Laser cutting end face quality control method and system

Abstract

The invention discloses a laser cutting end surface quality control method, which comprises the following steps: S1, data acquisition: acquiring the thickness, material spectrum, surface roughness, laser power and auxiliary gas pressure data of a workpiece by using a multi-mode sensing module, performing dimension reduction on hyperspectral data through a principal component analysis algorithm, extracting a material feature vector, and calculating the material feature vector; and generating an initial process parameter request in combination with the thickness information. Through a multi-parameter dynamic collaborative optimization mechanism, a coupling rule among parameters in the laser cutting process is deeply excavated, parameters such as laser power, cutting speed, focus position and auxiliary gas pressure are brought into a unified regulation and control system, and through a parameter coupling relation model and a multivariable collaborative regulation strategy, compared with a traditional single-parameter regulation mode, the adjustment efficiency is greatly improved. The notch width consistency is improved by 65%, the fluctuation range is controlled within + / -0.02 mm, the end face roughness is reduced by 52%, the Ra value can reach 3.2 [mu] m or below, and the problems of thick plate slag residue, thin plate overburning deformation and the like are effectively solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of laser cutting, and particularly relates to a laser cutting end face quality control method and system. BACKGROUND

[0002] In the process of modern manufacturing towards high precision and intelligentization, laser cutting as a core processing technology is widely used in the fields of aerospace, automobile manufacturing, precision electronics, etc. However, the current laser cutting end face quality control technology for workpieces of different thicknesses and materials still has a bottleneck that is difficult to break through.

[0003] Most of the existing laser cutting equipment adopts a single parameter adjustment or independent control mode. For example, the traditional PID control system only adjusts the focal point position according to a single index of kerf width, completely ignoring the coupling relationship between laser power, cutting speed, focal point position and auxiliary gas parameters. This isolated adjustment mode, when cutting thick plates (such as carbon steel with a thickness of more than 10 mm), cannot cooperatively optimize each parameter to enhance the slag discharge capacity, resulting in a large amount of slag remaining at the bottom of the kerf, the kerf width fluctuation reaching ±0.2 mm, and the end face roughness Ra value exceeding 20 μm; while processing thin plates (such as aluminum alloy with a thickness of less than 1 mm), it cannot timely adjust the heat input parameters, easily causing overburning deformation, causing cutting size error to be out of tolerance, and seriously affecting product qualification rate.

[0004] Traditional cutting relies on manual preset parameters or simple material thickness mapping table, and when facing new materials with special optical properties and thermal physical properties, it cannot quickly and accurately identify material properties and match the optimal cutting parameters. When cutting carbon fiber reinforced composite materials, traditional cutting often causes material delamination, carbonization and other defects due to the inability to effectively control the synergistic effect of laser energy and auxiliary gas, resulting in high scrap rate. SUMMARY

[0005] In view of the problems in the prior art, the technical scheme is proposed as follows:

[0006] The laser cutting end face quality control method comprises:

[0007] S1, data acquisition:

[0008] The thickness, material spectrum, surface roughness, laser power and auxiliary gas pressure data of the workpiece are collected by using a multi-modal perception module, the hyperspectral data are reduced in dimension by a principal component analysis algorithm, the material feature vector is extracted, and the initial process parameter request is generated in combination with the thickness information;

[0009] S2, workpiece pre-cutting:

[0010] Short-range pre-cutting is performed on the edge of the workpiece, the spectral changes of the cutting area are collected by a hyperspectral imager, the temperature field of the molten pool is analyzed, the edge detection algorithm is used to calculate the kerf width and taper, the initial parameters are iteratively optimized at least three times based on the Bayesian optimization algorithm, and the formal cutting parameters are determined;

[0011] S3, workpiece cutting:

[0012] During the cutting process, the kerf image, power fluctuation and gas pressure data are collected in real time at a frequency of not less than 200HZ, based on the pre-constructed parameter coupling relationship model, when the kerf width deviation exceeds ±0.02mm or the end face roughness Ra value exceeds the predetermined threshold, a multivariate collaborative adjustment strategy is adopted to adjust the laser power, cutting speed, focal point position and auxiliary gas pressure in linkage;

[0013] S4, data storage:

[0014] After each cutting is completed, the actual parameters, process data and quality evaluation results are stored in the process knowledge base; the data is classified by using the DBSCAN clustering algorithm, and the retraining based on the deep reinforcement learning model is triggered every 500 groups of effective data, and the parameter optimization model is updated.

