Image modeling and microstructure identification method and control system

Abstract

The invention belongs to the technical field of image detection, and particularly relates to an image modeling and microstructure identification method and a control system, and the method comprises the steps: generating a three-dimensional surface type error distribution diagram through an initial interference pattern of a to-be-processed target object, and carrying out the inversion construction of a target object surface type control parameter space through combining with a preset processing parameter-defect removal rate mapping database; precisely defining the residence time, the pressure, the speed and the regionalized control path form of each defect position point; a fuzzy control algorithm is introduced based on a parameter space to dynamically generate an ultrasonic hybrid control instruction, and hybrid ultrasonic parameters are regulated and controlled in real time; surface precision and surface quality evaluation errors are further extracted through the secondary interference pattern, a parameter mapping database is iteratively optimized through closed-loop feedback, and a control parameter space is adjusted till a preset polishing index is met; according to the method, the problem of balancing surface type correction and surface roughness in a traditional process is solved, and the defect removal efficiency and precision of the target object are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image detection, and in particular relates to an image modeling and microstructure recognition method and a control system. Background Art

[0002] In the field of fluoride target processing, traditional processing methods have many limitations. For the surface accuracy control of fluoride crystals, the existing traditional processing methods rely on manual experience to repair the defective areas, and there is a lack of quantitative basis for the control path and residence time, which is prone to defect residue or over-polishing. For example, in the surface detection and subsequent processing after rough defect treatment, the traditional method simply divides the defects into several levels and sets a unified pressure, speed and travel path according to the level. It cannot accurately process defects of different positions and types, resulting in the surface shape of the processed crystal still difficult to meet high-precision requirements; moreover, the traditional method does not fully consider the relationship between material removal rate and processing parameters, and cannot achieve quantitative defect removal based on the surface shape diagram; in addition, when evaluating the processing effect, the use of a single indicator can easily lead to performance imbalance and cannot fully reflect the surface accuracy and surface quality of the crystal. With the increasing requirements for the surface accuracy and surface quality of fluoride crystals in the optical field, extremely strict standards are put forward for indicators such as the surface error (PV value, RMS value) and surface roughness (Ra) of the target object in applications such as extreme ultraviolet optics and infrared optics. Existing processing technology can no longer meet the needs of these high-end applications. A more accurate, intelligent and comprehensive crystal dynamic detection and control method is urgently needed. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention proposes an image modeling and microstructure recognition method and control system. The method generates a three-dimensional surface error distribution map through the initial interference pattern of the target object to be processed, and combines the preset processing parameter-defect removal rate mapping database to inversely construct the target object surface control parameter space, accurately defining the residence time, pressure, speed and regional control path form of each defect position point; based on the parameter space, a fuzzy control algorithm is introduced to dynamically generate ultrasonic hybrid control instructions, and the hybrid ultrasonic parameters are adjusted in real time; the secondary interference pattern is further used to extract the surface accuracy and surface quality assessment error, and the parameter mapping database is iteratively optimized through closed-loop feedback and the control parameter space is adjusted until the preset polishing indicators are met; the present invention breaks through the difficult problem of balancing surface correction and surface roughness in traditional processes, and achieves an improvement in the polishing efficiency and accuracy of the target object.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] An image modeling and microstructure recognition method, comprising:

[0006] Based on the initial interference pattern of the target object to be processed, a three-dimensional error distribution map of the target object surface is obtained;

[0007] Based on the target object surface three-dimensional error distribution map and combined with a preset processing parameter-defect removal rate mapping database, a target object surface control parameter space is obtained by inversion; the target object surface control parameter space includes each defect location point and type, the residence time length of the defect removal at each location point, pressure, speed, and the control path form of different defect areas of the target object surface;

[0008] Generate ultrasonic hybrid control instructions based on the target object surface control parameter space combined with fuzzy control algorithm to perform real-time defect treatment on the target object to be processed, and obtain a secondary interference map after defect treatment;

[0009] Based on the secondary interference pattern after defect processing and the preset target object surface evaluation index, an evaluation error value is obtained, and the evaluation error value is fed back to the processing parameter-defect removal rate mapping database to adjust the target object surface control parameter space in real time until the preset target object surface accuracy requirement is met;

[0010] The target object surface shape evaluation index includes the valley-peak value and RMS value corresponding to the target object surface shape error.

[0011] Specifically, the process of obtaining the target object surface three-dimensional error distribution map includes:

[0012] Perform an initial scan on the target object using a laser interferometer with dual-wavelength module parameters and initial scanning parameters to obtain an initial surface gradient map;

[0013] Based on the initial surface gradient map and the Hessian matrix, the local curvature is calculated to obtain the high curvature area and the low curvature area in the initial surface gradient map and label the areas;

[0014] Based on the labeled high curvature area and low curvature area information, the high curvature point cloud spacing and the low curvature point cloud spacing are set, and based on the set point cloud spacing and the regional curvature type, the labeled high curvature area point cloud map and the low curvature area point cloud map are generated;

[0015] Based on the labeled point cloud images of high curvature areas and low curvature areas, the Delaunay triangulation algorithm and the spiral path algorithm are combined to search for the scanning path along the direction of maximum curvature to generate a spiral scanning path.

[0016] Specifically, the process of obtaining the target object surface three-dimensional error distribution map also includes:

[0017] The target object to be processed is scanned based on a spiral scanning path, and the surface gradient map is updated every time 10% of the area of ​​the target object's surface is scanned. The high curvature area and low curvature area of ​​the target object to be processed are re-divided once until the entire scan of the target object to be processed is completed, thereby obtaining an initial interference map of the target object's surface.

[0018] The initial interferogram of the target surface is filtered and processed, and the initial phase space of the interferogram is obtained through the phase shift interferometry algorithm;

[0019] Based on the initial phase space of the interferogram and the geometric information of the target object, the surface shape of the target object to be processed is iteratively solved by the phase unwrapping algorithm combined with the least squares method. Combined with the three-dimensional coordinate system with the center of gravity of the target object as the origin, the three-dimensional defect distribution map of the target object surface is obtained.

