Intelligent adjustment method, device and equipment for metal laser cutting

Abstract

The present invention relates to the field of laser cutting technology, and discloses an intelligent adjustment method, device and equipment for metal laser cutting. The method comprises: collecting data on the laser power, cutting area temperature and cutting head position of the laser cutting equipment to obtain a multidimensional data stream; inputting the multidimensional data stream into a deep subdomain adaptation network for feature extraction to obtain cutting feature data; calculating the local maximum average difference of the cutting feature data to obtain the local maximum average difference data; constructing a multi-objective optimization model including cutting error minimization, surface roughness minimization and cutting efficiency maximization according to the local maximum average difference data, and solving the optimal cutting parameter combination through a particle swarm algorithm. The present invention realizes precise adjustment of the cutting parameters and overcomes the problem of serious parameter coupling in traditional control methods.

Description

Technical Field

[0001] The present invention relates to the field of laser cutting technology, and in particular to an intelligent adjustment method, device and equipment for metal laser cutting. Background Art

[0002] Metal laser cutting technology is widely used in the processing of various metal materials due to its advantages of high precision, high efficiency and low pollution. However, in the actual cutting process, due to the influence of uncertain factors such as laser power fluctuations, changes in material properties, and interference from environmental factors, the cutting quality is often difficult to maintain stable, which brings great challenges to the optimization and control of cutting process parameters.

[0003] Traditional laser cutting control methods mainly rely on expert experience and simple feedback control, which is difficult to cope with complex and changeable cutting conditions. Although some studies have introduced machine learning methods to optimize cutting parameters, these methods are mostly based on offline training data and cannot effectively handle the dynamic changes in the actual cutting process. At the same time, existing control strategies often treat cutting parameters as independent variables, ignoring the coupling relationship between parameters, resulting in unsatisfactory control effects. Summary of the invention

[0004] The present invention provides an intelligent adjustment method, device and equipment for metal laser cutting, which realizes precise adjustment of cutting parameters and overcomes the problem of serious parameter coupling in traditional control methods.

[0005] In a first aspect, the present invention provides an intelligent adjustment method for metal laser cutting, the intelligent adjustment method for metal laser cutting comprising: Collect data on the laser power, cutting area temperature and cutting head position of the laser cutting equipment to obtain a multi-dimensional data stream; Inputting the multidimensional data stream into a deep subdomain adaptation network for feature extraction to obtain cutting feature data; Performing local maximum average difference calculation on the cutting feature data to obtain local maximum average difference data; A multi-objective optimization model including minimization of cutting error, minimization of surface roughness and maximization of cutting efficiency is constructed according to the local maximum average difference data, and the optimal cutting parameter combination is obtained by solving the model through the particle swarm algorithm.

[0006] In a second aspect, the present invention provides an intelligent adjustment device for metal laser cutting, the intelligent adjustment device for metal laser cutting comprising: The data acquisition module is used to collect data on the laser power, cutting area temperature and cutting head position of the laser cutting equipment to obtain a multi-dimensional data stream; A feature extraction module, used for inputting the multi-dimensional data stream into a deep sub-domain adaptation network for feature extraction to obtain cutting feature data; A difference calculation module, used for performing local maximum average difference calculation on the cutting feature data to obtain local maximum average difference data; The solution module is used to construct a multi-objective optimization model including minimizing cutting error, minimizing surface roughness and maximizing cutting efficiency according to the local maximum average difference data, and obtain the optimal cutting parameter combination through particle swarm algorithm.

[0007] The third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the above-mentioned intelligent adjustment method for metal laser cutting.

[0008] In the technical solution provided by the present invention, by introducing a deep subdomain adaptation network, effective migration of source domain data to target domain data is achieved, the distribution difference problem between the theoretical model and the actual cutting data is solved, and the accuracy of feature extraction is improved. The local maximum average difference calculation method is adopted to perform refined subdomain division and difference quantification of cutting features, which overcomes the problem of rough parameter space division in traditional methods and provides a more accurate evaluation basis for parameter optimization. The double-layer collaborative particle swarm algorithm is innovatively designed. Through the hierarchical structure of outer global search and inner local optimization, and the mutation strategy based on the local maximum average difference, the problem of traditional particle swarm algorithm easily falling into local optimality is effectively avoided. By establishing a multi-layer control architecture, combined with strategies such as dynamic parameter decoupling, hierarchical predictive control and neural network compensation, precise adjustment of cutting parameters is achieved, overcoming the serious problem of parameter coupling in traditional control methods. A cutting quality evaluation method based on multi-sensor fusion is introduced. Through the combination of three-dimensional scanning and metallographic analysis, a comprehensive evaluation of cutting quality is achieved, overcoming the limitation of insufficient information of a single detection method. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0010] Figure 1 A schematic diagram of the steps of the intelligent adjustment method for metal laser cutting in an embodiment of the present invention; Figure 2 It is a structural schematic diagram of an intelligent adjustment device for metal laser cutting in an embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] Embodiments of the present invention provide an intelligent adjustment method, device and equipment for metal laser cutting. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0012] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 , an embodiment of the intelligent adjustment method for metal laser cutting in the embodiment of the present invention includes: Step S1, collecting data on the laser power, cutting area temperature and cutting head position of the laser cutting equipment to obtain a multi-dimensional data stream; It is understandable that the execution subject of the present invention may be an intelligent adjustment device for metal laser cutting, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject as an example.

[0013] Specifically, a light intensity sensor is installed at the output end of the laser. The sensor detects the power intensity of the laser beam in real time and converts it into a digital signal to obtain laser power data. At the same time, an infrared array sensor is arranged around the cutting head. The sensor detects the thermal radiation intensity of the cutting area through the infrared band and forms a temperature distribution image. By extracting and analyzing these data, the temperature changes of various parts of the material during the cutting process are obtained. A grating scale displacement sensor is installed on the cutting head guide rail. This type of sensor uses optical encoding technology to obtain high-precision displacement information and converts it into three-dimensional coordinate data of the cutting head to grasp the spatial position of the cutting head at different time points. The laser power data, temperature distribution data and three-dimensional coordinate data of the cutting head are transmitted to the data processing unit via industrial Ethernet to form a raw data stream. The high-speed communication capability of industrial Ethernet ensures stable data transmission, and can effectively reduce delays and ensure data synchronization. At the same time, in the process of data transmission to the processing unit, a timestamp is added to each set of data to ensure that all data can be processed according to a unified time reference. The introduction of timestamps can ensure the accuracy of the data and help subsequent time series analysis, thereby avoiding the problem of data asynchrony. After the time stamp is completed, the data is rearranged in chronological order to form a synchronous data stream. The synchronous data stream is formatted. Since the data formats collected by different types of sensors are different, for example, the laser power data output by the light intensity sensor is expressed as current or voltage values, while the temperature data collected by the infrared array sensor exists in the form of a two-dimensional matrix, and the three-dimensional coordinate data of the grating scale displacement sensor uses an independent position information storage format, a unified standardized data format is required to integrate this information. During the formatting process, all data is converted into a unified standard data structure to ensure that subsequent deep learning models and optimization algorithms can smoothly parse this information. For example, laser power, temperature data, and cutting head position are stored as multidimensional arrays, where each record corresponding to a timestamp contains a laser power value, a set of temperature distribution data, and a set of three-dimensional coordinate data. Through the standardization process, a complete multidimensional data stream is formed.