[0015] As a preferred embodiment of the above technical solution, S11, in the multi-modal perception module, the measurement accuracy of the laser power sensor is ±1%, the measurement accuracy of the non-contact laser triangulation thickness gauge is ±0.01mm, the hyperspectral imager covers the 400-1000nm band, and the material type and surface state can be identified within 0.1 seconds, and the dynamic pressure sensor monitors the accuracy of the auxiliary gas pressure ±0.1bar.

[0016] As a preferred embodiment of the above technical solution, S31, the construction method of the parameter coupling relationship model is:

[0017] Through a large number of laser cutting tests of different materials and thickness workpieces, the laser power, cutting speed, focal point position, auxiliary gas pressure, kerf width, end face roughness and other data during the cutting process are collected;

[0018] The collected data is trained by using a machine learning algorithm to establish a nonlinear mapping relationship between each parameter and the cutting quality index.

[0019] As a preferred embodiment of the above technical solution, S32, the multivariate collaborative adjustment strategy is specifically:

[0020] When the kerf width is widened, the laser power is reduced and the cutting speed is increased in the proportion of 0.8-0.9 and 1.1-1.3 respectively, and the focal point is moved up by 0.03-0.07mm / time;

[0021] When the end face roughness Ra value exceeds the standard, the auxiliary gas pressure and cutting speed are automatically optimized according to the pressure-power-speed linkage formula ΔP = 0.08-0.12×ΔRa, ΔV =-0.04-0.06×ΔRa;

[0022] Wherein, ΔP is the power adjustment amount, and ΔV is the speed adjustment amount.

[0023] As a preferred embodiment of the above technical solution, in S41, the deep reinforcement learning model is fused with:

[0024] The LSTM neural network is used to process time series data in the cutting process, and to mine the rules and potential relationships of the parameters changing over time.

[0025] The Q-learning algorithm is used to find the optimal parameter combination in a complex parameter space, and to realize online updating of the parameter optimization model through continuous trial and error and learning.

[0026] The laser cutting end face quality control system comprises:

[0027] The multi-modal perception module comprises a laser power sensor, a non-contact laser triangulation thickness gauge, a hyperspectral imager, and a dynamic pressure sensor, and is used to collect data such as the thickness material spectrum, surface roughness, laser power, and auxiliary gas pressure of the workpiece.

[0028] The intelligent decision-making module is internally provided with a deep reinforcement learning model, the deep reinforcement learning model is fused with an LSTM neural network and a Q-learning algorithm, takes the data collected by the multi-modal perception module as input, outputs the optimal parameter combination of the laser power, cutting speed, focal point position, auxiliary gas type, and pressure, and has online learning capability.

[0029] The collaborative execution module comprises a piezoelectric ceramic driven fast zoom cutting head, a three-way independent gas supply system, and a laser power adjustment module, the piezoelectric ceramic driven fast zoom cutting head has a focal point position adjustment response speed of 1000 Hz, the three-way independent gas supply system realizes millisecond-level switching of the gas type and pressure gradient control through proportional electromagnetic valves, the pressure adjustment precision is ±0.05 bar, and the laser power adjustment module adopts pulse width modulation and power slope control technology, the power adjustment step is 0.05 kW, and the response time is less than 10 ms.

[0030] The system is used to execute the laser cutting end face quality control method.

[0031] As a preferred embodiment of the above technical solution, the intelligent decision-making module comprises a process parameter database, and stores an initial process parameter mapping table corresponding to different materials and different thicknesses.

[0032] As the preferred technical scheme of the above, the cooperative execution module further comprises a servo motor driven numerical control motion platform for controlling the translational motion of the laser cutting head in the X and Y axis directions, with a positioning accuracy of less than or equal to ±0.02 mm, so as to realize accurate cutting of different positions of the workpiece.

[0033] As the preferred technical scheme of the above, the system further comprises an abnormality monitoring and emergency handling module for monitoring various parameters in the cutting process in real time.