[0020] Based on the target object's surface three-dimensional defect distribution map and a preset multi-scale defect recognition and positioning model, the defects on the target object's surface three-dimensional distribution map are identified and located, and the defect type identified and located on the target object's surface three-dimensional distribution map, the corresponding position point of each defect, and the gradient distribution and height distribution are obtained;

[0021] The multi-scale defect recognition and positioning model is constructed and pre-trained in an integrated manner by combining a gradient threshold constructed by combining the surface gradient amplitude with the light source wavelength and a preset defect detection algorithm under different gradient amplitudes.

[0022] Specifically, the process of obtaining the target object surface three-dimensional error distribution map also includes:

[0023] Based on the defect types identified and located on the three-dimensional distribution map of the target surface, the location points corresponding to each defect, the gradient distribution and the height distribution, combined with the PV value of the target surface and the wavelength of the light source, regional clustering is performed using the DBSCAN clustering algorithm combined with the region growing method to obtain the target surface regional defect distribution map;

[0024] Based on the three-dimensional regional defect distribution map and the three-dimensional distribution map of the standard target object, a three-dimensional regional error distribution map of the target object surface is obtained.

[0025] Specifically, the process of constructing the machining parameter-defect removal rate mapping database includes:

[0026] Obtaining the ultrasonic processing parameter sequence and the second treatment solution processing parameter sequence and combining them with the preset variable bin values ​​to construct an ideal experimental variable orthogonal table;

[0027] The ultrasonic processing parameter sequence includes frequency, amplitude, pressure, and dwell time length of each position point; the second processing solution processing parameter sequence includes the second processing solution concentration, flow rate per unit time, and corresponding equivalent pressure;

[0028] Obtain a target material of the same material as the target object to be processed and select a step-by-step control parameter sequence corresponding to each ideal experiment based on a random algorithm using an orthogonal table of ideal experimental variables; the step-by-step control parameter sequence includes frequency, amplitude, pressure, dwell time length at each position point, and equivalent pressure;

[0029] The obtained grading control parameter sequence is input into the ultrasonic mixing system to carry out an ideal polishing experiment, and the defect removal rate corresponding to different defect types per unit time under each ideal experiment is obtained in real time.

[0030] Specifically, the process of constructing the processing parameter-defect removal rate mapping database also includes:

[0031] The defect removal rate is equal to the ratio of the defect height removed within the corresponding residence time length to the original defect height and the residence time length;

[0032] Based on the defect removal rates of different defect types per unit time under each ideal experiment, the optimal bin control parameter sequence with the maximum defect removal rate for each defect type is obtained;

[0033] Based on the optimal control parameter sequence and corresponding defect removal rate corresponding to each type and height of defects, the processing parameter-defect removal rate mapping function and the corresponding binned control parameter values ​​are obtained through convolution and function fitting;

[0034] Based on the processing parameter-defect removal rate mapping function and the corresponding grading control parameter values, a processing parameter-defect removal rate mapping database is obtained through a graph database.

[0035] Specifically, the process of generating ultrasonic hybrid control instructions includes:

[0036] Obtain a real-time target object surface three-dimensional regional error distribution map, combine it with the processing parameter-defect removal rate mapping database, and use a matching algorithm to invert and obtain the optimal binning control parameter sequence set corresponding to all position points of the current target object surface;

[0037] Based on the defect type, position, PV value and gradient amplitude information corresponding to each defect cluster area in the three-dimensional regional error distribution map of the current target object surface, combined with the path algorithm optimized by the ant colony algorithm, the corresponding defect position control path and normal position control path in each cluster area on the target object surface are obtained.

[0038] Specifically, the process of generating the ultrasonic hybrid control instruction further includes:

[0039] Based on the optimal bin control parameter sequence set corresponding to all position points of the current target object surface type and the corresponding defect control path and normal position control path in each cluster area on the target object surface type, the normal position control instruction and defect position control instruction are obtained through the fuzzy control algorithm;

[0040] Perform defect processing on the target object based on the normal position control instruction and the defect position control instruction, and detect and obtain the valley peak value and RMS value corresponding to the target object after defect processing;

[0041] The valley-peak value and RMS value corresponding to the target object after defect processing are fed back to the processing parameter-defect removal rate mapping database and the path algorithm optimized by the ant colony algorithm to perform real-time optimization on the defect position control path and the normal position control path corresponding to each cluster area on the target object surface and the optimal grading control parameter sequence set until the valley-peak value and RMS value corresponding to the target object surface meet the preset processing threshold position.

[0042] An image modeling and microstructure control system, comprising: a three-dimensional modeling module, an inversion module, a control module and an evaluation module;

[0043] The three-dimensional modeling module obtains a three-dimensional error distribution map of the target object surface based on the initial interference map of the target object to be processed;

[0044] The inversion module obtains a target object surface control parameter space based on the target object surface three-dimensional error distribution map combined with a preset processing parameter-defect removal rate mapping database; the target object surface control parameter space includes each defect location point and type, the length of the defect residence time at each location point, pressure, speed, and control path forms for different defect areas of the target object surface;

[0045] The control module generates ultrasonic hybrid control instructions based on the target object surface control parameter space and fuzzy control algorithm to perform real-time defect treatment on the target object to be processed, and obtains a secondary interference map after the defect treatment;