[0014] Step S2, inputting the multidimensional data stream into the deep subdomain adaptation network for feature extraction to obtain cutting feature data; Specifically, the multidimensional data stream is divided into source domain data and target domain data, where the source domain data is the annotated data accumulated in the historical cutting process, including laser power, temperature distribution, cutting head position and its corresponding cutting quality evaluation index, while the target domain data is the current real-time collected cutting process data, which has not been annotated, so feature adaptation is performed through deep learning methods to ensure that the model can effectively transfer knowledge and make real-time predictions. The source domain data is input into the first feature extraction layer of the deep subdomain adaptation network. In this layer, feature mapping is performed through a five-layer convolution structure to gradually extract local spatial features and time series features in the data. This five-layer convolution structure consists of multiple two-dimensional convolution kernels, and each layer extracts more abstract and deep cutting features, such as laser energy distribution, temperature change pattern in the cutting area, and slight offset of the cutting head trajectory. At the same time, the target domain data is input into the second feature extraction layer of the deep subdomain adaptation network, which uses the same five-layer convolution structure as the first feature extraction layer to ensure the mapping consistency of the source domain and target domain data in the feature space, and ensure that the model can effectively learn the cutting characteristics of the target domain. After completing the feature mapping, the features extracted from the source domain data are input into the first domain adaptation layer of the deep subdomain adaptation network, which contains two layers of fully connected networks. It reduces the distribution difference between the source domain and the target domain in the feature space through feature transformation, so that the knowledge of the source domain can be transferred to the target domain. At the same time, the features extracted from the target domain data are input into the second domain adaptation layer of the deep subdomain adaptation network, which consists of two layers of fully connected networks, and aligns the data of the target domain with the data of the source domain through feature transformation, so that the feature expressions of the two are as similar as possible, enhancing the adaptability of the model. The core idea of ​​this domain adaptation process is to use the deep subdomain adaptation method to enable the model to learn the key factors of cutting quality from the source domain data and apply them to the target domain data. Even if the specific processing environment or material properties of the target domain are different, the accuracy of the prediction results can be guaranteed. The source domain adaptation features and the target domain adaptation features are input into the task classification layer of the deep subdomain adaptation network, and feature fusion is performed through a layer of fully connected network to integrate the data features of the source domain and the target domain, and finally output the same number of fusion features as the cutting quality evaluation indicators. These features include the influence of variables such as laser power, temperature distribution, and cutting head motion trajectory on cutting quality. The fused features are normalized, and the value ranges of different features are adjusted to the same scale using a standardization method to obtain standardized features. The standardized features are dimensionally reconstructed, that is, the data is reorganized according to the time and space dimensions, so that the cutting feature data can maintain its physical meaning and practical engineering application value.

[0015] Step S3, calculating the local maximum average difference of the cutting feature data to obtain the local maximum average difference data; Specifically, according to the three parameters of laser power, cutting speed and focal length, the cutting feature data is divided into subdomains and multiple subdomain feature sets, each of which corresponds to a specific set of cutting parameter combinations. This division method enables each subdomain to represent a process feature under specific cutting conditions and reflect the cutting performance under these conditions. A network hierarchical analysis is performed for each subdomain feature set. The corresponding feature vectors are extracted from the five-layer convolution structure that has been trained with a deep subdomain adaptation network. Through this analysis method, the feature representation of each subdomain at different levels is obtained. These feature vectors are obtained by extracting the original data layer by layer through the convolution layer, and contain the deep relationship between parameters such as laser power, cutting speed and focal length and cutting quality. Each layer of convolution structure abstracts the data from different dimensions to obtain a multi-level feature representation, which can more finely capture the subtle differences under different cutting conditions. The multi-layer feature vectors are input into the nearest neighbor search module. By applying the k-nearest neighbor algorithm, the Euclidean distance between samples in each subdomain is calculated to obtain the intra-class distance matrix. This matrix reflects the similarity between samples under different cutting conditions in the same subdomain, that is, the spatial distance between different feature data under the same cutting condition. Through this calculation, the closeness of samples in each subdomain is effectively evaluated, which helps to reveal the stability of the cutting process and the trend of quality change under the same parameter conditions. The weighted distance matrix in the subdomain is calculated adaptively. Based on the features extracted by each convolutional layer, these features are assigned corresponding weight coefficients, reflecting the importance of each feature in cutting quality optimization. The contribution of different features in the calculation is automatically adjusted through adaptive methods to ensure that important features can get more attention in subsequent calculations, thereby improving the optimization effect. The cross-subdomain distance calculation is performed on the multi-layer weighted features, the features under different cutting parameter combinations are paired, and the feature difference between subdomains is calculated by Mahalanobis distance. As an effective measurement method for measuring the difference between data points in multidimensional space, Mahalanobis distance considers the correlation between features according to the distribution structure of the data, so as to more accurately evaluate the difference between subdomains. This step helps to identify the trend and sensitivity of cutting quality under different cutting parameter combinations, providing richer differentiation information for the optimization algorithm. The intra-subdomain distance matrix and the inter-subdomain difference matrix are combined to obtain the local maximum average difference data. This data reflects the degree of feature difference within each subdomain and between subdomains under different cutting parameter conditions. By combining the intra-class and inter-class difference information, the parameter adjustment in the cutting process can be more effectively guided to ensure that while optimizing the cutting quality, it can reduce errors and improve cutting efficiency.

[0016] Step S4: construct a multi-objective optimization model including minimizing cutting error, minimizing surface roughness and maximizing cutting efficiency according to the local maximum average difference data, and obtain the optimal cutting parameter combination through particle swarm algorithm.

[0017] Specifically, subdomain mapping analysis is performed on the local maximum average difference data. By analyzing the distribution characteristics of the distance matrix within the subdomain and the difference matrix between subdomains, the dynamic adjustment range of each cutting parameter is calculated. The ranges of the three key parameters of laser power, cutting speed and focal length are divided into N subintervals, M subintervals and K subintervals respectively. The division of these subintervals can accurately reflect the changing trend of cutting quality under different cutting conditions and provide constraint conditions for parameter adjustment for subsequent optimization. Through subdomain mapping analysis, it is ensured that different cutting parameters are adjusted within a reasonable range, and the optimization direction under each cutting condition can be effectively constrained. An adaptive meshing model is constructed based on parameter adjustment constraints. The model dynamically meshes the cutting parameter space, and the mesh density is proportional to the local maximum average difference value. In areas with large local maximum average difference values, the mesh density will be denser, forming a fine-grained mesh, while in areas with small difference values, the mesh density will be sparser. The dynamic meshing method ensures that more refined search and optimization can be performed in areas with large changes in cutting quality, while unnecessary calculations are reduced in areas with small changes, thereby improving optimization efficiency. In this way, the parameter search space is obtained, which provides an effective search area for the particle swarm algorithm. The parameter search space is input into the objective function construction module. In this module, the cutting error function, surface roughness function and cutting efficiency function are weighted and combined into a fitness function. In order to make the optimization process more refined and targeted, the local maximum average difference value is added to the fitness function as a weight adjustment factor, so that the area with a large difference value can be given priority in the optimization process to achieve more efficient cutting quality optimization. The design of this fitness function enables the optimization process to better adapt to the changes of factors such as materials and temperature during the cutting process, thereby improving cutting accuracy and efficiency. The double-layer collaborative particle swarm algorithm is initialized, in which the outer particle swarm is responsible for global parameter search and the inner particle swarm performs local parameter optimization. The task of the outer particle swarm is to extensively explore the parameter space and find possible global optimal solutions, while the inner particle swarm focuses on more refined optimization in local areas. The collaborative work between the two layers of particle swarms is achieved by dynamically adjusting the search direction. Through the local maximum average difference data, the search direction of the two-layer particle swarm is dynamically adjusted, so that the outer particle swarm can effectively explore the parameter space, while the inner particle swarm can perform deeper local optimization in the area with large differences. In order to improve the efficiency and accuracy of the particle swarm algorithm, the outer particle swarm adopts a mutation degree adaptive strategy, dynamically adjusts the mutation probability according to the local maximum average difference value at each iteration, and ensures that the algorithm can adapt to the optimization needs of different parameter areas. The inner particle swarm adopts a local search enhancement strategy, focusing on fine search of high-difference areas to improve the accuracy and efficiency of the cutting process.After multiple iterative calculations, the optimization results are multi-level, which means that through multiple searches and local optimizations, the particle swarm algorithm finds multiple possible cutting parameter combinations. In order to select the optimal solution, the Pareto dominance relationship is used to sort all non-inferior solutions, and the global optimal solution is selected based on the sorting results. Pareto dominance relationship is a method used in multi-objective optimization. By comparing the pros and cons between different solutions, no other solution can be better than it at all objectives at the same time. In this way, a global optimal cutting parameter combination is obtained, including laser power, cutting speed and focal length, so as to achieve the goals of minimizing cutting errors, minimizing surface roughness and maximizing cutting efficiency. Through the multi-level optimization calculation and Pareto dominance relationship screening of the double-layer collaborative particle swarm algorithm, the optimal cutting parameter combination obtained can ensure that the cutting quality in the metal laser cutting process reaches the best level, while improving the cutting efficiency.