[0034] The present application has the following beneficial effects:

[0035] 1. The present application uses a multi-parameter dynamic cooperative optimization mechanism to deeply mine the coupling rules between various parameters in the laser cutting process, and integrates laser power, cutting speed, focal point position, auxiliary gas pressure and other parameters into a unified control system. Through the parameter coupling relationship model and the multi-variable cooperative regulation strategy, compared with the traditional single-parameter regulation mode, the cut width consistency is improved by 65%, the fluctuation range is controlled within ±0.02 mm, the end face roughness is reduced by 52%, and the Ra value can be below 3.2 μm, effectively solving the problems of thick plate slag residue and thin plate deformation;

[0036] 2. The parameter optimization model based on deep reinforcement learning of the present application integrates LSTM neural network and Q-learning algorithm, and through deep training of the cutting sample data, it can quickly adapt to new material and new process requirements, and the accuracy rate of process parameter matching for new materials is greatly improved from 70% to above 95%. At the same time, the real-time dynamic optimization function of the system can complete parameter adjustment within 100 ms according to the data feedback in the cutting process, so as to improve the cutting efficiency by 40%, shorten the product processing cycle and improve the production benefit;

[0037] 3. The multi-modal perception module of the present application integrates laser power sensor, non-contact laser triangulation thickness gauge, hyperspectral imager and dynamic pressure sensor to build a comprehensive data acquisition system. This module can complete accurate identification of 0.1-100 mm thickness range and more than 20 kinds of metal / non-metal materials within 0.1 seconds, automatically match the optimal initial parameters, and combine with the self-evolution mechanism of the process knowledge base. The system can accumulate cutting experience of new materials and complex working conditions through DBSCAN clustering algorithm and model retraining, realize efficient processing of workpieces of different materials and thicknesses with "one-key switching", and reduce the process debugging cost. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The flowchart of the laser cutting end face quality control method in the embodiment is shown.

[0039] Figure 2 The flowchart of the laser cutting end face quality control system in the embodiment is shown. DETAILED DESCRIPTION

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments.

[0041] Example 1

[0042] like Figure 1 As shown, the laser cutting end face quality control method includes data acquisition, workpiece pre-cutting, workpiece cutting, and data storage. Data acquisition utilizes a multimodal sensing module to collect workpiece thickness, material spectrum, surface roughness, laser power, and auxiliary gas pressure data. Principal component analysis is used to reduce the dimensionality of the hyperspectral data, extracting material feature vectors, and combining this with thickness information to generate initial process parameter requests. Workpiece pre-cutting involves short-range pre-cutting at the workpiece edge. A hyperspectral imager is used to collect spectral changes in the cutting area, analyze the molten pool temperature field, and use an edge detection algorithm to calculate the cut width and taper. Based on a Bayesian optimization algorithm, the initial parameters are iteratively optimized at least three times to determine the final cutting parameters. Parameters: During the workpiece cutting process, cut images, power fluctuations, and gas pressure data are acquired in real time at a frequency of no less than 200 Hz. Based on a pre-built parameter coupling model, when the cut width deviation exceeds ±0.02 mm or the end face roughness Ra value exceeds a predetermined threshold, a multi-variable collaborative adjustment strategy is adopted to adjust the laser power, cutting speed, focal position, and auxiliary gas pressure in a coordinated manner. Data storage: After each cutting is completed, the actual parameters, process data, and quality evaluation results are stored in the process knowledge base. The DBSCAN clustering algorithm is used to classify the data. Every 500 sets of valid data are accumulated to trigger retraining based on the deep reinforcement learning model and update the parameter optimization model.

[0043] Specifically, the DBSCAN clustering algorithm is as follows:

[0044] Data Classification and Organization: After each cutting task is completed, the system stores the actual parameters used, real-time data during the cutting process, and the final quality evaluation results (such as kerf width, roughness, and slag residue) in the process knowledge base. The DBSCAN clustering algorithm can automatically classify this large amount of complex data. Based on the density connectivity between data points, it groups data points with similar densities in the feature space into the same cluster. For example, for cutting data of workpieces of different materials and thicknesses, the DBSCAN algorithm can classify them according to the similarity of parameter combinations and cutting quality, making the data in the process knowledge base more organized and facilitating subsequent management and retrieval.

[0045] Trigger model retraining: every 500 groups of valid data accumulated, the system will trigger the model retraining mechanism. The DBSCAN clustering algorithm classifies the data, providing a good data basis for model retraining. By analyzing the clustered data, the Q-table of the parameter optimization model based on deep reinforcement learning can be updated more targeted. For example, for a certain type of data with similar cutting characteristics, it is found that the model has a certain deviation in predicting parameters. Through clustering, this part of data can be concentrated for learning and optimization, thereby further improving the accuracy of the model in predicting parameters under various working conditions and enhancing the adaptability of the system to new materials and complex working conditions.

