Machine learning based adaptive machining control method and system for numerical control machine tool

Abstract

This invention relates to the field of machine tool control technology, and particularly to a machine learning-based adaptive machining control method and system for CNC machine tools. First, a digital material sample is constructed based on the workpiece material grade and heat treatment state, establishing a quantitative mapping relationship set between microstructure and macroscopic properties. Then, a digital twin is constructed by combining the tool and machine tool models to predict microstructure evolution and update material properties. Next, using multi-objective optimization and material micro-fluctuations as variables, an optimal robustness parameter package and parameter sensitivity map are obtained. During actual machining, this parameter package is loaded, and an online fine-tuning module generates fine-tuning instructions based on real-time sensor data and the sensitivity map. Finally, the instructions are executed to complete the machining, and the digital twin is updated based on the machining data. By implementing this invention, the consistency of machining quality and the stability of machining efficiency for high-end parts can be significantly improved.

Description

Technical Field

[0001] This invention relates to the field of machine tool control technology, and in particular to a machine learning-based adaptive machining control method and system for CNC machine tools. Background Technology

[0002] In the field of high-precision manufacturing, the quality stability of CNC machining heavily depends on the uniformity of the workpiece material. However, even for metal materials of the same grade, there are inherent fluctuations in their internal microstructure, such as grain size and orientation, which leads to changes in the mechanical response during the cutting process, causing problems such as cutting force fluctuations, abnormal tool wear, and surface quality variations. Traditional adaptive control methods are mostly based on real-time sensor feedback during the machining process for passive response, such as the scheme disclosed in CN120408868B. Their optimization focus is on compensating for macroscopic cutting forces or deformations, and they cannot proactively address the source disturbances caused by the microscopic inhomogeneity of the material. On the other hand, existing process optimization based on digital twins, such as the scheme disclosed in CN120124316B, although focusing on the correlation between macroscopic deformation and microstructure, rely heavily on historical production data to build a process database. This results in long modeling cycles, high costs, and optimization objectives that are mostly focused on avoiding defects rather than actively seeking robust process parameters that are insensitive to microscopic fluctuations. They do not take reducing the sensitivity of process results to small changes in the microstructure of the material as a core optimization objective. This means that even if an optimal parameter is found, it may become invalid if the material batches are slightly different, leading to quality fluctuations.

[0003] Therefore, existing technologies struggle to identify and avoid process-sensitive areas caused by differences in the microstructure of materials before processing. They lack a forward-looking parameter optimization and adaptive control method that starts from the intrinsic properties of materials and is guided by robust design, which restricts the further improvement of the consistency of processing quality of high-end parts. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies in lacking a forward-looking parameter optimization and adaptive control method based on the intrinsic properties of materials and guided by robust design, by providing a machine learning-based adaptive machining control method and system for CNC machine tools.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a machine learning-based adaptive machining control method for CNC machine tools, comprising: constructing a digital material specimen reflecting the initial microstructural characteristics of the target workpiece material based on the grade and initial heat treatment state of the target workpiece material, and establishing a quantitative mapping relationship library between the microstructural characteristics and macroscopic mechanical / thermal performance parameters as a material-performance mapping relationship set; Based on the material-property mapping relationship set, as well as the preset tool geometry model and machine tool dynamics model, a digital twin is constructed. The digital twin can predict the microstructure evolution of the workpiece material based on the thermal-mechanical load history of local processing and update the material properties based on the microstructure evolution. With processing quality, processing efficiency and microstructure evolution indicators as optimization targets, and the inherent fluctuations of the microstructure characteristics of the target workpiece material as uncertainty variables, the processing parameters of the target processing task are optimized based on the digital twin to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map. In actual processing, the optimal robustness parameter package is loaded, and the online fine-tuning module is started simultaneously. The online fine-tuning module, based on real-time sensor data, infers the equivalent microstate offset of the workpiece processing area through the built-in microstate observer, and generates parameter fine-tuning instructions within the preset robustness envelope by combining the parameter sensitivity map. The optimal robustness parameter package and the real-time generated parameter fine-tuning instructions are executed to complete the processing, and the digital twin is updated based on the processing data.

[0006] Optionally, based on the grade and initial heat treatment state of the target workpiece material, a digital material specimen reflecting the initial microstructural characteristics of the target workpiece material is constructed, including: Obtain the grade of the target workpiece material and its corresponding initial heat treatment process; Based on the grade, the material database is queried to obtain the standard chemical composition, phase diagram and basic physical property parameters of the target workpiece material; If measured data of the target workpiece material exists, import the electron backscatter diffraction or metallographic image of the target workpiece material, and digitally reconstruct the real microstructure model containing grains, grain boundaries and second phase through image segmentation and vectorization processing. If no measured data of the target workpiece material is available, then based on the chemical composition, initial heat treatment process and material phase transformation kinetics principle, a virtual microstructure model that is statistically representative is generated by cellular automata or Monte Carlo method. Based on the real microstructure model or the virtual microstructure model, a digital model expressed in the form of a three-dimensional voxel mesh is constructed for the target workpiece material as the digital material specimen, wherein each mesh in the digital material specimen is associated with a corresponding material phase and crystallographic direction.

[0007] Optionally, a quantitative mapping relationship library between the microstructural features and macroscopic mechanical / thermal performance parameters is established as a material-property mapping relationship set, including: Using the digital material sample as input, the macroscopic mechanical / thermal properties under different combinations of microstructure features are predicted by performing crystal plasticity finite element simulation calculations. The macroscopic mechanical property parameters include at least yield strength, flow stress and strain hardening index, and the macroscopic thermal property parameters include at least thermal conductivity and specific heat capacity. By using microstructural features as independent variables and the macroscopic mechanical / thermal properties obtained from simulation as dependent variables, a structured and queryable mapping relationship database is constructed as the material-property mapping relationship set.

[0008] Optionally, based on the material-property mapping relationship set, and the preset tool geometry model and machine tool dynamics model, a digital twin is constructed, including: Obtain the preset tool geometry model and machine tool dynamics model; Based on the aforementioned material-property mapping relationship set, the preset tool geometry model, and the machine tool dynamics model, a digital twin of the macroscopic thermo-mechanical coupling machining process is established; Within each macroscopic material property calculation unit of the digital twin, a microstructure evolution calculation unit is nested. The microstructure evolution calculation unit calculates the microscopic property data within the macroscopic material property calculation unit by calling the embedded crystal plasticity finite element model based on the macroscopic property data provided by the macroscopic material property calculation unit to which it is located. The macroscopic performance data includes local strain, strain rate, and temperature history data; the microscopic performance data includes the evolution of grain size, dislocation density, and phase fraction. A data transmission channel is established from the macroscopic material property calculation unit to the microstructure evolution calculation unit for real-time transmission of the macroscopic property data; Simultaneously, a data feedback channel is established to return from the microstructure evolution calculation unit to the macromaterial property calculation unit, which is used to update the material properties in the macromaterial property calculation unit based on the calculated microstructure property data.

[0009] Optionally, with processing quality, processing efficiency, and microstructure evolution indicators as optimization objectives, and the inherent fluctuations of the microscopic characteristics of the target workpiece material as uncertainty variables, the processing parameters of the target processing task are optimized based on the digital twin to obtain an optimal robustness parameter package and a corresponding parameter sensitivity map, including: Define an optimization multi-objective function, with objective terms including at least processing quality, processing efficiency, and microstructure evolution indicators; Among them, the processing quality target is characterized by the surface roughness and residual stress predicted by the digital twin, the processing efficiency target is characterized by the material removal rate, and the microstructure evolution index is characterized by the degree of subsurface grain refinement and the phase change layer depth. The statistical distribution of key microscopic features extracted from the digital material specimens is set as an interval-type uncertainty variable in the optimization process. Based on the digital twin, a multi-objective robust optimization algorithm is used to perform global optimization within the processing parameter space; During the optimization process, for each set of candidate processing parameters, multiple sampling simulations are performed within the range of the uncertainty variables to calculate the mean and variance of the performance of each objective item. The optimal robustness parameter package is the combination of processing parameters with the best overall performance and the smallest variance in the Pareto front. The parameter sensitivity map calculates the gradient of the influence of each processing parameter on each sub-indicator in the multi-objective function and presents it in the form of a visual data matrix.

[0010] Optionally, the preset robustness envelope is defined in the following way: Based on the aforementioned optimal robustness parameter package, parameter perturbation analysis is performed in the digital twin; To obtain the maximum positive and negative adjustment range of each key processing parameter while keeping all processing quality indicators and microstructure evolution indicators in line with preset standards; The maximum positive and maximum negative adjustment amplitudes together form a super-rectangular region in a multi-dimensional parameter space, and the super-rectangular region is used as the preset robustness envelope.

[0011] Optionally, the construction of the microstate observer includes: Based on the digital twin, a dataset containing various simulated microscopic state fluctuations and corresponding sensor response characteristics is generated; A lightweight machine learning model is trained using the dataset. The input of the machine learning model includes time-frequency domain features of real-time cutting force, vibration, and acoustic emission signals. The output includes estimates of the equivalent hardness and equivalent toughness state of the current machining area of ​​the target workpiece. The trained machine learning model is deployed as the microstate observer.

[0012] Optionally, based on the parameter sensitivity map, parameter fine-tuning instructions are generated within a preset robustness envelope, including: The parameter sensitivity map records the gradient of the influence of each processing parameter on the processing quality index. When the microstate observer infers that the equivalent microstate has shifted, it queries the parameter sensitivity map based on the direction and magnitude of the shift to identify the adjustment parameters and their adjustment directions that can most effectively compensate for the impact of the shift on key quality indicators. Based on the adjustment parameters and their adjustment direction, and in accordance with predefined compensation rules, parameter fine-tuning instructions are generated under the constraints of the robustness envelope.