[0046] The evaluation module obtains an evaluation error value based on the secondary interference pattern after defect processing and the preset target object surface evaluation index, and feeds the evaluation error value back to the processing parameter-defect removal rate mapping database to adjust the target object surface control parameter space in real time until the preset target object surface accuracy requirement is met.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] In response to the deficiencies of the prior art, the present invention significantly improves the spatial resolution accuracy and quantitative characterization capability of surface defects by constructing a precise mapping relationship between the initial interference pattern and the three-dimensional error distribution map, and realizes the digital modeling of sub-micron errors. The inversion mechanism based on the parameter-removal rate mapping database effectively integrates multi-source process data, and can dynamically match the optimal polishing parameter combination for different defect morphologies, effectively overcoming the technical bottleneck of low efficiency of parameter optimization in the traditional trial-and-error method. The polishing instruction generation system constructed by introducing the fuzzy control algorithm realizes the coordinated regulation of polishing pressure, speed and path through multi-modal sensor feedback, and enhances the adaptability of the process system to nonlinear material removal processes. The closed-loop evaluation system based on the secondary interference pattern is combined with the online feedback correction mechanism to establish a self-consistent association between the error evolution model and the process parameters, ensuring that the polishing process has real-time error compensation function. This data-driven intelligent polishing control architecture breaks through the limitations of the traditional open-loop processing mode, greatly improves the robustness and convergence efficiency of the process system, and provides a reliable technical approach for high-precision deterministic polishing of complex crystal components. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of an image modeling and microstructure recognition method of the present invention;

[0050] Figure 2 This is a module diagram of an image modeling and microstructure control system of the present invention. DETAILED DESCRIPTION

[0051] Example 1

[0052] See also Figure 1 The present invention provides an embodiment of an image modeling and microstructure recognition method, which is applied to the polishing control process of the fluoride crystal surface, comprising the following steps:

[0053] S1. Based on the initial interference pattern of the target object to be processed, a three-dimensional error distribution map of the target object surface is obtained;

[0054] S2. Inverting the target object surface shape control parameter space based on the target object surface shape three-dimensional error distribution map in combination with a preset processing parameter-defect removal rate mapping database; the target object surface shape control parameter space includes each defect location point and type, the residence time length, pressure, speed, and control path forms for different defect areas of the target object surface shape at each location point;

[0055] S3, generating an ultrasonic hybrid control instruction based on the target object surface control parameter space combined with a fuzzy control algorithm to perform real-time defect treatment on the target object to be processed, and obtaining a secondary interference map after the defect treatment;

[0056] S4. Based on the secondary interference pattern after defect processing and the preset target object surface evaluation index, an evaluation error value is obtained, and the evaluation error value is fed back to the processing parameter-defect removal rate mapping database to adjust the target object surface control parameter space in real time until the preset target object surface accuracy requirement is met.

[0057] Furthermore, the process of obtaining the target object surface three-dimensional error distribution map in this embodiment includes:

[0058] Perform an initial scan on the target object using a laser interferometer with dual-wavelength module parameters and initial scanning parameters to obtain an initial surface gradient map;

[0059] Furthermore, the initial surface gradient map of this embodiment is a 1024×1024 pixel matrix, and each pixel stores a defect height gradient value;

[0060] Furthermore, the dual-wavelength module parameters set in this embodiment include the laser vertical resolution and the scanning range coverage diameter; the initial scanning parameters in this embodiment include the spiral progressive scanning path, pitch, scanning speed, and phase measurement triggered every 5mm interval, etc., to ensure full surface coverage without omission;

[0061] Based on the initial surface gradient map and the Hessian matrix, the local curvature is calculated to obtain the high curvature area and the low curvature area in the initial surface gradient map and label the areas;

[0062] Furthermore, the process of obtaining the high curvature region and the low curvature region in this embodiment includes:

[0063] The local curvature K is obtained by calculating the local curvature based on the initial surface gradient map and the Hessian matrix.

[0064] When K ≥ k μm -1 is a high curvature area, otherwise it is a low curvature area, where k is the high and low curvature classification threshold preset in this embodiment;

[0065] Based on the labeled high curvature area and low curvature area information, the high curvature point cloud spacing and the low curvature point cloud spacing are set, and based on the set point cloud spacing and the regional curvature type, a labeled high curvature area point cloud map and a labeled low curvature area point cloud map are generated. Furthermore, this embodiment uses a region growing algorithm to merge adjacent high curvature points to generate a labeled point cloud map; wherein the labeled point cloud map includes a high curvature area point cloud map and a low curvature area point cloud map;

[0066] Based on the labeled high-curvature area point cloud map and low-curvature area point cloud map, the Delaunay triangulation algorithm and the spiral path algorithm are combined to search the scanning path along the maximum curvature direction to generate a spiral scanning path;

[0067] Furthermore, the process of generating the spiral scanning path in this embodiment includes:

[0068] The point cloud is converted into a topology-optimized triangulated mesh using the Delaunay triangulation algorithm, and the empty circle criterion is used to ensure geometric stability.

[0069] It should be noted that in this process, the discrete point cloud is converted into a non-overlapping triangular mesh using the Delaunay triangulation algorithm, and the empty circle criterion is used to ensure the optimality of the mesh topology and avoid geometric distortion caused by narrow triangles. It should be noted that the empty circle criterion specifically means that the circumcircle of each triangle does not contain other input points.

[0070] A curvature-driven path generation algorithm is used to prioritize tracking the direction of maximum curvature along the edges of the triangle mesh, and a dynamic spacing adjustment mechanism is combined to adapt to the scanning accuracy requirements of high and low curvature areas. It should be noted that in this embodiment, curvature-driven path generation is based on the discrete curvature estimates at the vertices of the triangle mesh, and the scanning path is preferentially generated along the direction of maximum principal curvature.

[0071] Path optimization technology is integrated with the A* algorithm to achieve obstacle avoidance, and B-spline curve fitting is used to enhance path continuity and smoothness;

[0072] It should be noted that this embodiment runs an improved A* algorithm on a triangular mesh, using mesh edge weights as path costs to search for a global collision-free path. The path nodes are then fitted using cubic B-spline curves. Control point interpolation is used to constrain curvature continuity, eliminate sudden changes in direction, and ensure smooth motion of the scanning robot arm. The path cost is constructed from the curvature weights of the corresponding meshes on the triangular mesh.