[0046] Mining potential laws of data: DBSCAN clustering algorithm helps to discover hidden potential laws and patterns in data. In the laser cutting process, there is a complex relationship between different material properties, process parameters and cutting quality. Through clustering analysis, some previously unnoticed associations between parameter combinations and cutting quality can be discovered. For example, it may be found that certain specific material properties and parameter combinations can achieve better cutting results. The mining of these laws can provide important basis for further optimizing the cutting process and enriching and perfecting the process knowledge base.

[0047] In the multi-modal perception module, the measurement accuracy of the laser power sensor is ±1%, the non-contact laser triangulation thickness gauge has a measurement accuracy of ±0.01mm based on the double-beam interference principle, the hyperspectral imager covers the 400-1000nm band and can complete material type and surface state recognition within 0.1 seconds, and the dynamic pressure sensor monitors the accuracy of the auxiliary gas pressure to ±0.1bar.

[0048] The construction method of the parameter coupling relationship model is as follows:

[0049] Through a large number of laser cutting tests of different materials and thickness workpieces, laser power, cutting speed, focal point position, auxiliary gas pressure, kerf width, end face roughness and other data during the cutting process are collected;

[0050] Use machine learning algorithms to train the collected data and establish a nonlinear mapping relationship between each parameter and the cutting quality index.

[0051] The multivariate collaborative regulation strategy is as follows:

[0052] When the kerf width becomes wider, the laser power is reduced by a factor of 0.8-0.9, the cutting speed is increased by a factor of 1.1-1.3, and the focal point is moved up by 0.03-0.07mm / time;

[0053] When the end face roughness Ra value exceeds the standard, the auxiliary gas pressure and cutting speed are automatically optimized according to the pressure-power-speed linkage formula ΔP = 0.08-0.12 x ΔRa, ΔV =-0.04-0.06 x ΔRa; wherein, ΔP is the power adjustment amount, and ΔV is the speed adjustment amount.

[0054] The deep reinforcement learning model is fused with an LSTM neural network and a Q-learning algorithm. The LSTM neural network is used to process time series data in the cutting process, mine the rules and potential relationships of parameters changing over time, and the Q-learning algorithm is used to find the optimal parameter combination in a complex parameter space. Through continuous trial and error and learning, the online update of the parameter optimization model is realized.

[0055] Specifically, the LSTM neural network is as follows:

[0056] Mining parameter time series rules: the laser cutting process is a dynamic time series process, and the changes of various process parameters (laser power, cutting speed, focal point position, etc.) over time and their interactions have a complex and critical impact on the final cutting quality. The LSTM neural network can deeply process the time series data in the cutting process due to its unique memory cell structure. For example, when cutting complex-shaped workpieces, the parameter settings at different times need to be coordinated with each other as the cutting path changes. The LSTM neural network can learn how the cutting speed and focal point position should change accordingly after adjusting the laser power at a certain time, thereby mining the potential rules of these parameters changing over time and providing strong support for intelligent decision-making.

[0057] Processing long-distance dependency: in laser cutting, the cutting quality at the current time is not only related to the immediate process parameters, but also may be affected by the changes of parameters in a long period of time before. For example, when cutting thick plates, the energy input and slag discharge in the early stage will have a lasting effect on the stability and quality of subsequent cutting. The long-term memory capability of the LSTM neural network enables it to effectively process such long-distance dependencies, selectively retain and update historical information through the collaborative work of the forget gate, input gate and output gate. It can remember the key states in the previous cutting process, such as the degree of accumulated thermal influence and the trend of molten pool changes, and combine these information with the current state to more accurately judge the current working condition and provide a more comprehensive basis for parameter optimization.

[0058] Co-optimization with Q-learning algorithm: This technical solution combines LSTM neural networks with Q-learning algorithms to form powerful intelligent decision-making capabilities. The LSTM neural network is responsible for feature extraction and analysis of the time-series data collected by the multimodal sensing module, uncovering complex relationships and potential patterns in the data, providing a more representative state representation for the Q-learning algorithm. The Q-learning algorithm, based on the data processed by the LSTM, explores and learns in a complex parameter space to find the optimal parameter adjustment strategy. The two complement each other: the LSTM neural network provides higher-quality input to the Q-learning algorithm, enabling it to learn the optimal strategy more efficiently; the Q-learning algorithm, based on the state information provided by the LSTM neural network, continuously optimizes the parameter combination, jointly achieving precise control of laser cutting parameters, improving the quality of the cut end face and processing efficiency.