[0013] Optionally, the optimal robustness parameter package and the real-time generated parameter fine-tuning instructions are executed to complete the processing, and the digital twin is updated based on the processing data, including: In the actual processing, the processing control system loads and executes the optimal robustness parameter package as the basic processing instructions, and at the same time receives and integrates the parameter fine-tuning instructions generated in real time by the online fine-tuning module to form the final execution instruction sequence and perform the processing. By deploying a sensor network on the machine tool and the target workpiece, the actual cutting force, vibration, acoustic emission and temperature data are collected synchronously throughout the entire machining cycle, which serves as the actual machining process dataset. After processing, the target workpiece is subjected to quality inspection to obtain actual data on the surface integrity, dimensional accuracy and microstructure of the target workpiece; The actual processing data set and the actual result data are input into the digital twin for post-processing simulation verification, and the prediction results of the digital twin are compared with the actual result data. If the prediction deviation exceeds the preset calibration threshold, the actual result data is used to calibrate the parameters of the corresponding material-property mapping relationship or microstructure evolution calculation logic in the digital twin.

[0014] Secondly, the present invention provides a machine learning-based adaptive machining control system for CNC machine tools, comprising: The mapping relationship acquisition module is used to construct a digital material specimen reflecting the initial microstructure characteristics of the target workpiece material based on the grade and initial heat treatment state of the target workpiece material, and to establish a quantitative mapping relationship library between the microstructure characteristics and macroscopic mechanical / thermal performance parameters as a material-performance mapping relationship set. The digital twin construction module is used to construct a digital twin based on the material-property mapping relationship set and the preset tool geometry model and machine tool dynamics model. The digital twin can predict the microstructure evolution of the workpiece material according to the thermal-mechanical load history of local processing and update the material properties based on the microstructure evolution. The optimal robustness parameter acquisition module is used to optimize the processing parameters of the target processing task based on the digital twin, with processing quality, processing efficiency and microstructure evolution indicators as optimization objectives and the inherent fluctuation of the micro-characteristics of the target workpiece material as uncertainty variables, to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map. The parameter fine-tuning instruction generation module is used to load the optimal robust parameter package during actual processing and simultaneously start the online fine-tuning module. The online fine-tuning module, based on real-time sensor data, infers the equivalent microstate offset of the workpiece processing area through the built-in microstate observer, and generates parameter fine-tuning instructions within the preset robustness envelope by combining the parameter sensitivity map. The digital twin update module is used to execute the optimal robustness parameter package and the parameter fine-tuning instructions generated in real time, complete the processing, and update the digital twin based on the processing data.

[0015] The present invention has at least the following beneficial effects: By implementing this invention, it is possible to construct a digital material specimen reflecting the initial microstructural characteristics of the target workpiece material based on its grade and initial heat treatment state, and to establish a quantitative mapping relationship library between the microstructural characteristics and macroscopic mechanical / thermal performance parameters, serving as a material-performance mapping relationship set. This provides accurate and quantifiable material foundational data support for the subsequent construction of digital twins. The digital material specimen can realistically reflect the initial microscopic state of the material, ensuring the representativeness and accuracy of the data whether it is measured reconstruction or simulated generation. The quantitative mapping relationship library allows subsequent process optimization to accurately correlate the material's microscopic characteristics with its macroscopic processing performance, providing a data foundation for forward-looking optimization.

[0016] By implementing this invention, a digital twin can be constructed based on the material-property mapping relationship set and preset tool geometry and machine tool dynamics models. This digital twin can predict the microstructural evolution of the workpiece material based on the local thermal-mechanical load history during machining and update material properties based on this microstructural evolution. This achieves collaborative simulation and dynamic linkage between macroscopic and microscopic aspects during machining. Compared to traditional models that only focus on macroscopic or static microscopic aspects, this digital twin can capture the dynamic changes in microstructure during machining in real time and simultaneously update the macroscopic material properties. This allows process optimization and machining control to no longer be based on fixed material property assumptions but rather on changes in the material state during actual machining, significantly improving the accuracy of simulation predictions and their alignment with actual machining.

[0017] By implementing this invention, it is possible to optimize processing parameters for a target processing task based on a digital twin, using processing quality, processing efficiency, and microstructure evolution indicators as optimization targets, and the inherent fluctuations in the microscopic characteristics of the target workpiece material as uncertainty variables. This yields an optimal robust parameter package and a corresponding parameter sensitivity map. This solves the problems of traditional optimization schemes being sensitive to material microscopic fluctuations and having poor parameter stability. The optimal robust parameter package itself possesses the ability to resist interference from inherent material microscopic fluctuations, ensuring consistent processing quality even with minor differences between batches of materials. The parameter sensitivity map clearly presents the degree of influence of each processing parameter on the target indicators, providing a clear adjustment basis for subsequent online fine-tuning and avoiding blind adjustments.

[0018] By implementing this invention, the optimal robustness parameter package can be loaded during actual processing, and the online fine-tuning module can be started simultaneously. Based on real-time sensor data, the online fine-tuning module uses a built-in microstate observer to infer the equivalent microstate shift of the workpiece processing area, and, combined with the parameter sensitivity map, generates parameter fine-tuning instructions within a preset robustness envelope. This achieves dual control of basic robustness and real-time adaptive control. The optimal robustness parameter package ensures the basic stability of processing, while the online fine-tuning module dynamically compensates for sudden material microstate shifts during actual processing. This avoids quality risks caused by parameter adjustments exceeding the robustness range and responds promptly to real-time disturbances, compensating for the lag and limitations of traditional passive response control, and further improving the stability and quality consistency of the processing.

[0019] By implementing this invention, the optimal robustness parameter package and the real-time generated parameter fine-tuning instructions can be executed to complete the processing, and the digital twin can be updated based on the processing data, forming a closed-loop optimization mechanism. The actual processing data provides a real basis for the iterative optimization of the digital twin. Continuous calibration continuously improves the simulation prediction accuracy of the digital twin, and subsequent processing tasks can directly reuse the optimized model, achieving continuous iterative upgrades of the technology. Simultaneously, the combined execution of basic and fine-tuning instructions during processing ensures that the final processing quality, efficiency, and microstructure evolution indicators all meet the preset requirements.

[0020] In summary, by implementing this invention, the consistency of processing quality and the stability of processing efficiency of high-end parts can be significantly improved, and problems such as abnormal tool wear and surface quality variation caused by microscopic inhomogeneity of materials can be reduced, thus breaking through the limitations of traditional technologies in improving the consistency of processing quality. Attached Figure Description

[0021] Figure 1 A flowchart illustrating the machine learning-based adaptive machining control method for CNC machine tools provided by this invention; Figure 2 This is a schematic diagram of the structure of the machine learning-based adaptive machining control system for CNC machine tools provided by the present invention.

[0022] In the attached diagram, the components represented by each number are as follows: The module includes: 11 for obtaining mapping relationships, 12 for constructing a digital twin, 13 for obtaining optimal robustness parameters, 14 for generating parameter fine-tuning instructions, and 15 for updating the digital twin. Detailed Implementation

[0023] Example 1, as Figure 1 As shown, embodiments of the present invention provide a machine learning-based adaptive machining control method and system for CNC machine tools, including: S100: Based on the grade and initial heat treatment state of the target workpiece material, construct a digital material specimen that reflects the initial microstructure characteristics of the target workpiece material, and establish a quantitative mapping relationship library between the microstructure characteristics and macroscopic mechanical / thermal performance parameters as a material-performance mapping relationship set; S200: Based on the material-property mapping relationship set, as well as the preset tool geometry model and machine tool dynamics model, a digital twin is constructed. The digital twin can predict the microstructure evolution of the workpiece material according to the thermal-mechanical load history of local processing and update the material properties based on the microstructure evolution. S300: Taking processing quality, processing efficiency and microstructure evolution indicators as optimization targets, and taking the inherent fluctuation of the microstructure characteristics of the target workpiece material as an uncertainty variable, the processing parameters of the target processing task are optimized based on the digital twin to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map. S400: Load the optimal robustness parameter package during actual processing and simultaneously start the online fine-tuning module; the online fine-tuning module, based on real-time sensor data, infers the equivalent microstate offset of the workpiece processing area through the built-in microstate observer, and generates parameter fine-tuning instructions within the preset robustness envelope in combination with the parameter sensitivity map. S500: Execute the optimal robustness parameter package and the parameter fine-tuning instructions generated in real time to complete the processing, and update the digital twin based on the processing data.

[0024] In step S100 of this application embodiment, a digital material specimen reflecting the initial microstructural characteristics of the target workpiece material is constructed based on the grade and initial heat treatment state of the target workpiece material, including: Obtain the grade of the target workpiece material and its corresponding initial heat treatment process; Based on the grade, the material database is queried to obtain the standard chemical composition, phase diagram and basic physical property parameters of the target workpiece material; If measured data of the target workpiece material exists, import the electron backscatter diffraction or metallographic image of the target workpiece material, and digitally reconstruct the real microstructure model containing grains, grain boundaries and second phase through image segmentation and vectorization processing. If no measured data of the target workpiece material is available, then based on the chemical composition, initial heat treatment process and material phase transformation kinetics principle, a virtual microstructure model that is statistically representative is generated by cellular automata or Monte Carlo method. Based on the real microstructure model or the virtual microstructure model, a digital model expressed in the form of a three-dimensional voxel mesh is constructed for the target workpiece material as the digital material specimen, wherein each mesh in the digital material specimen is associated with a corresponding material phase and crystallographic direction.

[0025] In step S100 of this embodiment, the purpose of the above steps is to digitize and model the initial microstructural features of the target workpiece material. This lays the foundation for establishing a quantitative mapping relationship between microstructural features and macroscopic mechanical and thermal properties, as well as for constructing a digital twin. This allows for precise material microscopic level data support for subsequent processing parameter optimization and adaptive control, thereby solving the problem of processing quality fluctuations caused by material microscopic inhomogeneity.