[0073] Ultimately, based on motion control coding principles, the path geometry parameters are converted into numerical control instructions, outputting a high-precision spiral scanning trajectory. This entire technology combines curvature adaptive analysis, dynamic obstacle avoidance planning, and kinematic coding to achieve autonomous generation of high-fidelity scanning paths for complex curved surfaces.

[0074] The target object to be processed is scanned based on a spiral scanning path, and the surface gradient map is updated every time 10% of the surface area of ​​the target object to be processed is scanned. The high curvature area and low curvature area of ​​the target object to be processed are re-divided once until the entire scan of the target object to be processed is completed, thereby obtaining an initial interference map of the target object surface.

[0075] Furthermore, in this embodiment, the process of dynamic scanning and updating the surface gradient map includes:

[0076] The spiral path scanning is performed in stages through periodic triggered scanning, and the local surface gradient map is dynamically updated based on real-time interferometric measurement data;

[0077] The curvature recalculation and area repartitioning algorithms are used to adjust the high-low curvature labels according to the updated gradient map, and the adaptive path planning algorithm is combined to dynamically encrypt the scan density of the defect area.

[0078] An incremental synthesis algorithm is used to fuse the phased gradient image sequence into a global surface interference image, and a closed-loop feedback mechanism is used to iteratively optimize the scanning accuracy and defect coverage.

[0079] The entire technology is driven by real-time surface data, adaptively adjusts the dynamic path, and uses incremental fusion principles to synchronize the scanning process and surface evolution, ensuring accurate capture of complex surface defects and high-fidelity reconstruction of global interference patterns.

[0080] The initial interferogram of the target surface is filtered and processed, and the initial phase space of the interferogram is obtained through the phase shift interferometry algorithm;

[0081] Based on the initial phase space of the interferogram and the geometric information of the target object, the surface shape of the target object to be processed is iteratively solved by the phase unwrapping algorithm combined with the least squares method. Combined with the three-dimensional coordinate system with the center of gravity of the target object as the origin, the three-dimensional defect distribution map of the target object surface is obtained.

[0082] Furthermore, the process of phase unwrapping and generating a three-dimensional defect distribution map in this embodiment includes:

[0083] Based on the initial phase space of the interferogram and the geometric information of the target, the nonlinear median filter algorithm is used to suppress the impulse noise interference, and the Gaussian low-pass filter algorithm is combined to smooth the high-frequency noise and improve the signal-to-noise ratio of the interferogram;

[0084] The phase-shifted image sequence is acquired based on a multi-frame phase-shifting interferometry algorithm. The wrapped phase distribution is extracted using a phase unwrapping algorithm. The continuous phase field corresponding to the target object surface is then solved and reconstructed based on the fast Poisson equation.

[0085] Utilize 3D geometric mapping technology to convert the continuous phase field into a surface height distribution based on the geometric center of the target object. Combined with gradient field analysis and surface parameter quantification model, a 3D defect distribution map of the target object is generated, including coordinates, gradients, peak-valley values, and defect types.

[0086] In this embodiment, the process realizes visualization and quantitative analysis from noisy interference patterns to high-precision three-dimensional defect features through the collaborative process of noise suppression, phase field reconstruction and geometric mapping.

[0087] The target object's three-dimensional surface defect distribution map is combined with a preset multi-scale defect recognition and positioning model to identify and locate defects on the target object's three-dimensional surface defect distribution map, and obtain the defect type identified and located on the target object's three-dimensional surface defect distribution map, the corresponding position point of each defect, and the gradient distribution and height distribution;

[0088] Further, the process of multi-scale defect recognition and localization in this embodiment includes:

[0089] The three-dimensional defects are divided into macroscopic, intermediate-frequency, and microscopic levels according to physical scale characteristics through a hierarchical gradient threshold segmentation algorithm. The wavelength characteristics of intermediate-frequency defects are extracted based on multi-scale wavelet analysis technology, and microscopic subsurface cracks are identified by combining cross-scale detection algorithms. Further, in this embodiment, the macroscopic defect scale is set as PV > λ / 20, the intermediate-frequency defect scale is λ / 50 < PV < λ / 20, and the microscopic defect scale is PV < λ / 50, where λ is the laser wavelength.

[0090] A multi-feature fusion classification algorithm is used to integrate the geometric shape, gradient distribution, and depth parameters of the target object, and a supervised learning model is used to train high-dimensional feature vectors to achieve accurate discrimination of defect types.

[0091] Based on the pattern recognition and physical property mapping algorithm, a defect feature table containing multi-dimensional attributes is generated. Further, in this embodiment, the defect attribute table preferably includes the position, type, PV value, gradient, and height distribution of each defect.

[0092] This process realizes the automatic analysis and structured output of full-scale defect parameters from three-dimensional surface data through the cooperation of hierarchical threshold segmentation, cross-scale feature extraction, and multi-modal classification.

[0093] The multi-scale defect recognition and localization model is constructed and pre-trained by combining the surface gradient amplitude with hierarchical gradient thresholds and preset defect detection algorithms at different gradient amplitudes.

[0094] Based on the defect types identified and located on the three-dimensional distribution map of the target object surface, the position points corresponding to each defect, the gradient distribution, and the height distribution, combined with the PV value of the target object surface and the light source wavelength, regional clustering is performed by combining the DBSCAN clustering algorithm with the region growing method to obtain the regional defect distribution map of the target object surface.

[0095] Further, the process of regional clustering by combining the DBSCAN clustering algorithm with the region growing method in this embodiment includes: core area clustering, transition area processing, and edge optimization.

[0096] Further, the process of core area clustering in this embodiment includes:

[0097] First, the DBSCAN algorithm is combined with the spatial distance matrix to perform core area clustering on defect points, and dense core areas are identified based on the defect spatial aggregation characteristics and density thresholds to exclude noise interference.