[0059] The Q-learning algorithm is as follows:

[0060] Parameter Space Exploration and Optimization: In the laser cutting process, parameters such as laser power, cutting speed, focal point position, and auxiliary gas type and pressure constitute a complex high-dimensional parameter space. The Q-learning algorithm uses data collected by the multimodal sensing module (such as workpiece thickness, material spectrum, real-time cutting status, etc.) as state input, defining different parameter adjustments as actions. The algorithm continuously "trial and error" within this parameter space, trying different parameter combinations and receiving corresponding reward feedback based on the quality evaluation results after cutting (such as kerf width, end face roughness, slag residue, etc.). For example, if cutting with a certain set of parameters results in a highly consistent kerf width and a smooth end face, the algorithm will award a higher reward value; conversely, if problems such as slag residue occur, a lower reward will be given. In this way, the algorithm gradually explores the optimal parameter combination under different working conditions, thereby optimizing the cutting process.

[0061] Dynamically adapting to changing cutting conditions: In actual laser cutting, conditions constantly change, such as batch variations in materials and fluctuations in equipment performance. The Q-learning algorithm has online learning capabilities, updating the Q-table (which stores the expected cumulative reward for different actions in each state) in real time based on new states and reward information. During the cutting process, once a change in the cutting state is detected (such as uneven workpiece surface requiring adjustment of the focal position), the algorithm can quickly adjust parameters according to the updated Q-table, dynamically adapting to the new conditions and ensuring the stability of cutting quality. For example, when cutting different parts of the workpiece and experiencing slight changes in material thickness, the algorithm can promptly adjust the laser power and cutting speed to avoid affecting the cutting effect due to parameter mismatch.

[0062] The technical solution combines the Q-learning algorithm with the LSTM neural network. The LSTM neural network is good at processing time series data and can mine the rules of changes of parameters in the cutting process over time and the potential relationship between them, providing a more in-depth representation of state features for the Q-learning algorithm. The Q-learning algorithm focuses on finding the optimal parameter adjustment strategy through reinforcement learning based on the features extracted by the LSTM. The two work together to make the intelligent decision-making module not only utilize the processing capabilities of the LSTM for historical data and time series information, but also take advantage of the Q-learning in dynamic environments to optimize decisions, significantly improving the intelligent decision-making capabilities of the system for laser cutting parameters, thereby achieving high-precision cutting quality control for workpieces of different thicknesses and materials.

[0063] As shown in Figure 2 The technical solution combines the Q-learning algorithm with the LSTM neural network. The LSTM neural network is good at processing time series data and can mine the rules of changes of parameters in the cutting process over time and the potential relationship between them, providing a more in-depth representation of state features for the Q-learning algorithm. The Q-learning algorithm focuses on finding the optimal parameter adjustment strategy through reinforcement learning based on the features extracted by the LSTM. The two work together to make the intelligent decision-making module not only utilize the processing capabilities of the LSTM for historical data and time series information, but also take advantage of the Q-learning in dynamic environments to optimize decisions, significantly improving the intelligent decision-making capabilities of the system for laser cutting parameters, thereby achieving high-precision cutting quality control for workpieces of different thicknesses and materials.

[0064] It should be noted that the laser power sensor: the thermal electric pile type laser power meter probe of Ophir Company can be selected, such as the BB type probe with wide spectrum absorption material. It is based on the thermal electric pile structure, can convert light energy into heat and further convert into electric signal output, the measurement precision can reach ±1%, the probe plating damage domain value can reach 20KW / cm, the absorption rate is about 90%, the response curve from ultraviolet to infrared is smooth, can accurately measure the power change in the laser cutting process;

[0065] Non-contact laser triangulation thickness gauge: LAP polarization series laser distance sensor can be used as a suitable choice, which works on the principle of triangle, integrates DSP (digital signal processor) internally, and can achieve a data processing speed of up to 4kHz, ensuring fast and accurate measurement. For example, the laser thickness gauge ZTMS08 used for online detection is based on the principle of laser triangulation measurement, with a measurement accuracy of ±0.01mm, which can effectively obtain the thickness information of the workpiece;