[0026] To achieve the above objectives, it is first necessary to obtain the grade of the target workpiece material and the corresponding initial heat treatment process. First, the grade of the target workpiece material must be clearly identified, such as 45 steel, along with the corresponding initial heat treatment process, such as quenching and tempering. This is a prerequisite for constructing a digital material specimen, enabling the determination of the material's basic property range.

[0027] Next, based on the grade, the material database is queried to obtain the standard chemical composition, phase diagram, and basic physical property parameters of the target workpiece material; Based on the obtained grade, the material database is searched to obtain the standard chemical composition of the material, such as the carbon content of 45 steel being 0.42%-0.50% and the manganese content being 0.50%-0.80%; phase diagrams, such as the iron-carbon phase diagram; and basic physical property parameters, such as density and melting point. These data are the basis for constructing the microstructure model.

[0028] If measured data of the target workpiece material exists, import the electron backscatter diffraction or metallographic image of the target workpiece material, and digitally reconstruct the real microstructure model containing grains, grain boundaries and second phase through image segmentation and vectorization processing. Specifically, the first step is to import the electron backscatter diffraction image or metallographic image of the target workpiece material. The electron backscatter diffraction image can characterize the crystal orientation, while the metallographic image can observe the grain, grain boundary, and second phase morphology. The second step involves using image segmentation algorithms such as threshold segmentation and region growing to distinguish different microscopic components in the image. For example, grain regions, grain boundary regions, and second-phase particles are segmented into independent image regions. Among them, second-phase particles include alloy carbides and nitrides in 304 stainless steel.

[0029] The third step is to perform vectorization processing, which transforms the segmented image regions into digital information such as geometric coordinates and morphological parameters. For example, the outline of the grain boundary is described by the sequence of coordinate points, and the size of the second phase particles is characterized by diameter and area parameters. The fourth step is to digitally reconstruct a real microstructure model containing grains, grain boundaries, and a second phase based on the above information. This model can completely replicate the microstructure of the measured material. For example, if the grain size of a measured 304 stainless steel is between 20-50 μm, the reconstruction model will accurately restore this size distribution and the spatial arrangement of the grains.

[0030] If no measured data of the target workpiece material is available, then based on the chemical composition, initial heat treatment process and material phase transformation kinetics principle, a virtual microstructure model that is statistically representative is generated by cellular automata or Monte Carlo method. Specifically, the first step is to determine the key laws governing the evolution of the microstructure based on the previously obtained chemical composition and initial heat treatment process, combined with the principles of material phase transformation kinetics, such as the phase transformation rate equation and grain growth kinetic model. For example, it is found that the grain growth rate of TC4 titanium alloy during annealing is proportional to the square root of the holding time. The second step is to select cellular automata or Monte Carlo method as the simulation tool. For example, when using cellular automata method, the simulation area is divided into a large number of tiny cells, each cell representing a tiny volume unit of the material. The third step is to set simulation rules. For example, the state of a cell changes based on the state of adjacent cells, the distribution of chemical composition, and heat treatment process parameters. The cell state can represent austenite, ferrite, or grain orientation, and the heat treatment process parameters include holding temperature and time. The Monte Carlo method simulates the nucleation and growth process of grains through random sampling. Using cellular automata or the Monte Carlo method as simulation tools is existing technology and will not be elaborated further here.

[0031] The fourth step is to generate a statistically representative virtual microstructure model through simulation. For example, for a TC4 titanium alloy with unknown measured data, the virtual microstructure model generated by simulation has a grain size that conforms to the statistical distribution of TC4 titanium alloy under this type of annealing process, such as a mean of 30 μm and a standard deviation of 5 μm. The number density and morphology of the second phase particles are consistent with the thermodynamic stability state corresponding to the chemical composition.

[0032] Finally, based on the real microstructure model or the virtual microstructure model, a digital model expressed in the form of a three-dimensional voxel mesh is constructed for the target workpiece material as the digital material specimen, wherein each mesh in the digital material specimen is associated with a corresponding material phase and crystallographic direction.

[0033] The previously obtained real or virtual microstructure models are transformed into three-dimensional digital models in a unified format, ensuring that each micro-region has clear attribute relationships, which facilitates subsequent simulation calculations.

[0034] The specific implementation method is as follows: The first step is to use a real or virtual microstructure model as a blueprint and discretize the model using a three-dimensional voxel mesh. This means dividing the entire microstructure model into a large number of tiny three-dimensional cubic meshes, which are the voxels. The side length of each voxel can be set according to the modeling accuracy requirements, such as a voxel size of 1μm×1μm×1μm, to ensure accurate representation of microscopic details. The second step is to associate two key attributes with each voxel grid: one is the material phase, such as a voxel being associated with austenite, ferrite, or a second phase, to clarify the material composition of the region; the other is the crystallographic direction, such as

[100] ,

[110] ,

[111] , etc., which characterize the arrangement of crystals in the region. The third step is to integrate all the voxel meshes with related attributes to form a complete three-dimensional digital model, i.e., a digital material specimen. For example, a digital material specimen of 304 stainless steel contains millions of voxel meshes, each of which is clearly labeled as either austenitic phase or crystallographically

[100] or

[110] . Subsequently, the macroscopic mechanical / thermal properties of the material can be calculated by calling the attribute data of these meshes.

[0035] In step S100 of this application embodiment, establishing a quantitative mapping relationship library between the microstructural features and macroscopic mechanical / thermal performance parameters, as a material-performance mapping relationship set, includes: Using the digital material sample as input, the macroscopic mechanical / thermal properties under different combinations of microstructure features are predicted by performing crystal plasticity finite element simulation calculations. The macroscopic mechanical property parameters include at least yield strength, flow stress and strain hardening index, and the macroscopic thermal property parameters include at least thermal conductivity and specific heat capacity. By using microstructural features as independent variables and the macroscopic mechanical / thermal properties obtained from simulation as dependent variables, a structured and queryable mapping relationship database is constructed as the material-property mapping relationship set.

[0036] In step S100 of this application embodiment, the purpose of the above steps is to establish a precise correlation bridge between microstructure and macroscopic properties. Digital material samples only achieve digital characterization of the microstructure, while subsequent construction of digital twins and optimization of processing parameters rely on the material's macroscopic performance data. Through this quantitative mapping, abstract microscopic features such as grain size and crystallographic orientation can be transformed into calculable and applicable macroscopic performance parameters, such as yield strength and thermal conductivity, providing a core basis for performance calculations of the digital twin. Simultaneously, it allows subsequent optimization of processing parameters to accurately correlate the impact of microscopic material fluctuations on macroscopic processing results, ensuring that the optimization process not only conforms to the material's inherent properties but also specifically avoids process sensitivity issues caused by microscopic inhomogeneities.

[0037] To achieve the above objectives, the digital material sample is first used as input. By performing crystal plasticity finite element simulation calculations, the macroscopic mechanical / thermal properties under different combinations of microstructure characteristics are predicted. Among them, the macroscopic mechanical property parameters include at least yield strength, flow stress and strain hardening index, and the macroscopic thermal property parameters include at least thermal conductivity and specific heat capacity. Based on a digital microstructure model, and with the help of professional simulation methods, the macroscopic performance of materials under stress and heat is simulated, so as to realize the performance derivation from micro to macro and ensure that the performance data is directly linked to the microscopic characteristics.

[0038] The specific implementation method is as follows: The first step is to use the previously constructed digital material specimen as the core input and import it into crystal plasticity finite element simulation software, such as the crystal plasticity module of Abaqus or LS-DYNA. The second step is to set the simulation boundary conditions and calculation parameters, such as the stress loading method, temperature range, and simulation iteration step size during the cutting process. The third step is to perform simulation calculations and use software to solve the macroscopic response of the material under different combinations of microstructural features. For example, when the grain size in the digital material sample increases from 20 μm to 50 μm and the crystallographic orientation is dominated by

[110] , the corresponding macroscopic performance changes are calculated. The fourth step is to extract key simulation results, namely macroscopic mechanical and macroscopic thermal performance parameters. The macroscopic mechanical performance parameters include at least: yield strength (e.g., the yield strength of 304 stainless steel at a grain size of 30μm is approximately 205MPa); flow stress (e.g., the flow stress of TC4 titanium alloy at a strain of 0.2 is approximately 800MPa); and strain hardening index (e.g., the strain hardening index of 45 steel is approximately 0.2). The macroscopic thermal performance parameters include at least: thermal conductivity (e.g., the thermal conductivity of aluminum alloy is approximately 237W / (m·K); and specific heat capacity (e.g., the specific heat capacity of copper is approximately 385J / (kg·K)).

[0039] Next, the microstructural characteristics are used as independent variables, and the macroscopic mechanical / thermal properties obtained from the simulation are used as dependent variables to construct a structured and queryable mapping relationship database, which serves as the material-property mapping relationship set.

[0040] The corresponding data of microstructural features and macroscopic performance obtained from the simulation will be organized into a standardized and searchable database to facilitate rapid retrieval when calling digital twins and optimizing processing parameters.

[0041] The specific implementation method is as follows: The first step is to clarify the data relationships by setting microstructural features as independent variables. These microstructural features include grain size, grain orientation distribution, grain boundary density, second-phase particle size and number density, etc. For example, "grain size 20μm, second-phase particle diameter 1μm and number density 100 particles / m³" is a combination of independent variables. The macroscopic mechanical properties obtained from the simulation are set as dependent variables, such as yield strength 220MPa and thermal conductivity 200W / (m·K) corresponding to the above combination of independent variables. The second step is to design the database structure and store the data in a structured format, such as a table. Each row corresponds to a set of "microscopic feature combination - macroscopic performance" data, and the columns are labeled with key fields such as grain size, second phase number density, yield strength, and thermal conductivity. The third step is to complete the database construction by entering all the corresponding sets of data obtained from the simulation into the database, forming a set of material-property mapping relationships that can be quickly queried based on microstructural features. For example, by inputting the query conditions of grain size 30μm and crystallographic orientation

[100] as the main direction, the corresponding flow stress, specific heat capacity and other parameters can be directly obtained.