[0018] Based on the feature mapping matrix of the deep subdomain adaptation network, the coupling degree of the optimal cutting parameter combination is analyzed to understand the mutual influence between different cutting parameters, thereby revealing their contribution to the cutting quality. By calculating the independent adjustment components of each cutting parameter, a decoupled parameter matrix is ​​obtained, which reflects the independent role and contribution of each cutting parameter in the adjustment process. The decoupled parameter matrix helps to determine the optimal adjustment range of each cutting parameter, so that in actual operation, the change of each parameter can be more accurately controlled to avoid the problem of unstable cutting quality caused by excessive coupling. The decoupled parameter matrix is ​​subjected to hierarchical predictive control. Through a multi-level control structure, the various parameters in the cutting process are gradually optimized so that the final cutting quality reaches the best state. Through hierarchical control, optimization is performed separately at different levels to ensure that the parameters of each stage can be optimally adjusted under their specific constraints. The control sequence is predicted to predict future control requirements based on the current control state, and real-time adjustments are made based on the prediction results. Based on the hierarchical predictive control sequence, the dynamic characteristics of each actuator are learned online, and the control instructions are dynamically corrected according to real-time data to obtain the corresponding compensation control amount. This compensation control quantity effectively eliminates the deviation caused by environmental changes, equipment wear and other factors, ensuring the stability and efficiency of the cutting process. The cutting state deviation is fuzzified according to the compensation control quantity. The actual deviation data is converted into a fuzzy set to better reflect the degree and type of deviation. Through fuzzification, the control quantity with strong uncertainty and irregularity is converted into a form that is easy to process and analyze, which is convenient for subsequent control decisions. Based on the fuzzification results, a defuzzification operation is performed to convert the fuzzification results into specific adaptive control instructions. The key to the defuzzification process is to determine the most suitable control instructions according to the control objectives and system characteristics, and adjust the actions of the actuator in real time. Through this adaptive control instruction, the power, speed and optical path of laser cutting are dynamically adjusted according to the changes in actual working conditions, optimizing each link in the cutting process. The adaptive control instructions are sent to the power adjustment unit, speed control unit and optical path adjustment unit respectively through the field bus. These units adjust the laser power, cutting speed and focal position of the laser beam according to the received instructions to achieve precise control of the cutting process. In this way, the cutting equipment automatically adapts to the changes of different workpieces and materials to ensure that each cutting can be carried out under the best parameters. By filtering the data collected by the light intensity sensor, temperature sensor and displacement sensor, the noise components in the data are effectively removed to obtain more accurate measurement data. These filtered data are used as the input of the next control cycle to provide more accurate information support for the next dynamic adjustment. The data update mechanism ensures that the laser cutting system can continuously adapt to changes in the actual processing process and maintain continuous high performance. After the workpiece is cut, the adjusted cut workpiece is scanned in three dimensions and metallographically analyzed.Through 3D scanning, the size and shape of the cut workpiece are measured to check whether there are errors in the cutting process and ensure that the workpiece meets the design requirements. Metallographic analysis observes the microstructure of the cutting area to evaluate the surface quality, the morphology of the molten pool in the cutting area, and whether there are defects such as cracks in the cutting process.

[0019] The edge profile of the adjusted cut workpiece is scanned by 3D laser. By using 3D laser scanning technology, the edge shape of the cut workpiece is accurately captured, and the edge profile data is obtained to reflect any deformation, deviation or defect that occurs during the cutting process. The edge profile data is subjected to a five-point method verticality analysis. By uniformly sampling in the thickness direction of the workpiece and calculating the vertical deviation of each sampling point, it is evaluated whether the verticality of the cut edge meets the standard. The actual profile curve of the cut edge is fitted by the least squares method to obtain the verticality data, so as to more accurately measure the verticality error in the cutting process. The edge profile data is subjected to a light section method roughness analysis. The light section method effectively evaluates the roughness level of the cut surface by measuring the microstructure of the cut workpiece surface. This analysis method helps detect whether the cut surface is excessively rough or uneven. At the same time, the adjusted cut workpiece is subjected to metallographic microscopic analysis. Metallographic microscopic analysis observes the metal structure of the cut workpiece under a microscope to view the heat affected zone during the cutting process. The heat affected zone refers to the area where the microstructure of the workpiece metal changes due to high temperature during laser cutting. By observing the heat-affected zone, it is possible to find out whether there is excessive heat damage, cracks or other structural defects in the material during the cutting process. Metallographic microscopic analysis helps to evaluate the heat treatment effect of the material and reveal problems such as metal phase transformation and grain coarsening during laser cutting. Multi-scale morphological feature extraction is performed on the heat-affected zone data to extract the microstructure characteristics of the cut workpiece from multiple scale levels to obtain tissue characteristic data. These characteristic data include information such as grain size, phase distribution, and phase interface of the metal structure, which can provide a detailed perspective on the evolution of the metal structure during the cutting process. Based on verticality data, roughness data, and tissue characteristic data, a multi-dimensional cutting quality assessment is performed. The data of different dimensions are comprehensively analyzed to evaluate various aspects of the cutting quality, including cutting accuracy, surface finish, and the quality of the internal metal structure, and obtain cutting quality assessment data.

[0020] In the embodiment of the present invention, by introducing the deep subdomain adaptation network, the effective migration of source domain data to target domain data is realized, the distribution difference problem between the theoretical model and the actual cutting data is solved, and the accuracy of feature extraction is improved. The local maximum average difference calculation method is adopted to perform refined subdomain division and difference quantification of cutting features, which overcomes the problem of rough parameter space division in traditional methods and provides a more accurate evaluation basis for parameter optimization. The double-layer collaborative particle swarm algorithm is innovatively designed. Through the hierarchical structure of outer global search and inner local optimization, and the mutation strategy based on the local maximum average difference, the problem of traditional particle swarm algorithm easily falling into local optimality is effectively avoided. By establishing a multi-layer control architecture, combined with strategies such as dynamic parameter decoupling, hierarchical predictive control and neural network compensation, the precise adjustment of cutting parameters is realized, overcoming the serious parameter coupling problem in traditional control methods. A cutting quality evaluation method based on multi-sensor fusion is introduced. Through the combination of three-dimensional scanning and metallographic analysis, a comprehensive evaluation of cutting quality is achieved, overcoming the limitation of insufficient information of a single detection method.

[0021] In a specific embodiment, the process of executing step S1 may specifically include the following steps: Install a light intensity sensor at the output end of the laser to collect the laser power during the cutting process and obtain laser power data; Install infrared array sensors around the cutting head to scan the temperature of the cutting area and obtain temperature distribution data; Install a grating ruler displacement sensor on the cutting head guide rail to track the movement of the cutting head and obtain the three-dimensional coordinate data of the cutting head; The laser power data, temperature distribution data and cutting head three-dimensional coordinate data are transmitted to the data processing unit through the industrial Ethernet to obtain the original data stream, and the original data stream is time-stamped to obtain the marked data, and the marked data is sorted according to the time series to obtain the synchronous data stream; The synchronous data stream is formatted, and the data collected by different sensors are unified into a standard data format to obtain a multi-dimensional data stream.

