Prediction method of tool mark error on polishing component surface based on neural network
Through the neural network-based prediction method, the Resnet model is trained using interferometer to measure and process parameter data, and the complexity and timeliness of the prediction of the surface cutter of the optical element is solved, and the rapid and accurate prediction effect is achieved, which improves engineering friendliness sex.
Patent Information
- Application Number
- CN202411300904.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-18
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-09-18
AI Technical Summary
The existing method for predicting surface cutters of optical components is complex in calculation, has knowledge barriers, and the numerical simulation is not friendly in time, making it difficult to quickly and accurately predict the engineering task of large-aperture mirrors without analytical removal functions.
Using the method of predicting surface knife mark error of polishing elements based on neural network, the initial surface shape residual is measured through an interferometer, the removal function is measured, the processing trajectory is designed, the residence time distribution is calculated, and the processing parameter data set and the knife mark error information data set are established. , build a Resnet neural network for training, and generate a knife mark error prediction model.
Fast and accurate prediction of cutter mark errors on the surface of optical components is achieved, reducing calculation time and easy-to-understandness, and improving process parameter mobility.
Smart Images

Figure CN118821623B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of optical processing surface shape prediction, and in particular to a method for predicting tool mark errors on the surface of a polishing element based on a neural network. Background Art
[0002] The existing methods for predicting tool marks on the surface of optical components have complex calculation theories and certain knowledge barriers, which are not conducive to use by engineers.
[0003] There are four existing tool mark prediction methods, which can be summarized as: prediction based on analysis, experiment, numerical and AI; prediction based on analysis, such as the literature [Analytical and stochastic modeling of surface topography in time-dependent sub-aperture processing], the analytical method is suitable for low-order differentiable tool influence functions and surfaces, and can well approximate the prediction of tool marks. Prediction based on experiments, such as the literature [Development of multi-pitch tool path in computer-controlled optical surfacing processes], the experimental method consumes a lot of materials, and the empirical rules summarized can be used for reference. Numerical prediction, such as the literature [Modelling and simulation of mid-spatial-frequency error generation in CCOS], the numerical method needs to consider the computational feasibility in time and space. Prediction based on AI, such as the literature [Ultrafast laser ablation simulator using deep neural networks], the artificial intelligence-based method obtains the corresponding surface shape residual through extensive process practice, and attempts to establish a mapping relationship between the two with the help of complex mathematical models.
[0004] For engineering tasks involving large-aperture mirrors without analytical removal functions, such as magnetorheological finishing (MRF), and the lack of multiple experiments or large data sets for neural network training, numerical simulation methods are more desirable. [Tool mark prediction on the surface of large-aperture mirrors via magnetorheological finishing] introduces a detailed numerical calculation method, but its timeliness is not friendly to engineering. Therefore, numerical methods can be used to solve the problem of insufficient data sets, and neural network training is used to solve the problem of timeliness. In view of the problems existing in the existing technology, a blind box-style fast and accurate prediction method for tool mark errors on the surface of optical components is urgently needed to be developed. Summary of the invention
[0005] In order to solve the above problems, the present invention provides a method for predicting tool mark errors on the surface of a polishing element based on a neural network.
[0006] The present invention aims to provide a method for predicting the tool mark error on the surface of a polishing element based on a neural network, which specifically comprises the following steps:
[0007] S1. Measuring the surface shape of the workpiece to be processed by an interferometer to obtain a discretized initial surface shape residual on the workpiece to be processed;
[0008] S2. Selecting a processing method for the workpiece to be processed and measuring a removal function;
[0009] S3. Design the machining trajectory according to the size of the workpiece to be machined;
[0010] S4. Calculating the dwell time distribution of the tool according to the initial surface shape residual, the processing trajectory and the removal function;
[0011] S5. taking the initial surface shape residual, the processing trajectory, the removal function and the dwell time as input processing parameter data;
[0012] S6. Calculate the tool mark residual according to the current parameter data or the actual processing to obtain a set of error information data, and use it as the output tool mark error information data;
[0013] S7. Repeat steps S1 to S6 N times to obtain N sets of processing parameter data and error information data, and establish a processing parameter data set and a tool mark error information data set;
[0014] S8. Constructing a Resnet neural network; inputting a processing parameter data set and a tool mark error information data set to train the Resnet neural network to obtain a tool mark error prediction model for the surface of a polished optical element;
[0015] S9. Obtaining initial optical processing parameters of the optical element whose tool mark error is to be predicted, inputting the initial optical processing parameters into a tool mark error prediction model for the surface of the polished optical element, and obtaining the surface tool mark error of the optical element.