[0042] In step S200 of this application embodiment, a digital twin is constructed based on the material-property mapping relationship set and the preset tool geometry model and machine tool dynamics model, including: Obtain the preset tool geometry model and machine tool dynamics model; Based on the aforementioned material-property mapping relationship set, the preset tool geometry model, and the machine tool dynamics model, a digital twin of the macroscopic thermo-mechanical coupling machining process is established; Within each macroscopic material property calculation unit of the digital twin, a microstructure evolution calculation unit is nested. The microstructure evolution calculation unit calculates the microscopic property data within the macroscopic material property calculation unit by calling the embedded crystal plasticity finite element model based on the macroscopic property data provided by the macroscopic material property calculation unit to which it is located. The macroscopic performance data includes local strain, strain rate, and temperature history data; the microscopic performance data includes the evolution of grain size, dislocation density, and phase fraction. A data transmission channel is established from the macroscopic material property calculation unit to the microstructure evolution calculation unit for real-time transmission of the macroscopic property data; Simultaneously, a data feedback channel is established to return from the microstructure evolution calculation unit to the macromaterial property calculation unit, which is used to update the material properties in the macromaterial property calculation unit based on the calculated microstructure property data.

[0043] In this embodiment, the purpose of step S200 is to create a digital twin that is fully synchronized with the actual processing and can interact with it in real time. This digital twin must accurately simulate the thermo-mechanical coupling processing at the macroscopic scale and capture the dynamic evolution of the material's microstructure in real time, achieving bidirectional linkage between the macroscopic processing state and the microstructure evolution. Through this linkage, the digital twin can predict in advance the changes in material properties caused by microstructure changes during processing, providing a realistic simulation environment for subsequent processing parameter optimization. Simultaneously, it provides real-time predictive data support for online fine-tuning during actual processing, solving the problem that traditional models cannot simultaneously consider the dynamic correlation between macroscopic processing and microscopic evolution, ensuring the accuracy of the processing simulation and the foresight of adaptive control.

[0044] To achieve the above objectives, it is first necessary to obtain the preset tool geometry model and machine tool dynamics model; For example, the tool geometry model needs to accurately characterize the key structural parameters of the tool, such as the end mill's diameter of 10mm, helix angle of 30°, cutting edge rake angle of 5°, and cutting edge radius of 0.02mm. These parameters directly affect the force distribution and heat conduction during the cutting process. The machine tool dynamics model needs to include the machine tool's key dynamic characteristic parameters, such as the machine tool spindle's natural frequency of 200Hz, damping ratio of 0.05, and feed system stiffness of 5×10⁻⁶. 7 N / m is used to simulate the dynamic behavior of machine tools during machining, such as vibration and motion response. The methods for constructing the tool geometry model and machine tool dynamics model are existing technologies and can be directly obtained from tool manufacturers and machine tool manufacturers, and will not be elaborated here.

[0045] Next, based on the material-property mapping relationship set, the preset tool geometry model, and the machine tool dynamics model, a digital twin of the macroscopic thermo-mechanical coupling machining process is established; This involves integrating material-property mapping relationships, tool geometry models, and machine tool dynamics models to build a macroscopic simulation framework that can simulate the interaction of force and heat in actual machining.

[0046] For example, based on a material-property mapping set, macroscopic mechanical and thermal property parameters, such as yield strength of 205 MPa and thermal conductivity of 237 W / (m·K), are used as material properties input into the system. A tool geometry model is imported to define the contact method between the cutting edge and the workpiece, as well as the geometric boundaries of the cutting area. A machine tool dynamics model is integrated to simulate the dynamic changes in parameters such as cutting speed and feed rate caused by machine tool movement during machining. The cutting speed can be referenced to a spindle speed of 3000 r / min, and the feed rate can be referenced to a feed per tooth of 0.02 mm / z. A thermo-mechanical coupling simulation algorithm, such as the finite element method, is used to establish a macroscopic digital twin. This digital twin can simulate the distribution and changes in cutting force and cutting temperature during the cutting process, as well as the macroscopic machining state such as the macroscopic deformation of the workpiece. Specifically, the cutting force can be referenced to a main cutting force of 500 N, the cutting temperature can be referenced to a cutting area temperature of 600 °C, and the macroscopic deformation can be referenced to a maximum deformation of 0.01 mm.

[0047] Then, within each macroscopic material property calculation unit of the digital twin, a microstructure evolution calculation unit is nested. The microstructure evolution calculation unit calculates the microscopic property data within the macroscopic material property calculation unit by calling the embedded crystal plasticity finite element model based on the macroscopic property data provided by the macroscopic material property calculation unit to which it is located. In other words, within each macroscopic material property calculation unit of the macroscopic digital twin, such as a 1mm × 1mm × 1mm calculation region, a separate microstructure evolution calculation unit is nested. This microscopic calculation unit incorporates a crystal plasticity finite element model and can receive macroscopic property data transmitted from the macroscopic calculation unit. This macroscopic property data includes: local strain, such as a strain of 0.3 in the cutting region; strain rate, such as a strain rate of 10... 3 s -1 The system also includes historical temperature data, such as temperature change curves from room temperature to 700℃ and then down to 300℃. Based on this macroscopic data, the microscopic computing unit performs simulation calculations using a crystal plasticity finite element model, outputting microscopic performance data.

[0048] These microscopic performance data include: the evolution of grain size, such as grain refinement from 30 μm to 15 μm after processing; and the evolution of dislocation density, such as dislocation density from 10 μm to 15 μm after processing. 12 m -2 Increased to 10 14 m -2 ; and the evolution of phase fractions, for example, the austenite phase fraction decreases from 90% to 70%, and the ferrite phase fraction increases from 10% to 30%.

[0049] Furthermore, a data transmission channel is established from the macroscopic material property calculation unit to the microstructure evolution calculation unit for real-time transmission of the macroscopic property data; Simultaneously, a data feedback channel is established to return from the microstructure evolution calculation unit to the macromaterial property calculation unit, which is used to update the material properties in the macromaterial property calculation unit based on the calculated microstructure property data.

[0050] This involves constructing a two-way data transmission and feedback mechanism from macro to micro and from micro to macro to ensure that macro and micro computing units can interact with each other in real time, so that the simulation of the twin remains consistent with the dynamic process of actual processing.

[0051] The specific implementation method is as follows: First, a data transfer channel is established from the macroscopic material property calculation unit to the microstructure evolution calculation unit. This channel can synchronously transmit data such as local strain, strain rate, and temperature history calculated in real time by the macroscopic calculation unit to the corresponding microscopic calculation unit, providing real-time operating condition input for microstructure evolution calculation. Simultaneously, a data feedback channel is established from the microstructure evolution calculation unit back to the macroscopic material property calculation unit. When the microscopic calculation unit calculates changes in microscopic properties such as grain size refinement and increased dislocation density, this feedback channel is used to feed these changes back to the macroscopic calculation unit. Based on the feedback microscopic property data, the macroscopic calculation unit updates its own material property parameters. For example, grain refinement may increase the yield strength from 205 MPa to 250 MPa, ensuring that the macroscopic simulation can promptly reflect the changes in material properties caused by microstructure changes, achieving dynamic linkage between macro and micro scales.

[0052] In step S300 of this embodiment, processing quality, processing efficiency, and microstructure evolution indicators are used as optimization targets, and the inherent fluctuations of the microscopic characteristics of the target workpiece material are used as uncertainty variables. Based on the digital twin, the processing parameters of the target processing task are optimized to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map, including: Define an optimization multi-objective function, with objective terms including at least processing quality, processing efficiency, and microstructure evolution indicators; Among them, the processing quality target is characterized by the surface roughness and residual stress predicted by the digital twin, the processing efficiency target is characterized by the material removal rate, and the microstructure evolution index is characterized by the degree of subsurface grain refinement and the phase change layer depth. The statistical distribution of key microscopic features extracted from the digital material specimens is set as an interval-type uncertainty variable in the optimization process. Based on the digital twin, a multi-objective robust optimization algorithm is used to perform global optimization within the processing parameter space; During the optimization process, for each set of candidate processing parameters, multiple sampling simulations are performed within the range of the uncertainty variables to calculate the mean and variance of the performance of each objective item. The optimal robustness parameter package is the combination of processing parameters with the best overall performance and the smallest variance in the Pareto front. The parameter sensitivity map calculates the gradient of the influence of each processing parameter on each sub-indicator in the multi-objective function and presents it in the form of a visual data matrix.

[0053] In step S300 of this embodiment, the purpose of the above steps is to find the optimal robust parameter combination that balances processing quality, processing efficiency, and microstructure evolution, under the premise of inherent fluctuations in the microscopic characteristics of the material, and to clarify the degree of influence of each processing parameter on the target index. By treating the fluctuations in the microscopic characteristics of the material as an uncertainty variable, the problem of traditional optimization parameters failing when there are batch variations in the material is avoided, allowing the optimized optimal robust parameter package to maintain stable processing results even when facing microscopic inhomogeneities. At the same time, the parameter sensitivity map provides a clear basis for online fine-tuning in subsequent actual processing, ensuring that the processing can accurately respond to microscopic state deviations, ultimately improving the consistency and stability of the processing quality of high-end parts.