[0022] Specifically, a light intensity sensor is installed at the output end of the laser to monitor the laser power in real time. The light intensity sensor measures the output power of the laser beam and converts the data into an electrical signal, which is then output to the data processing system. The laser power determines the cutting speed, cutting depth, and quality of the cutting surface. For example, assuming the laser power is , the power change is expressed as: ; in, is the variation of laser intensity over time, and are the start and end time points of the measurement respectively. By collecting laser power data in real time, the change of laser power over time is obtained. Infrared array sensors are installed around the cutting head to scan the temperature distribution of the cutting area. These sensors sense the heat generated during the cutting process through infrared rays and convert the temperature data of the cutting area into digital signals for transmission. The temperature distribution data helps to evaluate the size of the heat affected zone, the degree of heat influence, and the thermal deformation of the metal during the cutting process. For example, suppose the sensor measures the temperature at a certain location as ,in are spatial coordinates, is time. The change in temperature is expressed as a temperature field: ; in, is the temperature change at that point. By monitoring the temperature data of the cutting area in real time, it helps to adjust the laser power and cutting speed to reduce material damage or uneven cutting caused by overheating. Install a grating scale displacement sensor on the cutting head guide rail to track the three-dimensional position of the cutting head. The grating scale sensor senses the displacement of the cutting head on the X, Y and Z axes through the change of the grating stripes and obtains the three-dimensional coordinate data of the cutting head. Assume that the position of the cutting head is ,in ,and are the displacements of the three coordinate axes of the cutting head in space. The grating ruler accurately provides the position information of the cutting head to ensure that the laser beam always cuts along the predetermined path. The laser power data, temperature distribution data and the three-dimensional coordinate data of the cutting head are transmitted via industrial Ethernet. Industrial Ethernet is a high-speed and efficient data transmission network that transmits the data collected by the sensor to the data processing unit in real time. These data are transmitted in the form of raw data streams. Assume that the raw data stream is , where each data point Corresponding to a specific measurement value of laser power, temperature or displacement. In order to achieve time synchronization of multi-sensor data. Timestamp the original data stream, record the acquisition time of each data point, and obtain the marked data stream. Sort the marked data stream according to the time series to obtain a synchronous data stream, ensuring that the data collected by each sensor is completely synchronized in time and avoiding data dislocation caused by time delay. Format the synchronous data stream and convert the data from different sensors into a standard data format. This process includes steps such as unit conversion, data cleaning and format unification. Assuming that the unit of laser power is watt, the unit of temperature is Celsius, and the unit of displacement is millimeter, all data will be converted into a unified standard unit (for example, convert all data into Sl units). Through formatting, all data will be unified into a standard data format, which is convenient for subsequent calculation and analysis.

[0023] In a specific embodiment, the process of executing step S2 may specifically include the following steps: Divide the multidimensional data stream into source domain data and target domain data; The source domain data is input into the first feature extraction layer of the deep subdomain adaptation network, and feature mapping is performed through the five-layer convolution structure of the first feature extraction layer to obtain source domain feature data; The target domain data is input into the second feature extraction layer of the deep subdomain adaptation network, and feature mapping is performed through the five-layer convolution structure of the second feature extraction layer to obtain the target domain feature data; The source domain feature data is input into the first domain adaptation layer of the deep subdomain adaptation network, and the feature transformation is performed through the two-layer fully connected network of the first domain adaptation layer to obtain the source domain adaptation feature; Input the target domain feature data into the second domain adaptation layer of the deep subdomain adaptation network, and perform feature transformation through the two-layer fully connected network of the second domain adaptation layer to obtain the target domain adaptation features; The source domain adaptation features and the target domain adaptation features are input into the task classification layer of the deep subdomain adaptation network, and the features are fused through a layer of fully connected network. The output dimension is the same as the number of cutting quality evaluation indicators to obtain the fused features. The fused features are normalized to obtain standardized features, and the standardized features are dimensionally reconstructed to reorganize the features according to the time and space dimensions to obtain the cutting feature data.

[0024] Specifically, the collected multidimensional data stream is divided into source domain data and target domain data. The source domain data refers to the original data directly related to the cutting process, such as laser power, temperature distribution, and three-dimensional position of the cutting head, while the target domain data refers to the data obtained by transforming or inferring the source domain data. These data reflect certain indicators of cutting quality, such as cutting error or surface roughness. Suppose the source domain data is , where each is a feature data of the source domain, and the target domain data is , where each is a feature data of the target domain. The source domain data and the target domain data are collected simultaneously by multiple sensors, and they are consistent in time and space. The source domain data is input into the first feature extraction layer of the deep subdomain adaptation network for feature mapping. The first feature extraction layer in the deep subdomain adaptation network consists of five convolutional layers, which are used to extract deep features from the source domain data through convolution operations. In the first feature extraction layer, each convolutional layer performs convolution operations on the input data to generate a feature map. Assume that the input data is , the output feature map is obtained through convolution operation : ; in, Indicates Each convolution operation extracts different features, thus helping the network learn more complex feature representations. The filter weights of these convolutional layers are automatically learned during the training process through the back propagation algorithm. Through the convolution operation, the source domain data is convolved through five layers to obtain the source domain feature data. Similarly, the target domain data The second feature extraction layer is input to the deep subdomain adaptation network. The second feature extraction layer is also composed of five convolutional layers, which performs similar feature mapping on the target domain data to obtain the target domain feature data. ,Right now: ; Source domain feature data and target domain feature data The source domain feature data are extracted by the deep network respectively, and these feature representations are further used for feature transformation in the domain adaptation layer. The input is sent to the first domain adaptation layer of the deep subdomain adaptation network for feature transformation. The first domain adaptation layer consists of two layers of fully connected networks. Its goal is to map and transform the source domain features to better adapt to the feature distribution of the target domain. Assume that the input of the first domain adaptation layer is , the source domain adaptation feature is obtained through the full connection operation : ; in, Indicates The input features are transformed through multiple layers of fully connected layers to obtain the adapted features of the source domain. Similarly, the target domain feature data Input the second domain adaptation layer of the deep subdomain adaptation network, and after being processed by two layers of fully connected networks, the target domain adaptation features are obtained. : ; Adaptive characteristics and Reflects the changes in the source domain and target domain features during the adaptation process. and target domain adaptation features Input the task classification layer of the deep subdomain adaptation network. The task classification layer fuses the adaptation features of the source domain and the target domain through a fully connected network to obtain the fused features. The fused features represent different quality indicators in the cutting process, such as cutting error, surface roughness, and cutting efficiency. The output dimension of the fused features should be the same as the number of cutting quality evaluation indicators. Assume that the number of cutting quality evaluation indicators is , fusion features The dimension is ,Right now: ; in, represents the feature fusion operation, is the fully connected operation of the task classification layer. Through feature fusion, the adaptation features of the source domain and the target domain are integrated into a unified output, which represents the evaluation of each cutting quality index. Normalization is performed to make it more comparable and stable. Normalization is achieved through standardization, that is, the features are zero-mean and unit-varianced according to each dimension: ; in, and The fusion features are After standardization, the fusion feature Perform dimensional reconstruction and reorganize in time and space dimensions. Combine temporal and spatial features to assess quality and make adjustments in real time during the cutting process. Reconstructed feature data As the final cutting feature data: ; The obtained cutting feature data Used in the subsequent cutting quality optimization and adjustment process to provide real-time feedback to the laser cutting equipment to ensure quality control and performance improvement during the cutting process.