[0066] Hyperspectral imager: X20P integrated airborne hyperspectral imager from Cubert GmbH / Germany can meet the needs of collecting spectral changes in the cutting area. This imager covers the 400-1000nm band and can quickly complete material type and surface state recognition within 0.1 seconds, providing key data support for subsequent parameter optimization. In addition, the double spectrum visible-near infrared hyperspectral imager from Jiangsu Shuangli Hespectrum Technology Co., Ltd. can also be used as a backup and has good performance in related application scenarios;

[0067] Dynamic pressure sensor: Alps Electric's "HSFPAR series" pressure detection sensor can be used, which uses a piezoelectric element based on MEMS technology to convert pressure changes into current changes for detection. Its size is small, with a surface package type of 2mm x 1.6mm x 0.66mm, capable of detecting 0.01N level of micro pressure, and can withstand more than 1 million pressure actions, with a monitoring auxiliary gas pressure accuracy of ±0.1bar.

[0068] The system is used to perform a laser cutting end face quality control method.

[0069] The intelligent decision module includes a process parameter database storing an initial process parameter mapping table corresponding to different materials and different thicknesses.

[0070] Specifically, the database has a self-updating function, which can optimize and update the initial process parameter mapping table according to new data generated during the process knowledge base self-evolution process.

[0071] The collaborative execution module also includes a servo motor driven numerical control motion platform for controlling the translational motion of the laser cutting head in the X and Y axis directions, with a positioning accuracy of ≤±0.02mm, achieving accurate cutting of different positions of the workpiece.

[0072] The laser cutting end face quality control system also includes an abnormality monitoring and emergency handling module for real-time monitoring of various parameters during the cutting process.

[0073] Need to explain, when the cut width deviation exceeds ± 0.1mm, roughness Ra> 25μm or auxiliary gas pressure is lower than 80% of the set value, trigger sound light alarm, and automatically reduce the cutting speed by 30% or pause cutting, while switching to standby gas path.

[0074] Example 2

[0075] Data acquisition and initial parameter determination:

[0076] After loading the 5mm stainless steel plate, the multi-modal perception module works, the non-contact laser triangulation thickness gauge measures the thickness to be 5.02mm, and the hyperspectral imager identifies it as 304 stainless steel through spectral analysis;

[0077] The intelligent decision-making module calls the initial parameters: power 1.5kW, speed 1.5m / min, focus +0.3mm, nitrogen pressure 10bar.

[0078] Pre-cutting and parameter optimization:

[0079] 5mm short-range pre-cutting is performed, and the hyperspectral imager finds that the molten pool temperature field is not uniform, resulting in slight molten collapse on the upper edge of the cut;

[0080] Using the Bayesian optimization algorithm, after 3 iterations, the power is adjusted to 1.4kW, and the focus is moved down to +0.2mm.

[0081] Official cutting and dynamic adjustment:

[0082] When the official cutting reaches the middle, the system detects that the cut width increases by 0.04mm, according to the multivariate coordinated adjustment strategy, the power is reduced to 1.35kW, the speed is increased to 1.6m / min, the focus is moved up by 0.05mm, and the nitrogen pressure is increased to 11bar. The final cut width fluctuation is controlled within ± 0.015mm, and the end face roughness Ra = 3.2μm.

[0083] Process knowledge base update:

[0084] After cutting is completed, the parameters, process data and quality evaluation results are stored in the process knowledge base, which is used for model retraining in the future.

[0085] Example 3

[0086] Initial parameter acquisition:

[0087] 12mm carbon steel thick plate is loaded, and the multi-modal perception module measures the thickness to be 12.05mm and identifies the material.

[0088] The intelligent decision-making module gives the initial parameters: power 3.0kW, speed 0.6m / min, focus -1.5mm, oxygen pressure 8bar.

[0089] Pre-cut optimization:

[0090] After pre-cut, it was found that there were many slag residues at the bottom of the cut and the width was uneven.

[0091] After Bayesian optimization, the power was adjusted to 3.2 kW, the speed was reduced to 0.5 m / min, the focal point was lowered to -1.8 mm, and the oxygen pressure was increased to 9 bar.