[0054] To achieve the above objectives, it is first necessary to define an optimization multi-objective function, with the objective terms including at least processing quality, processing efficiency, and microstructure evolution indicators. Among them, the processing quality target is characterized by the surface roughness and residual stress predicted by the digital twin, the processing efficiency target is characterized by the material removal rate, and the microstructure evolution index is characterized by the degree of subsurface grain refinement and the phase change layer depth. This involves identifying the core objective dimension for optimization, transforming abstract processing requirements into quantifiable and calculable specific indicators, and providing judgment criteria for subsequent optimization.

[0055] Specifically, a multi-objective function can be constructed containing three core objective terms, each corresponding to a specific quantitative indicator. The processing quality objective is characterized by at least the surface roughness and residual stress predicted by the digital twin; for example, the surface roughness objective is set to not exceed Ra 0.8 μm, and the residual stress objective is set to not exceed 300 MPa. The processing efficiency objective is characterized by the material removal rate; for example, the material removal rate objective is set to not less than 500 mm. 3 / min; the microstructure evolution index is characterized by at least the degree of subsurface grain refinement and the depth of the phase change layer. For example, the target for the degree of subsurface grain refinement is set to a grain size of not less than 10μm after processing, and the target for the depth of the phase change layer is set to not more than 50μm.

[0056] Then, based on the digital material specimens, the statistical distribution of key microscopic features is extracted and set as an interval-type uncertainty variable in the optimization process; That is, relying on the simulation capabilities of digital twins, professional optimization algorithms are used to search within a broad processing parameter space to ensure that the found parameter combinations not only meet the requirements of multiple objectives, but also have robustness against micro fluctuations.

[0057] Specifically, suitable multi-objective robust optimization algorithms can be selected, such as the non-dominated sorting genetic algorithm NSGA-III and multi-objective particle swarm optimization. The range of the machining parameter space is determined; for example, machining parameters include cutting speed, feed rate, and depth of cut. For instance, the cutting speed range is set to 80 m / min-200 m / min, the feed rate range to 0.01 mm / z-0.05 mm / z, and the depth of cut range to 0.5 mm-2 mm. Based on a digital twin, global optimization is performed within this machining parameter space, and each set of candidate machining parameters needs to be verified through multiple rounds of simulation.

[0058] Based on the digital twin, a multi-objective robust optimization algorithm is then used to perform global optimization within the processing parameter space; That is, for each set of candidate parameters, multiple sampling simulations are performed within the range of uncertainty variables, and the comprehensive performance and robustness of the parameters are evaluated by the mean and variance, so as to avoid optimization bias under a single working condition.

[0059] Specifically, for each set of candidate machining parameters, such as cutting speed of 150 m / min, feed rate of 0.03 mm / z, and depth of cut of 1 mm, multiple samplings can be performed within the range of interval-type uncertainty variables. For example, for grain size of 20 μm-50 μm and second-phase particle number density of 80 particles / mm². 3 -120 pieces / mm 3 Fifty different combinations of microscopic features were randomly generated. Simulations were performed on each sampled condition using a digital twin, and the mean and variance of each target performance item were calculated. For example, the mean surface roughness Ra was 0.6 μm and the variance was 0.05 μm. 2 The average material removal rate was 550 mm. 3 / min, variance 30mm 6 / min 2 The average subsurface grain refinement is 15 μm, and the variance is 2 μm. 2 .

[0060] Furthermore, during the optimization process, for each set of candidate processing parameters, multiple sampling simulations are performed within the range of the uncertainty variables to calculate the mean and variance of the performance of each objective item. The optimal robustness parameter package is the combination of processing parameters with the best overall performance and the smallest variance in the Pareto front. In other words, in the Pareto front of multi-objective optimization, the combination of processing parameters with the best overall performance and the smallest variance is selected to ensure that the parameters not only meet the requirements of various indicators, but also have the strongest resistance to fluctuations.

[0061] Specifically, through multiple rounds of optimization and sampling simulation, multiple sets of non-dominated solutions are obtained, forming a Pareto front. Within the Pareto front, the mean and variance of each machining parameter combination are comprehensively evaluated, and the machining parameter combination with the best overall performance and smallest variance is selected as the optimal robust parameter package. For example, the finally determined optimal robust parameter package is a cutting speed of 160 m / min, a feed rate of 0.025 mm / z, and a depth of cut of 1.2 mm. Under this combination, the mean surface roughness Ra is 0.5 μm, and the variance is 0.03 μm. 2 The average material removal rate was 580 mm. 3 / min, variance 25mm 6 / min 2 All indicators met the requirements and showed minimal fluctuations.

[0062] The parameter sensitivity map is calculated by the gradient of the influence of each processing parameter on each sub-indicator in the multi-objective function, and presented in the form of a visual data matrix.

[0063] This involves calculating the gradient of influence of each machining parameter on various sub-indices of machining quality, machining efficiency, and microstructure evolution through methods such as partial derivative calculations and the controlled variable method. For example, the gradient of influence of cutting speed on surface roughness is calculated to be -0.002 μm / (m / min), meaning that for every 1 m / min increase in cutting speed, the surface roughness decreases by an average of 0.002 μm; the gradient of influence of feed rate on material removal rate is 10000 mm. 3 / (min・mm / z), meaning that for every 0.01 mm / z increase in feed rate, the material removal rate increases by an average of 100 mm. 3 / min. These influence gradients are presented in the form of a visual data matrix to form a parameter sensitivity map. For example, rows represent processing parameters, columns represent target sub-indicators, and matrix elements represent the corresponding influence gradient values.

[0064] In step S400 of this application embodiment, the optimal robustness parameter package needs to be loaded during actual processing, and the online fine-tuning module is started simultaneously. The online fine-tuning module infers the equivalent microstate offset of the workpiece processing area in reverse through the built-in microstate observer based on real-time sensor data, and generates parameter fine-tuning instructions within the preset robustness envelope in combination with the parameter sensitivity map. In this embodiment, the purpose of step S400 is to ensure that, during actual processing, based on the optimal robustness parameter package, the microscopic state deviation of the workpiece processing area is dynamically addressed through online fine-tuning, thereby ensuring that processing quality, efficiency, and microstructure evolution remain within an ideal range. On one hand, the optimal robustness parameter package needs to be loaded to ensure the basic stability of the processing; on the other hand, an online fine-tuning module is needed to accurately capture material microscopic fluctuations based on real-time sensor data and a microscopic state observer. Combined with a parameter sensitivity map and a preset robustness envelope, safe and effective parameter fine-tuning instructions are generated to avoid a decline in processing quality due to microscopic state deviations. This achieves a dual guarantee of basic robustness and dynamic adaptation, further improving the consistency of processing quality.

[0065] To achieve the above objectives, it is first necessary to load the optimal robustness parameter package and start the online fine-tuning module. That is, the optimal robustness parameter package obtained from the initial optimization is used as the basic instructions for actual processing, and the online fine-tuning module is started simultaneously to prepare for subsequent dynamic adjustments.

[0066] For example, an optimal robustness parameter package, such as a cutting speed of 160 m / min, a feed rate of 0.025 mm / z, and a depth of cut of 1.2 mm, is loaded into the machine tool machining control system as the initial machining parameters for executing the machining task. Simultaneously, an online fine-tuning module is activated. This module is associated with the machine tool's sensor network, microstate observer, parameter sensitivity map, and robustness envelope data, and is ready to receive real-time data and adjust parameters as needed.

[0067] Next, a microstate observer needs to be constructed.

[0068] In step S400 of this application embodiment, the construction of the microstate observer includes: Based on the digital twin, a dataset containing various simulated microscopic state fluctuations and corresponding sensor response characteristics is generated; A lightweight machine learning model is trained using the dataset. The input of the machine learning model includes time-frequency domain features of real-time cutting force, vibration, and acoustic emission signals. The output includes estimates of the equivalent hardness and equivalent toughness state of the current machining area of ​​the target workpiece. The trained machine learning model is deployed as the microstate observer.

[0069] This involves generating a simulation dataset using a digital twin, training a lightweight machine learning model, and enabling the machine learning model to infer the microscopic state shift of the material based on real-time sensor signals, thus providing a basis for parameter fine-tuning.

[0070] The specific implementation method is as follows: Based on digital twins, various microscopic state fluctuation scenarios are simulated, such as grain size fluctuations between 20μm and 50μm and second-phase particle number density of 80 particles / mm. 3 -120 pieces / mm 3 The data changes between these parameters, and the corresponding sensor response characteristics for each scenario are recorded to form a dataset containing a large number of samples.

[0071] Choose a lightweight machine learning model, such as a support vector machine, random forest, or lightweight neural network, and train it using the aforementioned dataset. The model's input consists of time-frequency domain features of real-time sensor data, including cutting force, vibration, and acoustic emission signals. Time-frequency domain features include the peak value, RMS value, and peak frequency of the cutting force; the root mean square value and kurtosis of the vibration signal; and the energy count of the acoustic emission signal. The model's output is an estimate of the equivalent hardness and equivalent toughness of the target workpiece's current machining area, for example, an equivalent hardness estimate of 250HB-350HB and an equivalent toughness estimate of 50J / cm². 2 -80J / cm 2 .

[0072] The specific method for training the machine learning model based on the dataset is existing technology and will not be described in detail here.

[0073] The trained machine learning model that meets the accuracy requirements is deployed to the online fine-tuning module as a microstate observer to infer the microstate shift in real time.

[0074] Next, the robustness envelope needs to be defined.

[0075] In step S400 of this application embodiment, the preset robustness envelope is defined in the following way: Based on the aforementioned optimal robustness parameter package, parameter perturbation analysis is performed in the digital twin; To obtain the maximum positive and negative adjustment range of each key processing parameter while keeping all processing quality indicators and microstructure evolution indicators in line with preset standards; The maximum positive and maximum negative adjustment amplitudes together form a super-rectangular region in a multi-dimensional parameter space, and the super-rectangular region is used as the preset robustness envelope.