[0025] Among them, before the source domain adaptation features and the target domain adaptation features are input into the task classification layer of the deep subdomain adaptation network, it also includes: multi-scale feature reconstruction of the source domain adaptation features and the target domain adaptation features, channel grouping convolution through three layers of Ghost modules, each layer of Ghost module contains a main branch and an auxiliary branch, the main branch uses standard convolution to extract the main features, and the auxiliary branch uses deep separable convolution to enhance the features to obtain multi-scale feature data; the multi-scale feature data is input into the spatial attention module, the spatial attention map is generated through maximum pooling and average pooling, the spatial position information of the feature map is weighted to obtain the spatial enhanced features; channel attention calculation is performed on the spatial enhanced features, and the features of different channels are adaptively weighted through the squeeze-excitation mechanism to obtain the channel enhanced features; the channel enhanced features are input into the feature generalization module, and the residual connection and cross-layer feature fusion are performed. Perform feature generalization to obtain generalized feature data; construct a pyramid feature network based on the generalized feature data, and perform multi-scale feature fusion through top-down feature transfer and lateral connection to obtain a multi-level feature pyramid; construct a prediction head for the multi-level feature pyramid to predict the cutting area position, cutting parameter category and parameter confidence respectively to obtain predicted feature data; input the predicted feature data into the improved DeepSORT tracking module, combine the Kalman filter to predict the state of the cutting parameters, and perform parameter matching through the Hungarian algorithm to obtain the parameter tracking result; based on the parameter tracking result, perform feature enhancement on the source domain adaptation feature and the target domain adaptation feature to obtain the enhanced feature data, and the enhanced feature data is used to input the task classification layer of the deep subdomain adaptation network, and feature fusion is performed through a layer of fully connected network, and the output dimension is the same as the number of cutting quality evaluation indicators to obtain the fused feature.

[0026] In a specific embodiment, the process of executing step S3 may specifically include the following steps: According to the three parameters of laser power, cutting speed and focal length, the cutting feature data is divided into sub-domains to obtain a sub-domain feature set, and each sub-domain in the sub-domain feature set corresponds to a set of specific cutting parameter combinations; Perform network hierarchical analysis on each subdomain in the subdomain feature set, extract the corresponding feature vector from the five-layer convolution structure of the deep subdomain adaptation network, and obtain a multi-layer feature vector; The multi-layer feature vectors are input into the nearest neighbor search module, and the Euclidean distance between samples within each subdomain is calculated as the intra-class distance using the k-nearest neighbor algorithm to obtain the intra-subdomain distance matrix; Based on the distance matrix within the subdomain, weight adaptive calculation is performed, and corresponding weight coefficients are assigned to each convolution layer feature to obtain multi-layer weighted features; The multi-layer weighted features are used to calculate the cross-subdomain distance, and the features under different cutting parameter combinations are paired. The feature difference between subdomains is calculated by Mahalanobis distance to obtain the difference matrix between subdomains. The intra-subdomain distance matrix and the inter-subdomain difference matrix are combined to obtain the local maximum average difference data.

[0027] Specifically, according to the three key parameters of laser power, cutting speed and focal length, the cutting feature data is divided into different sub-domains. Each sub-domain corresponds to a specific set of cutting parameter combinations, for example, laser power Range is , cutting speed Range is ,focal length Range is In the parameter space, these three parameters are divided into multiple intervals. Suppose each parameter is divided into subintervals, cutting the subdomain feature set of feature data Include subdomains, each of which corresponds to a specific set of cutting parameter combinations , and in each subdomain, the feature data will show certain regularity and differences. For each subdomain, a deep subdomain adaptation network is used for feature extraction. The five-layer convolution structure of the deep subdomain adaptation network extracts multi-level deep features from the input cut feature data. In each convolution layer, the network learns local and global feature representations through convolution operations. For example, in the Convolutional layer, input features Perform convolution operation through the filter and output feature map : ; in, and They are The height and width of the layer convolution, is the number of channels of the input feature map, and the convolution kernel is obtained through learning. Through the continuous deepening of five layers of convolution, each layer of convolution abstracts different feature information at different levels. The final multi-layer feature vector Reflects the feature changes of the input data in each convolution layer. The multi-layer feature vector is input into the nearest neighbor search module for analysis. The nearest neighbor search module uses -Nearest neighbor algorithm calculates the Euclidean distance between samples within each subdomain , thus constructing the distance matrix within each subdomain. Assume that in a subdomain There are two samples and , their Euclidean distance The calculation formula is: ; in, is the dimension of the feature vector, and Respectively represent samples and In the By calculating the Euclidean distance between all samples, we can get the distance matrix in the subdomain , where each element represents the similarity between the corresponding samples. Based on the distance matrix within the subdomain, the weight of each layer of convolution features is adaptively calculated. The weight coefficient of each layer of features is dynamically adjusted according to the distribution of each sample in the subdomain. Assuming a feature vector In the The weight of the layer is , the weights are adaptively adjusted using the following formula: ; in, is a regulating factor, is the Euclidean distance between samples in the subdomain. In this way, samples farther from the cluster center will receive lower weights, while samples closer to the cluster center will receive higher weights. The weighted feature vector As the final feature output. The distance across subdomains is calculated based on the weighted multi-layer feature vector. The features under different cutting parameter combinations are paired and the difference between subdomains is calculated. In order to evaluate the difference between different subdomains, the commonly used distance metric is the Mahalanobis distance, which can consider the correlation between samples and give a more accurate difference metric. The Mahalanobis distance calculation formula is as follows: ; in, and are the eigenvectors of the two subdomains, is the sample covariance matrix, is the inverse matrix of the covariance matrix. By calculating the Mahalanobis distance, we can get the difference matrix between subdomains: , whose elements represent the degree of difference between the features of two subdomains. The distance matrix within the subdomain and the difference matrix between subdomains are combined to obtain the local maximum average difference data. By comprehensively considering the compactness within the subdomain and the difference between subdomains, the local maximum average difference data provides a refined analysis method based on the cutting parameter space, thereby optimizing the cutting parameter combination and improving the cutting accuracy and efficiency.

[0028] In this embodiment, before combining the distance matrix within the subdomain with the difference matrix between the subdomains to obtain the local maximum average difference data, it also includes: performing eigenvalue decomposition on the distance matrix within the subdomain, extracting the main eigenvector of each subdomain, and constructing a cutting parameter influence matrix based on the main eigenvector to obtain a parameter feature space; inputting the parameter feature space into a linear optimization module, taking the weighted sum of the Euclidean distance within the subdomain and the similarity of the cutting power curve as the optimization target, constructing a group of cutting parameter constraint equations to obtain a linear optimization model; constructing a Lagrangian function for the linear optimization model, introducing Lagrangian multipliers λ1, λ2 and λ3 corresponding to the constraints of the three parameters of laser power, cutting speed and focal length, respectively, to obtain a Lagrangian optimization function; constructing KK based on the Lagrangian optimization function T conditional equation group, alternately iteratively optimize the original variables and the dual variables to obtain the parameter optimization results; substitute the parameter optimization results into the simulation verification module, generate multiple groups of random cutting conditions through the Monte Carlo method, and theoretically verify the optimized cutting parameters to obtain a verification data set; conduct cutting experiment comparative analysis on the verification data set, obtain cutting quality data through physical experiments, and perform error analysis with theoretical calculation results to obtain double verification results; dynamically adjust the subdomain division boundary according to the double verification results, determine the new subdomain boundary threshold based on the verification error size, and obtain the optimized subdomain division scheme; use the optimized subdomain division scheme as the weight coefficient, perform weighted combination of the distance matrix within the subdomain and the difference matrix between subdomains, and obtain the local maximum average difference data.

[0029] In a specific embodiment, the process of executing step S4 may specifically include the following steps: Subdomain mapping analysis is performed on the local maximum average difference data. According to the distribution characteristics of the distance matrix within the subdomain and the difference matrix between subdomains, the dynamic adjustment range of each cutting parameter is calculated. The laser power range is divided into N sub-intervals, the cutting speed range is divided into M sub-intervals, and the focal length range is divided into K sub-intervals to obtain the parameter adjustment constraints. An adaptive meshing model is constructed according to the parameter adjustment constraints, and the cutting parameter space is dynamically meshed. The mesh density is proportional to the local maximum average difference value, and a fine-grained mesh is formed in the area with large difference values ​​to obtain the parameter search space. The parameter search space is input into the objective function construction module, the weighted sum of the cutting error function, the surface roughness function and the cutting efficiency function is used as the fitness function, and the local maximum average difference value is used as the weight adjustment factor to obtain an adaptive multi-objective optimization function; The double-layer collaborative particle swarm algorithm is initialized. The outer particle swarm is responsible for global parameter search, and the inner particle swarm performs local parameter optimization. The search directions of the two-layer particle swarms are dynamically adjusted through the local maximum average difference data to obtain a hierarchical optimization structure. Based on the hierarchical optimization structure, iterative calculation is performed. The outer particle swarm adopts the mutation degree adaptive strategy to dynamically adjust the mutation probability according to the local maximum average difference value. The inner particle swarm adopts the local search enhancement strategy to conduct a fine search for high-difference areas to obtain multi-level optimization results. The multi-level optimization results are screened and integrated, the non-inferior solution set is sorted through the Pareto dominance relationship, and the laser power, cutting speed and focal length corresponding to the global optimal position are taken as the optimal cutting parameter combination.