[0092] Cutting process adjustment:

[0093] The end face roughness during cutting exceeded the standard. According to the pressure-power-speed linkage formula, the power was increased by 0.2 kW, the speed was reduced by 0.05 m / min, and the oxygen pressure was increased by 0.5 bar. Finally, the cut width was uniform, the slag residue was small, and the end face roughness Ra was 6.3 μm.

[0094] Data storage: store the cutting data this time into the process knowledge base.

[0095] Example 4

[0096] Initial parameter setting:

[0097] After loading the 0.8 mm aluminum alloy sheet and measuring the thickness of 0.82 mm and identifying the material, the initial parameters were: power 0.8 kW, speed 3.0 m / min, focal point +0.2 mm, and argon pressure 12 bar.

[0098] Pre-cut adjustment:

[0099] During pre-cut, the sheet edge was slightly overburned. After Bayesian optimization, the power was reduced to 0.7 kW, the speed was increased to 3.2 m / min, and the focal point was adjusted to +0.3 mm.

[0100] Formal cutting control:

[0101] During formal cutting, the system was adjusted in real time, and finally achieved no overburning and no deformation cutting. The cut width fluctuation was controlled within ±0.01 mm, and the end face roughness Ra was 1.6 μm.

[0102] Data accumulation: after completing the cutting, the data was stored into the process knowledge base.

[0103] Comparative Example 1

[0104] A 5mm stainless steel plate was cut using a conventional single-parameter adjustment laser cutting system, and the focal position was adjusted only according to the kerf width. The initial parameter settings were 1.5kW of power, 1.5m / min of speed, +0.3mm of focal position, and 10bar of nitrogen pressure. During the cutting process, it was found that the upper edge of the kerf had a molten collapse, and the system only adjusted the focal position to +0.2mm, without adjusting the power and speed in coordination. The final kerf width fluctuation reached ±0.08mm, the end face roughness Ra was 8.5μm, and there was a small amount of slag residue, and the cutting quality was significantly lower than that of Example 2 of the present application.

[0105] Comparative Example 2

[0106] A 12mm carbon steel thick plate was cut using a laser cutting system without intelligent optimization algorithm, and the parameters were set by artificial experience as 3.0kW of power, 0.6m / min of speed, -1.5mm of focal position, and 8bar of oxygen pressure. During the cutting process, there was a large amount of slag residue at the bottom of the kerf and a rough end face. Due to the lack of intelligent algorithm optimization and parameter coordination adjustment mechanism, it was difficult for the artificial to quickly and accurately adjust the parameters. The final kerf width was uneven, the slag residue was serious, the end face roughness Ra was 15μm, and the cutting efficiency was low, which was far from the effect of Example 3 of the present application.

[0107] Comparative Example 3

[0108] A 0.8mm aluminum alloy thin plate was cut using a single sensor laser cutting system equipped with only a laser power sensor, and the initial parameters were set as 0.8kW of power, 3.0m / min of speed, +0.2mm of focal position, and 12bar of argon pressure. During the cutting process, the thin plate was overburned and deformed, but the system could only monitor the power and could not sense the comprehensive changes of the material properties and the kerf quality, and could not effectively adjust the parameters. The final kerf width fluctuation was large, the overburning was serious, the end face roughness Ra was 12μm, and the cutting quality was poor, which was in sharp contrast to Example 4 of the present application.

[0109] The above examples are only used to illustrate the technical solutions of the present application, and not to limit them.

Claims

1. A method of end face quality control for laser cutting, characterized in that, Comprise: S1, data acquisition: Collect the thickness, material spectrum, surface roughness, laser power and auxiliary gas pressure data of the workpiece by using the multi-modal perception module, reduce the dimensionality of the hyperspectral data by principal component analysis algorithm, extract the material feature vector, and generate the initial process parameter request combined with the thickness information; S2, workpiece pre-cutting: Short-range pre-cutting is performed on the edge of the workpiece, the spectral change of the cutting area is collected by the hyperspectral imager, the molten pool temperature field is analyzed, the kerf width and taper are calculated by using the edge detection algorithm, and the initial parameters are iteratively optimized at least three times based on the Bayesian optimization algorithm to determine the formal cutting parameters; S3, workpiece cutting: In the cutting process, the kerf image, power fluctuation and gas pressure data are collected in real time at a frequency of not less than 200HZ, based on the pre-constructed parameter coupling relationship model, when the kerf width deviation exceeds ±0.02mm or the end face roughness Ra value exceeds the predetermined threshold, the multi-variable collaborative adjustment strategy is adopted to link the laser power, cutting speed, focal point position and auxiliary gas pressure for linkage adjustment; S4, data storage: After each cutting is completed, the actual parameters, process data and quality evaluation results are stored in the process knowledge base; DBSCAN clustering algorithm is used for data classification, and every 500 groups of effective data trigger retraining based on deep reinforcement learning model to update the parameter optimization model.