[0076] That is, by analyzing parameter disturbances, the maximum adjustment range of each key processing parameter is determined under the premise of meeting the index requirements, forming a constraint range for parameter adjustment, and avoiding processing failure due to excessive parameter adjustment.

[0077] The specific implementation method is as follows: Based on the optimal robust parameter package, parameter perturbation analysis is performed in the digital twin. For each key machining parameter, such as cutting speed, feed rate, and depth of cut, the machining parameter values ​​are progressively adjusted in both positive and negative directions.

[0078] After each adjustment, the processing quality indicators and microstructure evolution indicators are verified by digital twin simulation to see if they meet the preset standards, such as surface roughness not exceeding Ra0.8μm, residual stress not exceeding 300MPa, and phase change layer depth not exceeding 50μm.

[0079] Record the maximum allowable positive and negative adjustment ranges for each key machining parameter. For example, the optimal cutting speed is 160 m / min, with a maximum positive adjustment range of 30 m / min and a maximum negative adjustment range of 40 m / min; the optimal feed rate is 0.025 mm / z, with a maximum positive adjustment range of 0.01 mm / z and a maximum negative adjustment range of 0.008 mm / z; and the optimal depth of cut is 1.2 mm, with a maximum positive adjustment range of 0.3 mm and a maximum negative adjustment range of 0.4 mm.

[0080] The maximum positive and maximum negative adjustment ranges of all key machining parameters together constitute a super-rectangular region in the multi-dimensional parameter space. This super-rectangular region is the preset robustness envelope, such as the cutting speed adjustment range of 120m / min-190m / min, the feed rate adjustment range of 0.017mm / z-0.035mm / z, and the depth of cut adjustment range of 0.8mm-1.5mm.

[0081] In step S400 of this embodiment, parameter fine-tuning instructions are generated within a preset robustness envelope, based on the parameter sensitivity map, including: The parameter sensitivity map records the gradient of the influence of each processing parameter on the processing quality index. When the microstate observer infers that the equivalent microstate has shifted, it queries the parameter sensitivity map based on the direction and magnitude of the shift to identify the adjustment parameters and their adjustment directions that can most effectively compensate for the impact of the shift on key quality indicators. Based on the adjustment parameters and their adjustment direction, and in accordance with predefined compensation rules, parameter fine-tuning instructions are generated under the constraints of the robustness envelope.

[0082] That is, by combining the micro-state offset with the parameter sensitivity map, under the robustness envelope constraint, precise parameter fine-tuning instructions are generated according to the compensation rules to achieve real-time compensation for micro-fluctuations.

[0083] The specific implementation method is as follows: During actual machining, the microstate observer continuously receives real-time sensor data, such as a peak cutting force of 2000 N, a root mean square value of 0.5 g for vibration signals, and an acoustic emission signal energy count of 1000. The equivalent microstate shift is then inferred through model calculations. For example, it can be inferred that the equivalent hardness of the current machining area is 30 HB higher than the baseline state, and the equivalent toughness is 10 J / cm² lower. 2 This means that the microscopic state shifts towards a direction with higher hardness and lower toughness.

[0084] Then, the parameter sensitivity map is queried, which records the gradient of the influence of each machining parameter on the machining quality indicators. For example, the query reveals that the gradient of the influence of cutting speed on surface roughness is -0.002 μm / (m / min), the gradient of the influence of feed rate on residual stress is 5000 MPa / (mm / z), and the gradient of the influence of depth of cut on phase transformation layer depth is 200 μm / mm. Based on the direction and magnitude of the micro-state offset, the adjustment parameters and their adjustment directions that can most effectively compensate for the influence of this micro-state offset on key quality indicators are identified.

[0085] For example, regarding the problem that the surface roughness may not be reduced sufficiently due to the increase in equivalent hardness, the adjustment of the cutting speed has the most significant impact on the surface roughness, and increasing the cutting speed can further reduce the surface roughness. Therefore, the adjustment parameter is determined to be the cutting speed, and the adjustment direction is positive.

[0086] Based on predefined compensation rules, such as a linear relationship between microstate offset and parameter adjustment, and prioritizing the adjustment of parameters with the largest gradient impact on the target index, the specific adjustment amount is calculated under a preset robustness envelope constraint. For example, according to the compensation rules, for every 10 HB increase in equivalent hardness, the cutting speed should be adjusted by 5 m / min. With a current offset of 30 HB, the speed should be adjusted by 15 m / min. The original cutting speed was 160 m / min, and the adjusted speed is 175 m / min. This value falls within the robustness envelope range of 120 m / min-190 m / min, meeting the constraint requirements.

[0087] Based on the above analysis, parameter fine-tuning instructions are generated, such as adjusting the cutting speed from 160m / min to 175m / min, to ensure that the adjusted parameters can effectively compensate for the adverse effects of micro-state deviation and maintain stable machining quality.

[0088] In step S500 of this embodiment, the optimal robustness parameter package and the real-time generated parameter fine-tuning instructions are executed to complete the processing, and the digital twin is updated based on the processing data, including: In the actual processing, the processing control system loads and executes the optimal robustness parameter package as the basic processing instructions, and at the same time receives and integrates the parameter fine-tuning instructions generated in real time by the online fine-tuning module to form the final execution instruction sequence and perform the processing. By deploying a sensor network on the machine tool and the target workpiece, the actual cutting force, vibration, acoustic emission and temperature data are collected synchronously throughout the entire machining cycle, which serves as the actual machining process dataset. After processing, the target workpiece is subjected to quality inspection to obtain actual data on the surface integrity, dimensional accuracy and microstructure of the target workpiece; The actual processing data set and the actual result data are input into the digital twin for post-processing simulation verification, and the prediction results of the digital twin are compared with the actual result data. If the prediction deviation exceeds the preset calibration threshold, the actual result data is used to calibrate the parameters of the corresponding material-property mapping relationship or microstructure evolution calculation logic in the digital twin.

[0089] In this embodiment, the purpose of step S500 is to achieve closed-loop iterative optimization of the digital twin while completing the actual processing task. On the one hand, by integrating the optimal robustness parameter package with real-time fine-tuning instructions, it ensures that the actual processing can accurately respond to microscopic fluctuations in materials, guaranteeing processing quality and efficiency. On the other hand, by using the collected actual processing data and quality inspection results, the predictive accuracy of the digital twin is verified, and parts with excessive deviations are calibrated, continuously improving the simulation accuracy and reliability of the digital twin, providing more accurate model support for subsequent similar processing tasks, and forming a virtuous cycle.

[0090] To achieve the above objectives, the optimal robustness parameter package must first be loaded and executed by the machining control system as the basic machining instructions during the actual machining process. At the same time, the system receives and integrates the parameter fine-tuning instructions generated in real time by the online fine-tuning module to form the final execution instruction sequence and execute the machining. Based on the optimal robustness parameter package, it integrates real-time generated parameter fine-tuning instructions to form the final execution instructions that balance stability and flexibility, ensuring that the processing can dynamically respond to changes in microscopic state.

[0091] For example, the machining control system first loads an optimal robust parameter package as the basic machining instructions, such as a cutting speed of 160 m / min, a feed rate of 0.025 mm / z, and a depth of cut of 1.2 mm. During machining, it receives parameter fine-tuning instructions from the online fine-tuning module in real time, such as an instruction to adjust the cutting speed to 175 m / min due to microscopic state deviations. The control system integrates the basic instructions and the fine-tuning instructions to form a continuous sequence of final execution instructions, such as a smooth transition of the cutting speed from 160 m / min to 175 m / min while other parameters remain unchanged, driving the machine tool to complete the entire machining process.

[0092] Next, a sensor network deployed on the machine tool and the target workpiece is needed to synchronously collect the actual cutting force, vibration, acoustic emission and temperature data throughout the entire machining cycle, as a dataset of the actual machining process. This involves comprehensively collecting dynamic data during the manufacturing process through a sensor network, providing real process input data for subsequent digital twin verification, and ensuring the objectivity of the verification.

[0093] For example, a sensor network, including cutting force sensors, vibration sensors, acoustic emission sensors, and temperature sensors, can be deployed at key locations such as the machine tool spindle, tool post, and worktable, as well as near the machining area of ​​the target workpiece. Throughout the machining cycle, various sensor data are collected synchronously. For instance, cutting force data: peak value 2100N, average value 1800N; vibration data: root mean square value 0.45g, kurtosis 4.2; acoustic emission data: energy count 1200, peak frequency 500kHz; temperature data: highest temperature in the cutting area 680℃, average temperature 550℃. This data is then organized and archived according to timestamps to form a complete dataset of the actual machining process.

[0094] After processing, the target workpiece is subjected to quality inspection to obtain actual data on the surface integrity, dimensional accuracy and microstructure of the target workpiece; That is, after processing is completed, the actual results of the workpiece's quality and microstructure are obtained through professional testing methods, which serve as the core basis for verifying the prediction accuracy of the digital twin.

[0095] For example, multi-dimensional inspections are performed on the processed target workpiece. Surface integrity is inspected using a roughness meter and a residual stress tester, for example, measuring a surface roughness Ra of 0.48 μm and a residual stress of 280 MPa; dimensional accuracy is inspected using a coordinate measuring machine, for example, measuring a critical dimension error of ±0.005 mm; microstructure is inspected using a metallographic microscope and an electron backscatter diffraction device, for example, measuring a subsurface grain size of 14 μm and a phase transformation layer depth of 45 μm. These inspection results are then integrated to form the actual result data.

[0096] Furthermore, the actual processing data set and the actual result data are input into the digital twin for post-processing simulation verification, and the prediction results of the digital twin are compared with the actual result data. The actual processing data is input into the digital twin for simulation. The simulation prediction results are compared with the actual detection results to quantify the prediction deviation of the digital twin.