[0030] Specifically, the subdomain mapping analysis of the local maximum average difference data is carried out, and the dynamic adjustment range of each cutting parameter is determined according to the distribution characteristics of the distance matrix within the subdomain and the difference matrix between the subdomains. , cutting speed and focal length Adjust the three parameters and divide them into multiple sub-intervals. Divided into Sub-intervals, cutting speed Divided into sub-intervals, focal length Divided into subintervals, a multidimensional subinterval structure of the laser cutting parameter space is obtained, and the obtained parameter adjustment constraints As the basis for subsequent optimization processes. The design of parameter adjustment constraints ensures the accuracy and diversity of parameter selection during laser cutting, so that the entire operating space can be better covered during optimization, thereby improving cutting quality and efficiency. An adaptive meshing model is constructed based on parameter adjustment constraints. The density of the mesh is dynamically adjusted according to the local maximum average difference data. The mesh density is proportional to the local maximum average difference value. In areas with larger difference values, finer meshes are used, while in areas with smaller difference values, coarser meshes are used. Assuming that the local maximum average difference value is , the mesh density is calculated using the following formula: ; In the area with large differences, a finer grid will be formed to facilitate more detailed exploration of the characteristics of the area and obtain more accurate optimization results. The parameter search space obtained by dynamic meshing is input into the objective function construction module. In order to comprehensively consider the cutting error, surface roughness and cutting efficiency, the weighted sum of these three factors is used as the fitness function for optimization. Assume that the cutting error function is , the surface roughness function is , the cutting efficiency function is , fitness function It is expressed as: ; in, ,and are the weight coefficients of the three objectives, respectively, through the local maximum average difference value As a weight adjustment factor, it is dynamically adjusted. For example, the weight of each objective function is dynamically adjusted through the following formula: ; Local maximum average difference Different objective functions can be flexibly adjusted according to the size of the difference. The objective function is optimized using a double-layer collaborative particle swarm algorithm. The double-layer collaborative particle swarm algorithm introduces two layers of particle swarms for collaborative optimization: the outer particle swarm is responsible for global parameter search, while the inner particle swarm is responsible for local parameter optimization. During the search process of the outer particle swarm, the mutation probability of the particle is based on the local maximum average difference value. Dynamic adjustment. Assume that the mutation probability is , and its adjustment formula is expressed as: ; in, is the regulating factor, is the local maximum average difference value. As the difference value As the value of increases, the probability of mutation decreases, thereby enhancing the search accuracy in areas with smaller differences. The inner particle swarm adopts a local search enhancement strategy to perform a fine search on high-difference areas to ensure that a better combination of cutting parameters can be obtained in these areas. After multiple iterative calculations, a non-inferior solution set containing multiple optimization results is obtained. In order to screen out the optimal solution, the Pareto dominance relationship is used to sort the non-inferior solution set. The basic idea of ​​the Pareto dominance relationship is that if a solution is not inferior to another solution in all objectives and is superior to another solution in at least one objective, then the solution is considered to dominate the other solution. By sorting the non-inferior solution set, a globally optimal cutting parameter combination is determined. The laser power, cutting speed and focal length corresponding to the obtained global optimal position are the optimal cutting parameter combination.

[0031] In a specific embodiment, the intelligent adjustment method for performing metal laser cutting further includes the following steps: Based on the feature mapping matrix of the deep subdomain adaptation network, the coupling degree of the optimal cutting parameter combination is analyzed, and the independent adjustment components of each cutting parameter are calculated to obtain the decoupled parameter matrix. Perform hierarchical predictive control on the decoupling parameter matrix to obtain a hierarchical predictive control sequence, and based on the hierarchical predictive control sequence, perform online learning on the dynamic characteristics of each actuator and dynamically correct the control instructions to obtain the compensation control amount; According to the compensation control amount, the fuzzification result of the cutting state deviation is defuzzified to obtain the adaptive control instruction; The adaptive control instructions are respectively sent to the power adjustment unit, the speed control unit and the optical path adjustment unit through the field bus to obtain the execution adjustment results; Based on the execution adjustment result, the data collected by the light intensity sensor, the temperature sensor and the displacement sensor are subjected to noise suppression and state estimation to obtain filtered data, and the filtered data is used as the input of the next control cycle to obtain the adjusted cutting workpiece; The adjusted cut workpiece is subjected to three-dimensional scanning and metallographic analysis to obtain cutting quality evaluation data.

[0032] Specifically, the coupling degree of the optimal cutting parameter combination is analyzed based on the feature mapping matrix of the deep subdomain adaptation network. Assuming that the optimal laser cutting parameter combination includes laser power , cutting speed and focal length , these parameters represent the settings for optimal cutting quality under specific conditions. In order to analyze the coupling between these parameters, the deep subdomain adaptation network processes the input data through multiple feature extraction layers to extract a high-dimensional feature mapping matrix related to the cutting parameters. This mapping matrix can represent the relationship between the various cutting parameters. Assume that the matrix is , whose elements Indicates cutting parameters , and By analyzing the coupling degree of these features, the independent adjustment component of each parameter is obtained, that is, the independent contribution of each parameter to the cutting effect. At this time, the decoupled parameter matrix Calculated by the following formula: ; in, is a diagonal matrix, is the coupling coefficient vector, which indicates the adjustment strength of each cutting parameter. Through the decoupling operation, the influence of each cutting parameter on the cutting quality is analyzed separately to avoid interference between multiple parameters. The decoupling parameter matrix is ​​hierarchically predicted and controlled, and the system is optimized step by step through a multi-level control structure. , design an independent predictive control model for each cutting parameter. Assume that a model-based control system is set up, in which the predictive control sequence is , indicating that at time The goal of the system control is to make each cutting parameter adjustment accurately in a dynamic environment. Through the hierarchical control strategy, each level of the controller is adjusted according to the output of the previous level controller. For example, at the outer control level, the global optimization goal is used for control, and the inner control level is fine-tuned according to the local feedback data. Based on the predictive control sequence , the dynamic characteristics of the actuator are incorporated into the control model to achieve online learning and dynamic correction of control instructions. At this time, the dynamic characteristics of the actuator are modeled, assuming that the dynamic response of the actuator is ,in Indicates the state of the actuator. Indicates time. The control system dynamically modifies the control instructions based on the response information. , ensuring the system can operate stably under different operating conditions. The generation of can realize the dynamic adjustment of cutting state deviation. Assuming that the correction value output by the control system The system state converges to the target state. In order to ensure the accuracy of feedback control, the fuzzy control mechanism is used to fuzzify the cutting state deviation, and then the final adaptive control instruction is obtained through defuzzification operation. The fuzzification process is expressed using the following formula: ; in, Indicates the state deviation Fuzzy processing is performed to obtain the compensation control quantity. The fuzzy controller adjusts the parameters such as the rule base and the membership function so that the control instructions can adapt to various changing operating environments. It is sent to each execution unit for control through the field bus system, including the power adjustment unit, speed control unit and optical path adjustment unit. The role of the field bus is to ensure the coordinated work between the various execution units so that each cutting parameter can be accurately adjusted. After the control instruction is adjusted, the corresponding execution adjustment result is generated, which indicates the real-time adjustment of the cutting process by the system. Based on the execution adjustment result, the cutting process is fully monitored through the sensor network. The light intensity sensor, temperature sensor and displacement sensor respectively collect data such as light intensity, temperature and displacement during the cutting process. In order to ensure the accuracy and reliability of the data, the data collected by these sensors are subjected to noise suppression and state estimation. Noise suppression methods include Kalman filtering, mean filtering, etc. Assuming that the Kalman filter is used for data filtering, its state update formula is: ; in, represents the estimated system state, is the observed value, is the Kalman gain, is the observation matrix, is the predicted state. The filtered data is used as the input for the next control cycle, and finally the adjusted cut workpiece is generated. The adjusted cut workpiece is subjected to 3D scanning and metallographic analysis to evaluate the cutting quality. 3D scanning can accurately capture the geometry of the cutting edge, and metallographic analysis can reveal the heat-affected zone and microstructural changes during the cutting process. Through this step, the cutting quality evaluation data is obtained. .