[0016] Preferably, the training process in step S8 is as follows:
[0017] S81. Perform matrix concatenation processing on the feature matrices of the initial surface shape residual, the removal function, the processing trajectory and the dwell time; let the feature matrix of the initial surface shape residual, the feature matrix of the removal function, the feature matrix of the processing trajectory and the feature matrix of the dwell time be , , , , let the feature matrix , , , The shapes are [a,b], [c,d], [e,f], [g,h];
[0018] Find the maximum shape of all feature matrices [m,n] where , ; Expand the initial feature matrix by filling the edges with 0 so that its shape becomes [m,n], and then generate a matrix with the shape [m,n] accordingly , , , ;Finally, define the multiple matrix concatenation operation as , get the feature matrix of the input Resnet neural network W : ; The processed input feature matrix W The shape of is [4,m,n];
[0019] S82. Record the actual tool mark error information obtained by calculation or actual processing as a tool mark error matrix ;
[0020] S83. The feature matrix W The input is connected to the tool mark error matrix in the Resnet neural network, and the optimization objective function is set to MAE:
[0021] ;
[0022] In the formula, Num is the total number of samples, is the actual tool mark residual, is the predicted tool mark residual.
[0023] Preferably, the algorithm optimized in step S82 is the Adam algorithm; and Num in step S7>200.
[0024] Preferably, step S1 specifically comprises: using an interferometer to measure the surface shape of the workpiece to be processed, and comparing it with the target surface shape to obtain the initial surface shape residual of the workpiece to be processed, and obtaining discrete points, , M is the number of discrete points on the surface.
[0025] Preferably, step S2 specifically includes the following sub-steps:
[0026] S21. Selecting a test piece of the same material as the workpiece to be processed, and measuring the initial surface shape of the test piece;
[0027] S22. The test piece is polished at a fixed point and time, and the surface shape of the test piece after polishing is measured to obtain the polished surface shape;
[0028] S23. Subtract the initial surface shape from the polished surface shape, and then divide the result by the polishing time to obtain the removal function.
[0029] Preferably, step S3 specifically includes: setting a polishing trajectory according to the surface sampling size of the workpiece to be processed, obtaining a processing trajectory, and obtaining discrete points, , N is the number of discrete points of the trajectory.
[0030] Compared with the prior art, the present invention can achieve the following beneficial effects:
[0031] The present invention proposes a prediction method for tool mark errors on the surface of polished optical elements based on a neural network, takes previous initial optical processing parameters (including initial surface shape error, removal function, polishing trajectory, and dwell time) and surface shape residuals with tool marks as data sets, and uses a neural network to mine the intrinsic relationship between the initial optical processing parameters and the surface shape residuals with tool marks; when a new tool mark prediction task comes, it is only necessary to input the new initial optical processing parameters into the network to quickly obtain the surface shape residuals with tool marks. Compared with the traditional complex formula method, the method proposed by the present invention has obvious advantages in calculation time and ease of understanding, and the process parameter transferability is good. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Figure 1 The present invention is a flowchart of a method for predicting tool mark errors on the surface of a polishing element based on a neural network provided in accordance with an embodiment of the present invention.
[0033] Figure 2 It is a Resnet neural network training flow chart provided according to an embodiment of the present invention.
[0034] Figure 3 It is a structural diagram of a Resnet neural network provided according to an embodiment of the present invention. DETAILED DESCRIPTION
[0035] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, the same modules are represented by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, the detailed description thereof will not be repeated.
[0036] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.
[0037] The present invention provides a method for predicting the tool mark error on the surface of a polishing element based on a neural network (the flow chart is as follows Figure 1 ), specifically including the following steps:
[0038] S1. Measuring the surface shape of the workpiece to be processed by an interferometer to obtain a discretized initial surface shape residual on the workpiece to be processed;
[0039] Specifically, the surface shape of the workpiece to be processed is measured by using an interferometer, and compared with the target surface shape to obtain the initial surface shape residual of the workpiece to be processed, and obtain discrete points, , M is the number of discrete points on the surface.
[0040] The parameters recorded include the type of imaging system, as well as the relative positions, placement angles, surface shapes, low-order surface shapes and substrate materials of all optical components in the imaging system.