[0098] Second, the region growing algorithm is used to process the transition area. Seed points are generated by meshing the non-core area, and neighborhood expansion and similar area merging are performed according to the defect height gradient threshold to capture potential continuous defects.

[0099] Third, by sampling contour control points and fitting B-spline curves to optimize edges, dense control points are used to maintain geometric details, and high-order continuous curves are used to ensure the smoothness of the processing path, realizing a complete processing flow from discrete defect points to engineering usable area division.

[0100] Based on the three-dimensional regional defect distribution map and the three-dimensional distribution map of the standard target object, a three-dimensional regional error distribution map of the target object surface is obtained.

[0101] This process uses the collaborative optimization of multi-dimensional technical means to systematically integrate dual-wavelength laser interferometers and adaptive spiral scanning strategies to expand the measurement range, and combines gradient-driven dynamic point cloud density adjustment to improve scanning efficiency. Based on the fusion of curvature quantization and region growing algorithms, physical constraint segmentation of defect spatial distribution is achieved, and triangulation and path search algorithms are integrated to generate curvature-guided spiral paths to optimize the topological structure. The phase unwrapping link introduces an optimization model and combines frequency domain acceleration to suppress jump errors. The multi-scale defect recognition model breaks through detection limitations by fusing hierarchical gradient thresholds with wavelet deep learning. Cascaded density clustering and region growing methods are combined with density thresholds and high continuity constraints to resolve clustering ambiguity. Edge optimization uses high-order spline fitting to ensure path continuity and differentiability. The closed-loop feedback mechanism of the entire process dynamically corrects the parameter mapping relationship and path weight, constructing a full-link precision control system covering multiple scales, significantly improving the efficiency of target surface detection and error analysis.

[0102] Furthermore, the process of constructing the processing parameter-defect removal rate mapping database in this embodiment includes:

[0103] Obtaining the ultrasonic processing parameter sequence and the second processing solution processing parameter sequence and combining them with the preset variable bin values, an ideal experimental variable orthogonal table is constructed; it should be noted that in this embodiment, the second processing solution is preferably a polishing solution;

[0104] The ultrasonic processing parameter sequence includes frequency, amplitude, pressure, and dwell time length of each position point; the second processing solution processing parameter sequence includes the second processing solution concentration, flow rate per unit time, and corresponding equivalent pressure;

[0105] Obtain a target material of the same material as the target object to be processed and select a binned control parameter sequence corresponding to each ideal experiment through an orthogonal table of ideal experimental variables using a random algorithm; the control parameter sequence includes frequency, amplitude, pressure, dwell time length at each position point, and equivalent pressure;

[0106] The obtained grading control parameter sequence is input into the ultrasonic mixing system to carry out an ideal polishing experiment, and the defect removal rate corresponding to different defect types per unit time under each ideal experiment is obtained in real time.

[0107] The defect removal rate is equal to the ratio of the defect height removed within the corresponding residence time length to the original defect height and the residence time length;

[0108] Based on the defect removal rates of different defect types per unit time under each ideal experiment, the optimal bin control parameter sequence with the maximum defect removal rate for each defect type is obtained;

[0109] Based on the optimal control parameter sequence and corresponding defect removal rate corresponding to each type and height of defects, the processing parameter-defect removal rate mapping function and the corresponding binned control parameter values ​​are obtained through convolution and function fitting;

[0110] Based on the processing parameter-defect removal rate mapping function and the corresponding grading control parameter values, a processing parameter-defect removal rate mapping database is obtained through a graph database.

[0111] Furthermore, the specific process of constructing the process parameter-defect removal rate mapping database in this embodiment is as follows:

[0112] First, a multi-factor orthogonal experimental design algorithm is used to construct a combination space of ultrasound and second treatment solution parameters, and a statistical sampling algorithm is used to achieve uniform parameter coverage. In this embodiment, the ultrasound parameters are preferably frequency, amplitude, pressure, and residence time; the second treatment solution parameters are preferably concentration, flow rate, and equivalent pressure. The orthogonal table corresponding to the orthogonal experiment is preferably a Taguchi L27 orthogonal table, covering 3 levels × 6 factors, to ensure that any two factor level combinations are evenly distributed. The statistical sampling algorithm is preferably a Latin hypercube sampling algorithm to supplement non-orthogonal experimental points and address the problem of insufficient boundary condition coverage.

[0113] Second, based on standard sample preparation technology, controllable artificial defects made of the same material as the target object to be inspected are generated. Laser interferometer is used to measure the three-dimensional morphology of the defects and dynamically collect defect removal rate data.

[0114] Third, anomaly detection and data normalization techniques are used to clean noise interference, and high removal rate parameter combinations are selected through hierarchical sorting and statistical optimization models;

[0115] It should be noted that in this embodiment, the hierarchical sorting and statistical optimization model is grouped by defect type and initial depth, and arranged in descending order by removal rate within each group, retaining the parameter combination of the previously preset ratio. In this embodiment, the hierarchical sorting and statistical optimization model is preferably a random forest algorithm optimized by principal component analysis;

[0116] Fourth, a quantitative correlation model between processing parameters and defect removal efficiency was established based on defect type classification and parameter mapping function fitting technology, forming a multi-dimensional process knowledge base. This process, through a collaborative process of experimental drive, data modeling, and parameter optimization, achieves a scientific mapping from complex process parameters to defect removal efficiency, providing data support for adaptive polishing decision-making.

[0117] It should be further noted that the quantitative correlation model between processing parameters and defect removal rate in this embodiment is constructed using the ResNet-18 network. It should be noted that the parameter mapping function is constructed by fitting the convolution kernel function using the process parameters, defect type, removal rate range, and the dwell time length corresponding to each type of high-defect.