[0033] In a specific embodiment, the process of executing step S6 may specifically include the following steps: Performing three-dimensional laser scanning on the edge profile of the adjusted cut workpiece to obtain edge profile data; The five-point verticality analysis of the edge profile data is performed, sampling is performed evenly in the thickness direction of the workpiece and the vertical deviation of each sampling point is calculated. The actual profile curve of the cutting edge is fitted by the least squares method to obtain the verticality data; Performing light section roughness analysis on edge profile data to obtain roughness data; Performing metallographic microscopic analysis on the adjusted cut workpiece to obtain heat-affected zone data, and extracting multi-scale morphological features from the heat-affected zone data to obtain organizational feature data; A multi-dimensional cutting quality assessment is performed based on verticality data, roughness data and tissue feature data to obtain cutting quality assessment data.

[0034] Specifically, the edge contour of the adjusted cut workpiece is scanned by 3D laser. 3D laser scanning technology scans the surface of the workpiece with a laser beam and constructs a 3D model of the workpiece surface by receiving the reflected light signal. Assume that multiple points are scanned on the edge of the workpiece to obtain edge contour data. , which represents the three-dimensional coordinate information of each point on the workpiece surface. The point cloud data obtained by scanning is expressed by the following formula: ; in, For the The three-dimensional coordinates of each scanning point indicate the position of the point in space. After obtaining the edge profile data, a five-point method verticality analysis is performed. Verticality refers to the deviation between the edge of the cut workpiece and the thickness direction of the workpiece. Multiple sampling points are uniformly selected in the thickness direction of the workpiece to calculate the vertical deviation between these points and the ideal edge. Set uniform sampling in the thickness direction of the workpiece. points, marked as in , the verticality analysis is performed by calculating the vertical deviation between each sampling point and the ideal edge. The specific calculation method is obtained by fitting the actual contour curve of the cutting edge by the least squares method. Assume that the equation of the ideal edge is , then the verticality error of each point is expressed as: ; in, Indicates The verticality error of each point is obtained by fitting the actual contour curve of the cutting edge using the least squares method. , which is the vertical deviation set of all sampling points. Perform light section roughness analysis. Roughness is an indicator of the irregularity and details of the workpiece surface, and is analyzed using the light section method. In this method, a light source is used to illuminate the workpiece surface, and the reflection difference between the light and the surface is used to calculate the surface roughness. Assume that the light section method can give the surface height of each point , the local roughness is calculated by the following formula: ; in, is the arithmetic mean roughness, is the average height of all sampling points, It is The surface height of the sampling points, is the total number of sampling points. Roughness data The above formula is used to calculate the roughness characteristics of the workpiece surface in different areas. Metallographic microscopic analysis is performed to study the heat affected zone of the workpiece. Metallographic microscopic analysis magnifies the workpiece surface through a microscope and observes its microstructure, especially near the laser cutting area, to reveal the structural changes in the heat affected zone. The heat affected zone refers to the area where the metal material undergoes microstructural changes due to high temperature during the cutting process. Images are taken through a microscope and processed to extract the characteristic data of the heat affected zone. These data can reflect the effect of heat input on the workpiece structure during cutting. In order to more comprehensively evaluate the changes in the heat-affected zone, multi-scale morphological feature extraction is performed. Through image processing techniques such as edge detection and morphological analysis, the organizational feature data at different scales are obtained. These data reflect the organizational structure of the workpiece surface at different scales. Based on all the above data, multi-dimensional cutting quality evaluation is performed. , Roughness data , organizational characteristics data and heat affected zone data , comprehensively evaluate the cutting quality. These different feature data are integrated through the weighted linear combination method to obtain the final cutting quality evaluation data , and its calculation formula is as follows: ; in, It is the weight coefficient of each index, which is adjusted according to different process requirements. The final cutting quality evaluation data It is a comprehensive evaluation value that reflects the overall cutting quality of the workpiece.

[0035] The above describes the intelligent adjustment method for metal laser cutting in the embodiment of the present invention. The following describes the intelligent adjustment device for metal laser cutting in the embodiment of the present invention. Figure 2 , an embodiment of the intelligent adjustment device for metal laser cutting in the embodiment of the present invention includes: The data acquisition module is used to collect data on the laser power, cutting area temperature and cutting head position of the laser cutting equipment to obtain a multi-dimensional data stream; A feature extraction module is used to input the multi-dimensional data stream into the deep sub-domain adaptation network for feature extraction to obtain cutting feature data; A difference calculation module is used to perform local maximum average difference calculation on the cutting feature data to obtain local maximum average difference data; The solution module is used to construct a multi-objective optimization model including minimization of cutting error, minimization of surface roughness and maximization of cutting efficiency according to the local maximum average difference data, and obtain the optimal cutting parameter combination through the particle swarm algorithm.

[0036] Through the collaboration of the above components and the introduction of deep subdomain adaptation network, the effective migration of source domain data to target domain data is realized, the distribution difference problem between theoretical model and actual cutting data is solved, and the accuracy of feature extraction is improved. The local maximum average difference calculation method is used to perform refined subdomain division and difference quantification of cutting features, which overcomes the problem of rough parameter space division in traditional methods and provides a more accurate evaluation basis for parameter optimization. The double-layer collaborative particle swarm algorithm is innovatively designed. Through the hierarchical structure of outer global search and inner local optimization, as well as the mutation strategy based on local maximum average difference, the problem of traditional particle swarm algorithm easily falling into local optimality is effectively avoided. By establishing a multi-layer control architecture, combined with dynamic parameter decoupling, hierarchical predictive control and neural network compensation strategies, the precise adjustment of cutting parameters is achieved, overcoming the serious parameter coupling problem in traditional control methods. The cutting quality evaluation method based on multi-sensor fusion is introduced. Through the combination of three-dimensional scanning and metallographic analysis, a comprehensive evaluation of cutting quality is achieved, overcoming the limitation of insufficient information of a single detection method.

[0037] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0038] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0039] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0040] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0041] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent adjustment method for metal laser cutting, characterized in that: The method comprises: Collect data on the laser power, cutting area temperature and cutting head position of the laser cutting equipment to obtain a multi-dimensional data stream; Inputting the multidimensional data stream into a deep subdomain adaptation network for feature extraction to obtain cutting feature data; Performing local maximum average difference calculation on the cutting feature data to obtain local maximum average difference data; A multi-objective optimization model including minimization of cutting error, minimization of surface roughness and maximization of cutting efficiency is constructed according to the local maximum average difference data, and the optimal cutting parameter combination is obtained by solving the model through the particle swarm algorithm.

2. The intelligent adjustment method for metal laser cutting according to claim 1, characterized in that: The laser power, cutting area temperature and cutting head position of the laser cutting equipment are collected to obtain a multi-dimensional data stream, including: Install a light intensity sensor at the output end of the laser to collect the laser power during the cutting process and obtain laser power data; Install infrared array sensors around the cutting head to scan the temperature of the cutting area and obtain temperature distribution data; Install a grating ruler displacement sensor on the cutting head guide rail to track the movement of the cutting head and obtain the three-dimensional coordinate data of the cutting head; The laser power data, the temperature distribution data and the three-dimensional coordinate data of the cutting head are transmitted to a data processing unit via industrial Ethernet to obtain an original data stream, and the original data stream is timestamped to obtain marked data, and the marked data is sorted according to a time series to obtain a synchronous data stream; The synchronous data stream is subjected to data formatting processing, and the data collected by different sensors are unified into a standard data format to obtain a multi-dimensional data stream.

