thermal property analysis unit: Calculates the minimum energy threshold based on the material's thermal conductivity and specific heat capacity to ensure that the cutting depth meets the standard; Laser pulse optimization strategy: Change the continuous laser to an adaptive pulse mode, and dynamically adjust the pulse frequency and duty cycle by the model to reduce energy consumption by ≥15%; Waste recycling feedback: Real-time monitoring of waste quality via weighing sensors optimizes cutting paths to reduce scrap, increasing material utilization by ≥8%; The AI ​​model library supports the following functions: Knowledge distillation technology: Extracting expert experience and historical data into lightweight rule models; Geometric feature matching engine: By comparing the curvature distribution of the new workpiece with that of the workpiece in the model library through the point cloud registration algorithm, it automatically recommends similar process parameters with an adaptation time of ≤2 minutes; Cross-material transfer learning: If the material of a new workpiece is not in the library, it is mapped to an existing parameter combination based on the similarity of its thermophysical properties.

2. The three-dimensional cutting dynamic parameter adjustment system based on deep learning according to claim 1, characterized in that: The magnetic anti-collision module further includes: Collision prediction algorithm: Based on the five-axis motion trajectory and the workpiece's three-dimensional point cloud data, the position offset of the cutting head within the next 5ms is predicted by Kalman filtering. If the offset exceeds the safety threshold, path replanning is triggered in advance. Multi-level response mechanism: Level 1 Response: In the event of a minor collision, adjust the cutting speed only to a safe level; Level 2 response: In the event of a severe collision, the system shuts down and initiates the A* algorithm to dynamically generate an obstacle avoidance path; Collision Scene Learning Module: Uploads abnormal data to the AI ​​model library to generate a collision avoidance rule library for active avoidance when cutting similar workpieces.

3. A method for adjusting dynamic parameters of three-dimensional cutting based on deep learning, applicable to the three-dimensional cutting dynamic parameter adjustment system based on deep learning as described in claim 2, characterized in that: The adjustment method includes the following steps: S1. Data Acquisition Stage: Real-time acquisition of plasma spectrum, capacitance value, magnetic contact status, workpiece geometric data and ambient temperature through a multimodal sensor cluster; thermal imaging and morphological images of the cutting area are obtained through the vision inspection subsystem. S2, Model Inference Stage: The collected data is input into the spatiotemporal dual-stream network, and the dynamic parameter combination of laser power, cutting speed, nozzle height, and focal position is output. The pulsed laser parameters are generated through the energy consumption optimization module. S3, Execution and Feedback Phase: The closed-loop control module drives the laser generator and the five-axis motion mechanism to adjust the execution parameters, and updates the model weights through the incremental learning module based on the cutting quality and energy consumption data; S4. Anomaly Handling Phase: If a collision risk or parameter exceedance is detected, a multi-level response mechanism is triggered, and the abnormal scenario is recorded in the AI ​​model library. S5: Cloud-based collaborative optimization: Local model parameters and abnormal data are encrypted and uploaded to the cloud, and a global optimization model is generated through federated learning and distributed to all devices; S6, Process Chain Traceability: Generates a unique ID for each cutting task, associates parameter settings, sensor data and quality reports, and supports full lifecycle quality analysis.

Images