[0097] Implementation details: The previously collected dataset of the actual machining process, including time-series data such as cutting force, vibration, and temperature, is input into the digital twin. A simulation replicating the actual machining conditions is performed to obtain the digital twin's predicted machining quality and microstructure results. For example, the predicted surface roughness Ra is 0.55 μm, residual stress is 310 MPa, subsurface grain size is 16 μm, and phase transformation layer depth is 48 μm. The predicted results are compared with the actual results one by one, and the deviation values ​​are calculated. For example, the surface roughness deviation is 0.07 μm, the residual stress deviation is 30 MPa, the subsurface grain size deviation is 2 μm, and the phase transformation layer depth deviation is 3 μm.

[0098] If the prediction deviation exceeds the preset calibration threshold, the actual result data is used to calibrate the parameters of the corresponding material-property mapping relationship or microstructure evolution calculation logic in the digital twin.

[0099] This involves setting calibration thresholds and adjusting the core model parameters of the digital twin for items with excessive deviations, thereby achieving iterative optimization of the model.

[0100] First, preset calibration thresholds, such as surface roughness deviation threshold ±0.05μm, residual stress deviation threshold ±25MPa, and microstructure parameter deviation threshold ±1.5μm.

[0101] Comparison revealed that the residual stress deviation of 30 MPa exceeded the calibration threshold of 25 MPa, and the subsurface grain size deviation of 2 μm exceeded the calibration threshold of 1.5 μm. For these out-of-tolerance items, the corresponding core modules in the digital twin were traced. For example, the residual stress prediction deviation stemmed from an inaccurate mapping of the yield strength parameter in the material-property mapping relationship, and the subsurface grain size deviation originated from an unreasonable grain growth rate model parameter in the microstructure evolution calculation logic. Actual results were used to calibrate the parameters of these core modules. For instance, the yield strength corresponding to a certain microstate in the material-property mapping relationship was adjusted from 250 MPa to 240 MPa, and the grain growth rate coefficient in the microstructure evolution calculation logic was adjusted from 0.003 to 0.0028. After calibration, the digital twin was updated to make its subsequent simulation predictions more closely match actual processing conditions.

[0102] Example 2, as Figure 2As shown, based on the same inventive concept as the machine learning-based adaptive machining control method for CNC machine tools provided in Embodiment 1, this embodiment of the invention also provides a machine learning-based adaptive machining control system for CNC machine tools, including: The mapping relationship acquisition module 11 is used to construct a digital material specimen reflecting the initial microstructure characteristics of the target workpiece material based on the grade and initial heat treatment state of the target workpiece material, and to establish a quantitative mapping relationship library between the microstructure characteristics and macroscopic mechanical / thermal performance parameters as a material-performance mapping relationship set. The digital twin construction module 12 is used to construct a digital twin based on the material-property mapping relationship set and the preset tool geometry model and machine tool dynamics model. The digital twin can predict the microstructure evolution of the workpiece material based on the thermal-mechanical load history of local processing and update the material properties based on the microstructure evolution. The optimal robustness parameter acquisition module 13 is used to optimize the processing parameters of the target processing task based on the digital twin, with processing quality, processing efficiency and microstructure evolution index as optimization objectives and the inherent fluctuation of the micro-characteristics of the target workpiece material as uncertainty variables, to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map. The parameter fine-tuning instruction generation module 14 is used to load the optimal robustness parameter package in actual processing and simultaneously start the online fine-tuning module; the online fine-tuning module, based on real-time sensor data, infers the equivalent microstate offset of the workpiece processing area through the built-in microstate observer, and generates parameter fine-tuning instructions within the preset robustness envelope in combination with the parameter sensitivity map. The digital twin update module 15 is used to execute the optimal robustness parameter package and the parameter fine-tuning instructions generated in real time, complete the processing, and update the digital twin based on the processing data.

[0103] Furthermore, the mapping relationship acquisition module 11 includes the following execution steps: Obtain the grade of the target workpiece material and its corresponding initial heat treatment process; Based on the grade, the material database is queried to obtain the standard chemical composition, phase diagram and basic physical property parameters of the target workpiece material; If measured data of the target workpiece material exists, import the electron backscatter diffraction or metallographic image of the target workpiece material, and digitally reconstruct the real microstructure model containing grains, grain boundaries and second phase through image segmentation and vectorization processing. If no measured data of the target workpiece material is available, then based on the chemical composition, initial heat treatment process and material phase transformation kinetics principle, a virtual microstructure model that is statistically representative is generated by cellular automata or Monte Carlo method. Based on the real microstructure model or the virtual microstructure model, a digital model expressed in the form of a three-dimensional voxel mesh is constructed for the target workpiece material as the digital material specimen, wherein each mesh in the digital material specimen is associated with a corresponding material phase and crystallographic direction.

[0104] Using the digital material sample as input, the macroscopic mechanical / thermal properties under different combinations of microstructure features are predicted by performing crystal plasticity finite element simulation calculations. The macroscopic mechanical property parameters include at least yield strength, flow stress and strain hardening index, and the macroscopic thermal property parameters include at least thermal conductivity and specific heat capacity. By using microstructural features as independent variables and the macroscopic mechanical / thermal properties obtained from simulation as dependent variables, a structured and queryable mapping relationship database is constructed as the material-property mapping relationship set.

[0105] Furthermore, the digital twin building module 12 includes the following execution steps: Obtain the preset tool geometry model and machine tool dynamics model; Based on the aforementioned material-property mapping relationship set, the preset tool geometry model, and the machine tool dynamics model, a digital twin of the macroscopic thermo-mechanical coupling machining process is established; Within each macroscopic material property calculation unit of the digital twin, a microstructure evolution calculation unit is nested. The microstructure evolution calculation unit calculates the microscopic property data within the macroscopic material property calculation unit by calling the embedded crystal plasticity finite element model based on the macroscopic property data provided by the macroscopic material property calculation unit to which it is located. The macroscopic performance data includes local strain, strain rate, and temperature history data; the microscopic performance data includes the evolution of grain size, dislocation density, and phase fraction. A data transmission channel is established from the macroscopic material property calculation unit to the microstructure evolution calculation unit for real-time transmission of the macroscopic property data; Simultaneously, a data feedback channel is established to return from the microstructure evolution calculation unit to the macromaterial property calculation unit, which is used to update the material properties in the macromaterial property calculation unit based on the calculated microstructure property data.

[0106] Furthermore, the optimal robustness parameter acquisition module 13 includes the following execution steps: Define an optimization multi-objective function, with objective terms including at least processing quality, processing efficiency, and microstructure evolution indicators; Among them, the processing quality target is characterized by the surface roughness and residual stress predicted by the digital twin, the processing efficiency target is characterized by the material removal rate, and the microstructure evolution index is characterized by the degree of subsurface grain refinement and the phase change layer depth. The statistical distribution of key microscopic features extracted from the digital material specimens is set as an interval-type uncertainty variable in the optimization process. Based on the digital twin, a multi-objective robust optimization algorithm is used to perform global optimization within the processing parameter space; During the optimization process, for each set of candidate processing parameters, multiple sampling simulations are performed within the range of the uncertainty variables to calculate the mean and variance of the performance of each objective item. The optimal robustness parameter package is the combination of processing parameters with the best overall performance and the smallest variance in the Pareto front. The parameter sensitivity map calculates the gradient of the influence of each processing parameter on each sub-indicator in the multi-objective function and presents it in the form of a visual data matrix.

[0107] Furthermore, the parameter fine-tuning instruction generation module 14 includes the following execution steps: The default robustness envelope is defined as follows: Based on the aforementioned optimal robustness parameter package, parameter perturbation analysis is performed in the digital twin; To obtain the maximum positive and negative adjustment range of each key processing parameter while keeping all processing quality indicators and microstructure evolution indicators in line with preset standards; The maximum positive and maximum negative adjustment amplitudes together form a super-rectangular region in a multi-dimensional parameter space, and the super-rectangular region is used as the preset robustness envelope.

[0108] The construction of the microstate observer includes: Based on the digital twin, a dataset containing various simulated microscopic state fluctuations and corresponding sensor response characteristics is generated; A lightweight machine learning model is trained using the dataset. The input of the machine learning model includes time-frequency domain features of real-time cutting force, vibration, and acoustic emission signals. The output includes estimates of the equivalent hardness and equivalent toughness state of the current machining area of ​​the target workpiece. The trained machine learning model is deployed as the microstate observer.

[0109] Based on the parameter sensitivity map, parameter fine-tuning instructions are generated within a preset robustness envelope, including: The parameter sensitivity map records the gradient of the influence of each processing parameter on the processing quality index. When the microstate observer infers that the equivalent microstate has shifted, it queries the parameter sensitivity map based on the direction and magnitude of the shift to identify the adjustment parameters and their adjustment directions that can most effectively compensate for the impact of the shift on key quality indicators. Based on the adjustment parameters and their adjustment direction, and in accordance with predefined compensation rules, parameter fine-tuning instructions are generated under the constraints of the robustness envelope.

[0110] Furthermore, the digital twin update module 15 includes the following execution steps: In the actual processing, the processing control system loads and executes the optimal robustness parameter package as the basic processing instructions, and at the same time receives and integrates the parameter fine-tuning instructions generated in real time by the online fine-tuning module to form the final execution instruction sequence and perform the processing. By deploying a sensor network on the machine tool and the target workpiece, the actual cutting force, vibration, acoustic emission and temperature data are collected synchronously throughout the entire machining cycle, which serves as the actual machining process dataset. After processing, the target workpiece is subjected to quality inspection to obtain actual data on the surface integrity, dimensional accuracy and microstructure of the target workpiece; The actual processing data set and the actual result data are input into the digital twin for post-processing simulation verification, and the prediction results of the digital twin are compared with the actual result data. If the prediction deviation exceeds the preset calibration threshold, the actual result data is used to calibrate the parameters of the corresponding material-property mapping relationship or microstructure evolution calculation logic in the digital twin.