2. The laser cut end face quality control method of claim 1, wherein, S11, in the multi-modal perception module, the measurement accuracy of the laser power sensor is ±1%, the measurement accuracy of the non-contact laser triangulation thickness gauge is ±0.01mm, the hyperspectral imager covers 400-1000nm band, and the material type and surface state can be identified within 0.1 seconds, and the dynamic pressure sensor monitors the accuracy of the auxiliary gas pressure is ±0.1bar.

3. The laser cut end face quality control method of claim 1, wherein, S31, the construction method of the parameter coupling relationship model is: Through a large number of laser cutting tests of different materials and thickness workpieces, the laser power, cutting speed, focal point position, auxiliary gas pressure, kerf width and end face roughness data in the cutting process are collected; Use machine learning algorithm to train the collected data, and establish the nonlinear mapping relationship between each parameter and cutting quality index.

4. The laser cut end face quality control method of claim 1, wherein, S32, the multi-variable collaborative adjustment strategy is specifically: When the kerf width becomes wider, the laser power is reduced by 0.8-0.9, the cutting speed is increased by 1.1-1.3, and the focal point is moved up by 0.03-0.07mm / time; When the end face roughness Ra value exceeds the standard, according to the pressure-power-speed linkage formula ΔP=0.08-0.12×ΔRa, ΔV=-0.04-0.06×ΔRa, the auxiliary gas pressure and cutting speed are automatically optimized; Wherein, ΔP is the power adjustment amount, ΔV is the speed adjustment amount.

5. The method of claim 1, wherein, S41, the deep reinforcement learning model is fused with: LSTM neural network, used to process time series data in the cutting process, and to mine the rules and potential relationships of each parameter changing with time; The Q-learning algorithm is used to find the optimal parameter combination in a complex parameter space, and the parameter optimization model is updated online through continuous trial and error and learning.

6. A laser cut end face quality control system, characterized by, The system comprises: A multi-modal perception module comprising a laser power sensor, a non-contact laser triangulation thickness gauge, a hyperspectral imager, and a dynamic pressure sensor, for collecting data such as the thickness of the workpiece, the material spectrum, the surface roughness, the laser power, and the auxiliary gas pressure; An intelligent decision-making module, which is internally provided with a deep reinforcement learning model, the deep reinforcement learning model fuses an LSTM neural network and a Q-learning algorithm, takes the data collected by the multi-modal perception module as input, and outputs the optimal parameter combination of the laser power, the cutting speed, the focal point position, the auxiliary gas type, and the pressure, and has an online learning capability; A collaborative execution module comprising a piezoelectric ceramic driven fast zoom cutting head, a three-way independent gas supply system, and a laser power adjustment module, the piezoelectric ceramic driven fast zoom cutting head has a focal point position adjustment response speed of 1000 Hz, the three-way independent gas supply system realizes millisecond-level switching of the gas type and gradient control of the pressure through proportional electromagnetic valves, the pressure adjustment precision is ±0.05 bar, the laser power adjustment module adopts pulse width modulation and power slope control technology, the power adjustment step is 0.05kW, and the response time is less than 10ms; The system is used to perform the laser cutting end face quality control method of any one of claims 1-5.

7. The laser-cut end face quality control system of claim 6, wherein, The intelligent decision-making module comprises a process parameter database, which stores an initial process parameter mapping table corresponding to different materials and different thicknesses.

8. The laser-cut end face quality control system of claim 6, wherein, The collaborative execution module further comprises a servo motor driven numerical control motion platform, which is used to control the translational motion of the laser cutting head in the X and Y axis directions, has a positioning accuracy of ≤±0.02mm, and realizes accurate cutting of different positions of the workpiece.

9. The laser-cut end face quality control system of claim 6, wherein, Further comprising an abnormality monitoring and emergency handling module, which is used to monitor various parameters in the cutting process in real time.

Images