[0118] This process combines orthogonal experimental design with a parameter binning mechanism to construct an efficient coverage model for the multidimensional variable space. Leveraging the balanced dispersion and neat comparability of orthogonal tables, this significantly reduces the curse of dimensionality in multi-factor coupled experiments, enabling full combinatorial feature extraction of the combined parameters of the ultrasound-second treatment solution within a limited number of experiments. A Latin hypercube sampling algorithm is used to dynamically generate control parameter sequences, breaking through the static parameter constraints of traditional orthogonal experiments. A space-filling optimization strategy ensures uniform distribution of parameter samples within the hypercube, effectively enhancing the process data's ability to characterize the nonlinear parameter space. A data cleaning mechanism based on the Grubbs test, combined with normalization of defect height groups, establishes a robust data preprocessing framework, eliminating anomalous data interference caused by instrument noise and process disturbances and improving the statistical significance of removal rate calculations. A dual grouping and sorting algorithm based on defect type and height interval is used to construct a correlation classification model between defect characteristics and process response, enabling targeted screening of optimal parameters for defects of different morphologies and providing refined data support for parameter-removal rate mapping. A convolution kernel function is used to fit the nonlinear relationship between process parameters and removal rate. The kernel function's multi-scale feature extraction capability captures the coupling mechanism of complex process factors such as high-frequency vibration and fluid dynamics on material removal, establishing a mathematical model that can analyze parameter interaction effects. This hybrid parameter optimization system, based on an orthogonal experimental framework, upgrades discrete process data to a continuous function space, forming a process knowledge base that is both physically interpretable and data-driven. This provides high-precision nonlinear mapping relationships for subsequent parameter inversion, while leveraging the dynamic association characteristics of a graph database to enable rapid retrieval and combinatorial optimization of process parameters.

[0119] Furthermore, the process of generating the ultrasonic hybrid control instruction in this embodiment includes:

[0120] Obtain a real-time target object surface three-dimensional regional error distribution map, combine it with the processing parameter-defect removal rate mapping database, and use a matching algorithm to invert and obtain the optimal binning control parameter sequence set corresponding to all position points of the current target object surface;

[0121] Based on the defect type, location, PV value, and gradient amplitude information corresponding to each defect cluster area in the current target object surface three-dimensional regional error distribution map, combined with the path algorithm optimized by the ant colony algorithm, the corresponding defect position control path and normal position control path in each cluster area on the target object surface are obtained. It should be noted that the control path here is the corresponding polishing path during the target object defect removal process;

[0122] Furthermore, in this embodiment, the defect position control path adopts a spiral path: starting from the defect center of mass, it unfolds like an Archimedean spiral, while the normal position control path adopts a grating path, reciprocating in the X-axis direction, with a Y-axis step of a mm and a speed of bmm / s.

[0123] Furthermore, in this embodiment, the process of optimizing the control path generation using the ant colony algorithm includes:

[0124] First, the surface of the target object is divided into a uniform node network through surface mesh discretization technology, and the defect attribute association is modeled as node weights. It should be noted that the surface mesh discretization technology in this embodiment preferably uses the Marching Cubes algorithm, which uses the target object surface point cloud data as input, adopts octree space segmentation, and combines normal consistency constraints to generate a topologically continuous triangular mesh to eliminate holes and islands.

[0125] The construction process of the uniform node network is as follows: the mesh vertices are projected onto a two-dimensional parameterized plane, path nodes are generated at equal arc length intervals, and an adjacency matrix is ​​established to store the reachability between nodes;

[0126] Second, based on the pheromone concentration difference initialization mechanism, the ant colony algorithm simulates the path search process, prioritizing the control path to cover high-defect areas. It should be noted that the pheromone concentration difference initialization mechanism uses the number and height of defects corresponding to high-defect areas, medium-defect areas, and low-defect areas to obtain corresponding defect weights to differentiate the initial pheromone concentration.

[0127] Third, we leverage the dynamic construction principle of heuristic factors to map defect severity to path selection probability, and combine this with the pheromone volatilization mechanism to balance global exploration and local development capabilities.

[0128] Fourth, spiral path generation technology is used to focus on the center of mass for progressive polishing in defective areas, while raster path technology is used to achieve efficient traversal in normal areas.

[0129] It should be noted that in this embodiment, through the anisotropy analysis of the target object, the direction of minimum Young's modulus is selected as the main scanning direction to reduce surface stress concentration; it should be noted that in the polishing of the normal area, the grating path is decomposed into sub-segments, and the shortest Hamiltonian path is solved by the genetic algorithm to construct the control path, thereby reducing the invalid movement time.

[0130] Fifth, the defect-priority path is dynamically combined with the global scanning path through a multimodal path fusion algorithm to achieve coordinated optimization of polishing efficiency and accuracy. It should be noted that this embodiment inserts a fifth-order polynomial transition curve at the intersection of the spiral and grating paths to ensure that the acceleration of the transition area corresponding to the defect area and the normal area is continuous;

[0131] Based on the optimal bin control parameter sequence set corresponding to all position points of the current target object surface type and the corresponding defect control path and normal position control path in each cluster area on the target object surface type, the normal position control instruction and defect position control instruction are obtained through the fuzzy control algorithm;

[0132] Perform defect processing on the target object based on the normal position control instruction and the defect position control instruction, and detect and obtain the valley peak value and RMS value corresponding to the target object after defect processing;

[0133] The valley-peak value and RMS value corresponding to the target object after defect processing are fed back to the processing parameter-defect removal rate mapping database and the path algorithm optimized by the ant colony algorithm to perform real-time optimization on the defect position control path and the normal position control path corresponding to each cluster area on the target object surface and the optimal grading control parameter sequence set until the valley-peak value and RMS value corresponding to the target object surface meet the preset processing threshold position.