3. The intelligent adjustment method for metal laser cutting according to claim 1, characterized in that: The step of inputting the multi-dimensional data stream into a deep sub-domain adaptation network for feature extraction to obtain cutting feature data comprises: Dividing the multidimensional data stream into source domain data and target domain data; Inputting the source domain data into the first feature extraction layer of the deep sub-domain adaptation network, performing feature mapping through the five-layer convolution structure of the first feature extraction layer, and obtaining source domain feature data; Inputting the target domain data into the second feature extraction layer of the deep sub-domain adaptation network, performing feature mapping through the five-layer convolution structure of the second feature extraction layer, and obtaining target domain feature data; Inputting the source domain feature data into the first domain adaptation layer of the deep subdomain adaptation network, performing feature transformation through the two-layer fully connected network of the first domain adaptation layer, and obtaining source domain adaptation features; Inputting the target domain feature data into the second domain adaptation layer of the deep subdomain adaptation network, performing feature transformation through a two-layer fully connected network of the second domain adaptation layer, and obtaining target domain adaptation features; Inputting the source domain adaptation features and the target domain adaptation features into the task classification layer of the deep subdomain adaptation network, performing feature fusion through a layer of fully connected network, and outputting the same dimension as the number of cutting quality evaluation indicators to obtain fused features; The fused features are normalized to obtain standardized features, and the standardized features are dimensionally reconstructed to reorganize the features according to time and space dimensions to obtain cutting feature data.

4. The intelligent adjustment method for metal laser cutting according to claim 3 is characterized in that: The step of calculating the local maximum average difference of the cutting feature data to obtain the local maximum average difference data includes: According to the three parameters of laser power, cutting speed and focal length, the cutting feature data is divided into sub-domains to obtain a sub-domain feature set, wherein each sub-domain in the sub-domain feature set corresponds to a set of specific cutting parameter combinations; Performing network hierarchical analysis on each subdomain in the subdomain feature set, extracting corresponding feature vectors from the five-layer convolution structure of the deep subdomain adaptation network, and obtaining multi-layer feature vectors; The multi-layer feature vector is input into the nearest neighbor search module, and the Euclidean distance between samples within each subdomain is calculated as the intra-class distance using the k-nearest neighbor algorithm to obtain the intra-subdomain distance matrix; Perform weight adaptive calculation based on the distance matrix in the subdomain, assign corresponding weight coefficients to each convolution layer feature, and obtain multi-layer weighted features; Performing cross-subdomain distance calculation on the multi-layer weighted features, pairing the features under different cutting parameter combinations, calculating the feature difference between subdomains by Mahalanobis distance, and obtaining the difference matrix between subdomains; The intra-subdomain distance matrix and the inter-subdomain difference matrix are combined and operated to obtain local maximum average difference data.

5. The intelligent adjustment method for metal laser cutting according to claim 1, characterized in that: The multi-objective optimization model including minimizing cutting error, minimizing surface roughness and maximizing cutting efficiency is constructed according to the local maximum average difference data, and the optimal cutting parameter combination is obtained by solving the model through the particle swarm algorithm, including: Perform subdomain mapping analysis on the local maximum average difference data, calculate the dynamic adjustment range of each cutting parameter according to the distribution characteristics of the distance matrix within the subdomain and the difference matrix between the subdomains, divide the laser power range into N subintervals, divide the cutting speed range into M subintervals, and divide the focal length range into K subintervals to obtain parameter adjustment constraints; An adaptive meshing model is constructed according to the parameter adjustment constraints, and the cutting parameter space is dynamically meshed, wherein the mesh density is proportional to the local maximum average difference value, and a fine-grained mesh is formed in the area with a large difference value to obtain a parameter search space; The parameter search space is input into the objective function construction module, the weighted sum of the cutting error function, the surface roughness function and the cutting efficiency function is used as the fitness function, and the local maximum average difference value is used as the weight adjustment factor to obtain an adaptive multi-objective optimization function; The double-layer collaborative particle swarm algorithm is initialized. The outer particle swarm is responsible for global parameter search, and the inner particle swarm performs local parameter optimization. The search directions of the two-layer particle swarms are dynamically adjusted through the local maximum average difference data to obtain a hierarchical optimization structure. Based on the hierarchical optimization structure, iterative calculations are performed, the outer particle swarm adopts a mutation degree adaptive strategy to dynamically adjust the mutation probability according to the local maximum average difference value, and the inner particle swarm adopts a local search enhancement strategy to perform a fine search on the high difference area to obtain a multi-level optimization result; The multi-level optimization results are screened and integrated, the non-inferior solution set is sorted by the Pareto dominance relationship, and the laser power, cutting speed and focal length corresponding to the global optimal position are used as the optimal cutting parameter combination.

6. The intelligent adjustment method for metal laser cutting according to claim 1, characterized in that: The intelligent adjustment method for metal laser cutting also includes: Based on the feature mapping matrix of the deep subdomain adaptation network, the coupling degree of the optimal cutting parameter combination is analyzed, and the independent adjustment component of each cutting parameter is calculated to obtain a decoupling parameter matrix; Performing hierarchical predictive control on the decoupling parameter matrix to obtain a hierarchical predictive control sequence, and based on the hierarchical predictive control sequence, performing online learning and dynamic correction of control instructions on the dynamic characteristics of each actuator to obtain a compensation control amount; According to the compensation control amount, a defuzzification operation is performed on the fuzzified result of the cutting state deviation to obtain an adaptive control instruction; The adaptive control instruction is sent to the power adjustment unit, the speed control unit and the optical path adjustment unit respectively through the field bus to obtain the execution adjustment result; Based on the execution adjustment result, the data collected by the light intensity sensor, the temperature sensor and the displacement sensor are subjected to noise suppression and state estimation to obtain filtered data, and the filtered data is used as the input of the next control cycle to obtain the adjusted cutting workpiece; The adjusted cut workpiece is subjected to three-dimensional scanning and metallographic analysis to obtain cutting quality evaluation data.

7. The intelligent adjustment method for metal laser cutting according to claim 6, characterized in that: The three-dimensional scanning and metallographic analysis of the adjusted cut workpiece to obtain cutting quality evaluation data includes: Performing three-dimensional laser scanning on the edge profile of the adjusted cut workpiece to obtain edge profile data; Performing a five-point verticality analysis on the edge profile data, uniformly sampling in the thickness direction of the workpiece and calculating the vertical deviation of each sampling point, fitting the actual profile curve of the cutting edge by the least squares method to obtain the verticality data; Performing light section roughness analysis on the edge profile data to obtain roughness data; Performing metallographic microscopic analysis on the adjusted cut workpiece to obtain heat-affected zone data, and performing multi-scale morphological feature extraction on the heat-affected zone data to obtain tissue feature data; A multi-dimensional cutting quality assessment is performed based on the verticality data, the roughness data and the tissue feature data to obtain cutting quality assessment data.

8. An intelligent adjustment device for metal laser cutting, characterized in that: The intelligent adjustment method for metal laser cutting according to any one of claims 1 to 7 is used, and the intelligent adjustment device for metal laser cutting comprises: The data acquisition module is used to collect data on the laser power, cutting area temperature and cutting head position of the laser cutting equipment to obtain a multi-dimensional data stream; A feature extraction module, used for inputting the multi-dimensional data stream into a deep sub-domain adaptation network for feature extraction to obtain cutting feature data; A difference calculation module, used for performing local maximum average difference calculation on the cutting feature data to obtain local maximum average difference data; The solution module is used to construct a multi-objective optimization model including minimizing cutting error, minimizing surface roughness and maximizing cutting efficiency according to the local maximum average difference data, and obtain the optimal cutting parameter combination through particle swarm algorithm.

9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and is characterized in that when the processor executes the computer program, the intelligent adjustment method for metal laser cutting described in any one of claims 1 to 7 is implemented.

Images