Claims

1. A machine learning-based adaptive machining control method for CNC machine tools, characterized in that, include: Based on the grade and initial heat treatment state of the target workpiece material, a digital material specimen reflecting the initial microstructure characteristics of the target workpiece material is constructed, and a quantitative mapping relationship library between the microstructure characteristics and macroscopic mechanical / thermal performance parameters is established as a material-performance mapping relationship set. Based on the material-property mapping relationship set, as well as the preset tool geometry model and machine tool dynamics model, a digital twin is constructed. The digital twin can predict the microstructure evolution of the workpiece material based on the thermal-mechanical load history of local processing and update the material properties based on the microstructure evolution. With processing quality, processing efficiency and microstructure evolution indicators as optimization targets, and the inherent fluctuations of the microstructure characteristics of the target workpiece material as uncertainty variables, the processing parameters of the target processing task are optimized based on the digital twin to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map. In actual processing, the optimal robustness parameter package is loaded, and the online fine-tuning module is started simultaneously. The online fine-tuning module, based on real-time sensor data, infers the equivalent microstate offset of the workpiece processing area through the built-in microstate observer, and generates parameter fine-tuning instructions within the preset robustness envelope by combining the parameter sensitivity map. The optimal robustness parameter package and the real-time generated parameter fine-tuning instructions are executed to complete the processing, and the digital twin is updated based on the processing data.

2. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, Based on the grade and initial heat treatment state of the target workpiece material, a digital material specimen reflecting the initial microstructural characteristics of the target workpiece material is constructed, including: Obtain the grade of the target workpiece material and its corresponding initial heat treatment process; Based on the grade, the material database is queried to obtain the standard chemical composition, phase diagram and basic physical property parameters of the target workpiece material; If measured data of the target workpiece material exists, import the electron backscatter diffraction or metallographic image of the target workpiece material, and digitally reconstruct the real microstructure model containing grains, grain boundaries and second phase through image segmentation and vectorization processing. If no measured data of the target workpiece material is available, then based on the chemical composition, initial heat treatment process and material phase transformation kinetics principle, a virtual microstructure model that is statistically representative is generated by cellular automata or Monte Carlo method. Based on the real microstructure model or the virtual microstructure model, a digital model expressed in the form of a three-dimensional voxel mesh is constructed for the target workpiece material as the digital material specimen, wherein each mesh in the digital material specimen is associated with a corresponding material phase and crystallographic direction.

3. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, Establish a quantitative mapping relationship library between the aforementioned microstructural features and macroscopic mechanical / thermal performance parameters, as a material-property mapping relationship set, including: Using the digital material sample as input, the macroscopic mechanical / thermal properties under different combinations of microstructure features are predicted by performing crystal plasticity finite element simulation calculations. The macroscopic mechanical property parameters include at least yield strength, flow stress and strain hardening index, and the macroscopic thermal property parameters include at least thermal conductivity and specific heat capacity. By using microstructural features as independent variables and the macroscopic mechanical / thermal properties obtained from simulation as dependent variables, a structured and queryable mapping relationship database is constructed as the material-property mapping relationship set.

4. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, Based on the aforementioned material-property mapping relationship set, and the preset tool geometry model and machine tool dynamics model, a digital twin is constructed, including: Obtain the preset tool geometry model and machine tool dynamics model; Based on the aforementioned material-property mapping relationship set, the preset tool geometry model, and the machine tool dynamics model, a digital twin of the macroscopic thermo-mechanical coupling machining process is established; Within each macroscopic material property calculation unit of the digital twin, a microstructure evolution calculation unit is nested. The microstructure evolution calculation unit calculates the microscopic property data within the macroscopic material property calculation unit by calling the embedded crystal plasticity finite element model based on the macroscopic property data provided by the macroscopic material property calculation unit to which it is located. The macroscopic performance data includes local strain, strain rate, and temperature history data; the microscopic performance data includes the evolution of grain size, dislocation density, and phase fraction. A data transmission channel is established from the macroscopic material property calculation unit to the microstructure evolution calculation unit for real-time transmission of the macroscopic property data; Simultaneously, a data feedback channel is established to return from the microstructure evolution calculation unit to the macromaterial property calculation unit, which is used to update the material properties in the macromaterial property calculation unit based on the calculated microstructure property data.

5. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, Using processing quality, processing efficiency, and microstructure evolution indicators as optimization objectives, and taking the inherent fluctuations of the microscopic characteristics of the target workpiece material as uncertainty variables, the processing parameters for the target processing task are optimized based on the digital twin to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map, including: Define an optimization multi-objective function, with objective terms including at least processing quality, processing efficiency, and microstructure evolution indicators; Among them, the processing quality target is characterized by the surface roughness and residual stress predicted by the digital twin, the processing efficiency target is characterized by the material removal rate, and the microstructure evolution index is characterized by the degree of subsurface grain refinement and the phase change layer depth. The statistical distribution of key microscopic features extracted from the digital material specimens is set as an interval-type uncertainty variable in the optimization process. Based on the digital twin, a multi-objective robust optimization algorithm is used to perform global optimization within the processing parameter space; During the optimization process, for each set of candidate processing parameters, multiple sampling simulations are performed within the range of the uncertainty variables to calculate the mean and variance of the performance of each objective item. The optimal robustness parameter package is the combination of processing parameters with the best overall performance and the smallest variance in the Pareto front. The parameter sensitivity map calculates the gradient of the influence of each processing parameter on each sub-indicator in the multi-objective function and presents it in the form of a visual data matrix.

6. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, The preset robustness envelope is defined in the following way: Based on the aforementioned optimal robustness parameter package, parameter perturbation analysis is performed in the digital twin; To obtain the maximum positive and negative adjustment range of each key processing parameter while keeping all processing quality indicators and microstructure evolution indicators in line with preset standards; The maximum positive and maximum negative adjustment amplitudes together form a super-rectangular region in a multi-dimensional parameter space, and the super-rectangular region is used as the preset robustness envelope.

7. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, The construction of the microstate observer includes: Based on the digital twin, a dataset containing various simulated microscopic state fluctuations and corresponding sensor response characteristics is generated; A lightweight machine learning model is trained using the dataset. The input of the machine learning model includes time-frequency domain features of real-time cutting force, vibration, and acoustic emission signals. The output includes estimates of the equivalent hardness and equivalent toughness state of the current machining area of ​​the target workpiece. The trained machine learning model is deployed as the microstate observer.

8. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, Based on the parameter sensitivity map, parameter fine-tuning instructions are generated within a preset robustness envelope, including: The parameter sensitivity map records the gradient of the influence of each processing parameter on the processing quality index. When the microstate observer infers that the equivalent microstate has shifted, it queries the parameter sensitivity map based on the direction and magnitude of the shift to identify the adjustment parameters and their adjustment directions that can most effectively compensate for the impact of the shift on key quality indicators. Based on the adjustment parameters and their adjustment direction, and in accordance with predefined compensation rules, parameter fine-tuning instructions are generated under the constraints of the robustness envelope.

9. The adaptive machining control method for CNC machine tools based on machine learning according to claim 1, characterized in that, Execute the optimal robustness parameter package and the real-time generated parameter fine-tuning instructions to complete the processing, and update the digital twin based on the processing data, including: In the actual processing, the processing control system loads and executes the optimal robustness parameter package as the basic processing instructions, and at the same time receives and integrates the parameter fine-tuning instructions generated in real time by the online fine-tuning module to form the final execution instruction sequence and perform the processing. By deploying a sensor network on the machine tool and the target workpiece, the actual cutting force, vibration, acoustic emission and temperature data are collected synchronously throughout the entire machining cycle, which serves as the actual machining process dataset. After processing, the target workpiece is subjected to quality inspection to obtain actual data on the surface integrity, dimensional accuracy and microstructure of the target workpiece; The actual processing data set and the actual result data are input into the digital twin for post-processing simulation verification, and the prediction results of the digital twin are compared with the actual result data. If the prediction deviation exceeds the preset calibration threshold, the actual result data is used to calibrate the parameters of the corresponding material-property mapping relationship or microstructure evolution calculation logic in the digital twin.

10. A machine learning-based adaptive machining control system for CNC machine tools, characterized in that, The system includes: The mapping relationship acquisition module is used to construct a digital material specimen reflecting the initial microstructure characteristics of the target workpiece material based on the grade and initial heat treatment state of the target workpiece material, and to establish a quantitative mapping relationship library between the microstructure characteristics and macroscopic mechanical / thermal performance parameters as a material-performance mapping relationship set. The digital twin construction module is used to construct a digital twin based on the material-property mapping relationship set and the preset tool geometry model and machine tool dynamics model. The digital twin can predict the microstructure evolution of the workpiece material according to the thermal-mechanical load history of local processing and update the material properties based on the microstructure evolution. The optimal robustness parameter acquisition module is used to optimize the processing parameters of the target processing task based on the digital twin, with processing quality, processing efficiency and microstructure evolution indicators as optimization objectives and the inherent fluctuation of the micro-characteristics of the target workpiece material as uncertainty variables, to obtain the optimal robustness parameter package and the corresponding parameter sensitivity map. The parameter fine-tuning instruction generation module is used to load the optimal robust parameter package during actual processing and simultaneously start the online fine-tuning module. The online fine-tuning module, based on real-time sensor data, infers the equivalent microstate offset of the workpiece processing area through the built-in microstate observer, and generates parameter fine-tuning instructions within the preset robustness envelope by combining the parameter sensitivity map. The digital twin update module is used to execute the optimal robustness parameter package and the parameter fine-tuning instructions generated in real time, complete the processing, and update the digital twin based on the processing data.

Images