[0134] This process dynamically matches real-time three-dimensional error distribution maps with a parameter-defect removal rate database to construct an adaptive inversion mechanism for defect characteristics and process parameters. Based on a multidimensional feature vector of defect type, PV value, and gradient amplitude, this method achieves precise mapping of parameter combinations, effectively resolving the local optimal solution trap problem in traditional empirical parameter setting. When using an ant colony algorithm to optimize path planning, the pheromone concentration gradient and the dynamic heuristic factor β are synergistically adjusted. In this embodiment, the β value is dynamically adjusted by the ratio of PV to maximum PV, giving the algorithm the ability to prioritize search areas with high errors. Combining the geometric characteristics of the spiral path in the defective area with the ergodic design of the grating path in the normal area, a spatial resolution path adaptive generation system is formed, significantly improving the coverage efficiency of complex morphological defects. The introduction of a fuzzy control algorithm fuses multidimensional control parameters such as dwell time, pressure, and speed through a membership function. A fuzzy inference rule library is used to achieve nonlinear decoupling and smooth transition of the parameter set, overcoming the lack of adaptability of traditional PID control to nonlinear, time-varying process systems. A dynamic correlation model between process parameters and machining effects is established based on a closed-loop feedback mechanism of PV / RMS values. Through an online data-driven parameter-path joint optimization strategy, the system is equipped with the ability to predict error evolution trends in real time. The setting of the pheromone volatility coefficient ρ balances the efficiency of path exploration and development, ensuring that the path topology structure can quickly converge to the optimal solution during the iteration process. This multimodal collaborative control architecture achieves spatial synchronous matching between the material removal rate model and motion trajectory planning through a three-layer coupling of parameter inversion, path optimization, and fuzzy decision-making, effectively suppressing secondary surface damage caused by parameter mutations during the machining process. At the same time, through submillimeter grid division and defect clustering analysis, the local correction capability of microscale errors is improved, providing technical support for the global consistency control of the surface shape of complex targets.

[0135] Example 2

[0136] See also Figure 2 , another embodiment provided by the present invention: an image modeling and microstructure control system, comprising: a three-dimensional modeling module, an inversion module, a control module and an evaluation module;

[0137] The three-dimensional modeling module obtains a three-dimensional error distribution map of the target object surface based on the initial interference map of the target object to be processed;

[0138] The inversion module obtains a target object surface control parameter space based on the target object surface three-dimensional error distribution map combined with a preset processing parameter-defect removal rate mapping database; the target object surface control parameter space includes each defect location point and type, the length of the defect residence time at each location point, pressure, speed, and control path forms for different defect areas of the target object surface;

[0139] The control module generates ultrasonic hybrid control instructions based on the target object surface control parameter space and fuzzy control algorithm to perform real-time defect treatment on the target object to be processed, and obtains a secondary interference map after the defect treatment;

[0140] The evaluation module obtains an evaluation error value based on the secondary interference pattern after defect processing and the preset target object surface evaluation index, and feeds the evaluation error value back to the processing parameter-defect removal rate mapping database to adjust the target object surface control parameter space in real time until the preset target object surface accuracy requirement is met.

[0141] Example 3

[0142] An electronic device includes a memory and a processor, wherein the memory stores a computer program and the processor implements an image modeling and microstructure recognition method when executing the computer program.

[0143] A computer-readable storage medium stores computer instructions, which, when executed, execute an image modeling and microstructure recognition method.

[0144] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the scope of protection of the purpose of the present invention and the claims, and all of these are protected by the present invention.

Claims

1. An image modeling and microstructure recognition method, characterized in that: include: Based on the initial interference pattern of the target object to be processed, a three-dimensional error distribution map of the target object surface is obtained; Based on the target object surface three-dimensional error distribution map and combined with a preset processing parameter-defect removal rate mapping database, a target object surface control parameter space is obtained by inversion; the target object surface control parameter space includes each defect location point and type, the residence time length of the defect removal at each location point, pressure, speed, and the control path form of different defect areas of the target object surface; Generate ultrasonic hybrid control instructions based on the target object surface control parameter space combined with fuzzy control algorithm to perform real-time defect treatment on the target object to be processed, and obtain a secondary interference map after defect treatment; Based on the secondary interference pattern after defect processing and the preset target object surface evaluation index, an evaluation error value is obtained, and the evaluation error value is fed back to the processing parameter-defect removal rate mapping database to adjust the target object surface control parameter space in real time until the preset target object surface accuracy requirement is met; The target object surface shape evaluation index includes the valley-peak value and RMS value corresponding to the target object surface shape error.

2. The image modeling and microstructure recognition method according to claim 1, wherein: The process of obtaining the target object surface three-dimensional error distribution map includes: Perform an initial scan on the target object using a laser interferometer with dual-wavelength module parameters and initial scanning parameters to obtain an initial surface gradient map; Based on the initial surface gradient map and the Hessian matrix, the local curvature is calculated to obtain the high curvature area and the low curvature area in the initial surface gradient map and label the areas; Based on the labeled high curvature area and low curvature area information, the high curvature point cloud spacing and the low curvature point cloud spacing are set, and based on the set point cloud spacing and the regional curvature type, the labeled high curvature area point cloud map and the low curvature area point cloud map are generated; Based on the labeled point cloud images of high curvature areas and low curvature areas, the Delaunay triangulation algorithm and the spiral path algorithm are combined to search for the scanning path along the direction of maximum curvature to generate a spiral scanning path.

3. The image modeling and microstructure recognition method according to claim 2, characterized in that: The process of acquiring the target object surface three-dimensional error distribution map further includes: Scanning the target object based on a spiral scanning path and updating the surface gradient map every time 10% of the target object's surface area is scanned, and re-dividing the target object into high curvature areas and low curvature areas until the entire target object is scanned, thereby obtaining an initial interference map of the target object's surface; The initial interferogram of the target surface is filtered and processed, and the initial phase space of the interferogram is obtained through the phase shift interferometry algorithm; Based on the initial phase space of the interferogram and the geometric information of the target object, the surface shape of the target object to be processed is iteratively solved by the phase unwrapping algorithm combined with the least squares method. Combined with the three-dimensional coordinate system with the center of gravity of the target object as the origin, the three-dimensional defect distribution map of the target object surface is obtained. Based on the target object's surface three-dimensional defect distribution map and a preset multi-scale defect recognition and positioning model, the defects on the target object's surface three-dimensional distribution map are identified and located, and the defect type identified and located on the target object's surface three-dimensional distribution map, the corresponding position point of each defect, and the gradient distribution and height distribution are obtained; The multi-scale defect recognition and positioning model is constructed and pre-trained in an integrated manner by combining a gradient threshold constructed by combining the surface gradient amplitude with the light source wavelength and a preset defect detection algorithm under different gradient amplitudes.

4. The image modeling and microstructure recognition method according to claim 3, wherein: The process of acquiring the target object surface three-dimensional error distribution map further includes: Based on the defect types identified and located on the three-dimensional distribution map of the target surface, the location points corresponding to each defect, the gradient distribution and the height distribution, combined with the PV value of the target surface and the wavelength of the light source, regional clustering is performed using the DBSCAN clustering algorithm combined with the region growing method to obtain the target surface regional defect distribution map; Based on the three-dimensional regional defect distribution map and the three-dimensional distribution map of the standard target object, a three-dimensional regional error distribution map of the target object surface is obtained.

5. The image modeling and microstructure recognition method according to claim 4, characterized in that: The process of constructing the processing parameter-defect removal rate mapping database includes: Obtaining the ultrasonic processing parameter sequence and the second treatment solution processing parameter sequence and combining them with the preset variable bin values ​​to construct an ideal experimental variable orthogonal table; The ultrasonic processing parameter sequence includes frequency, amplitude, pressure, and dwell time length of each position point; the second processing solution processing parameter sequence includes the second processing solution concentration, flow rate per unit time, and corresponding equivalent pressure; Obtain a target material of the same material as the target object to be processed and select a step-by-step control parameter sequence corresponding to each ideal experiment based on a random algorithm using an orthogonal table of ideal experimental variables; the step-by-step control parameter sequence includes frequency, amplitude, pressure, dwell time length at each position point, and equivalent pressure; The obtained grading control parameter sequence is input into the ultrasonic mixing system to carry out an ideal polishing experiment, and the defect removal rate corresponding to different defect types per unit time under each ideal experiment is obtained in real time.

6. The image modeling and microstructure recognition method according to claim 5, characterized in that: The process of constructing the processing parameter-defect removal rate mapping database further includes: The defect removal rate is equal to the ratio of the defect height removed within the corresponding residence time length to the original defect height and the residence time length; Based on the defect removal rates of different defect types per unit time under each ideal experiment, the optimal bin control parameter sequence with the maximum defect removal rate for each defect type is obtained; Based on the optimal control parameter sequence and corresponding defect removal rate corresponding to each type and height of defects, the processing parameter-defect removal rate mapping function and the corresponding binned control parameter values ​​are obtained through convolution and function fitting; Based on the processing parameter-defect removal rate mapping function and the corresponding grading control parameter values, a processing parameter-defect removal rate mapping database is obtained through a graph database.

7. The image modeling and microstructure recognition method according to claim 6, characterized in that: The process of generating the ultrasonic mixing control instruction includes: Obtain a real-time target object surface three-dimensional regional error distribution map, combine it with the processing parameter-defect removal rate mapping database, and use a matching algorithm to invert and obtain the optimal binning control parameter sequence set corresponding to all position points of the current target object surface; Based on the defect type, position, PV value and gradient amplitude information corresponding to each defect cluster area in the three-dimensional regional error distribution map of the current target object surface, combined with the path algorithm optimized by the ant colony algorithm, the corresponding defect position control path and normal position control path in each cluster area on the target object surface are obtained.

8. The image modeling and microstructure recognition method according to claim 7, wherein: The process of generating the ultrasonic mixing control instruction further includes: Based on the optimal bin control parameter sequence set corresponding to all position points of the current target object surface type and the corresponding defect control path and normal position control path in each cluster area on the target object surface type, the normal position control instruction and defect position control instruction are obtained through the fuzzy control algorithm; Perform defect processing on the target object based on the normal position control instruction and the defect position control instruction, and detect and obtain the valley peak value and RMS value corresponding to the target object after defect processing; The valley-peak value and RMS value corresponding to the target object after defect processing are fed back to the processing parameter-defect removal rate mapping database and the path algorithm optimized by the ant colony algorithm to perform real-time optimization on the defect position control path and the normal position control path corresponding to each cluster area on the target object surface and the optimal grading control parameter sequence set until the valley-peak value and RMS value corresponding to the target object surface meet the preset processing threshold position.

9. An image modeling and microstructure control system, which is implemented based on an image modeling and microstructure recognition method according to any one of claims 1 to 8, characterized in that: include: 3D modeling module, inversion module, control module and evaluation module; The three-dimensional modeling module obtains a three-dimensional error distribution map of the target object surface based on the initial interference map of the target object to be processed; The inversion module obtains a target object surface control parameter space based on the target object surface three-dimensional error distribution map combined with a preset processing parameter-defect removal rate mapping database; the target object surface control parameter space includes each defect location point and type, the length of the defect residence time at each location point, pressure, speed, and control path forms for different defect areas of the target object surface; The control module generates ultrasonic hybrid control instructions based on the target object surface control parameter space and fuzzy control algorithm to perform real-time defect treatment on the target object to be processed, and obtains a secondary interference map after the defect treatment; The evaluation module obtains an evaluation error value based on the secondary interference pattern after defect processing and the preset target object surface evaluation index, and feeds the evaluation error value back to the processing parameter-defect removal rate mapping database to adjust the target object surface control parameter space in real time until the preset target object surface accuracy requirement is met.

Images