[0041] S2. Select a processing method for the workpiece to be processed, including but not limited to small grinding head polishing, air bag polishing, magnetorheological polishing, ion beam polishing, etc., and measure the removal function; specifically including the following sub-steps:
[0042] S21. Select a test piece of the same material as the workpiece to be processed, and measure the initial surface shape of the test piece;
[0043] S22. The test piece is polished at a fixed point and time, and the surface shape of the test piece after polishing is measured to obtain the polished surface shape;
[0044] S23. Subtract the initial surface shape from the polished surface shape, and then divide the result by the polishing time to obtain a removal function.
[0045] S3. Design the machining trajectory according to the size of the workpiece to be machined: Set the polishing trajectory according to the surface sampling size of the workpiece to be machined, and obtain the machining trajectory. discrete points, , N is the number of discrete points of the trajectory.
[0046] S4. Calculate the dwell time distribution of the tool according to the machining trajectory, the initial surface shape residual and the removal function;
[0047] make ;
[0048] (1.1)
[0049] in, Represents discretized surface data points The removal amount at . Represents the stationary points of the trajectory discretization The residence time at . It is the removal rate generated based on the discrete convolution relationship between the surface data points and the residence points, denoted as .
[0050] This equation (1.1) can be simply described as , called the residence time matrix equation, R is the removal rate matrix generated by the surface data point grid and the residence point grid, M is the number of removal points, and N is the number of residence points;
[0051] (1.2)
[0052] (1.3)
[0053] (1.4)
[0054] (1.5)
[0055] (1.6)
[0056] (1.7)
[0057] In the formula, is the objective function, is the kth estimate of the objective function, is the gradient of the function, is to find the partial derivative of the i-th component of the function, is the dwell time vector, is the dwell time vector obtained at the kth iteration, E is the face removal amount vector, is the kth damping coefficient, the present invention sets the initial β is 0.1, is the kth iteration step length, the present invention sets the initial γ is 0.1, is a projection onto the set of positive real numbers, is the i-th component of the vector, is the lower limit constraint of the dwell time solution, which is set to 0.1s in the present invention. is the 2-norm of a matrix or vector, δ is the iterative convergence condition, and the present invention sets δ to 0.01. B is the condition set. For details, please refer to the paper “Algorithm for Solving the Dwell Time of Magnetorheological Processing of Large-Aperture Optical Elements”.
[0058] S5. Take the initial surface shape residual, the processing trajectory, the removal function and the dwell time as input processing parameter data. The data contained in the collected feature matrix include: the size of the initial surface shape residual, the surface shape, the error distribution, the error root mean square (RMS) value, the error peak valley (PV) value, the low-frequency error information, and the medium-frequency error information; the size of the removal function, the peak removal rate, the removal rate distribution, the volume removal rate, the RMS value, and the PV value; the trajectory form, sampling interval, and range of action of the processing trajectory; the distribution of the dwell time, the total dwell time, the RMS value, the PV value, and the uniformity of the dwell time.
[0059] S6. A set of error information data is obtained according to the tool mark residual calculated based on the current parameter data or obtained by actual processing, and used as the output tool mark error information data. The output tool mark error information data includes the error amplitude error, phase error, period error, angle relative to the coordinate system, surface power spectrum density or surface structure function.
[0060] S7. Repeat steps S1 to S6 Num times to obtain Num sets of processing parameter data and error information data, and establish a processing parameter data set and a tool mark error information data set; Num>200.
[0061] S8. Construct a Resnet neural network; input a processing parameter data set and a tool mark error information data set to train the Resnet neural network to obtain a tool mark error prediction model for the surface of a polished optical element; the specific process is as follows Figure 2 As shown; the training process is as follows:
[0062] S81. Perform matrix concatenation processing on the feature matrices of the initial surface shape residual, the removal function, the processing trajectory and the dwell time; let the feature matrix of the initial surface shape residual, the feature matrix of the removal function, the feature matrix of the processing trajectory and the feature matrix of the dwell time be , , , , let the feature matrix , , , The shapes are [a,b], [c,d], [e,f], [g,h] respectively; find the maximum shape of all feature matrices [m,n], where , ;
[0063] The initial feature matrix is expanded by padding the edges with 0 so that its shape becomes [m,n]. For example,
[0064]
[0065]
[0066] Then correspondingly, a matrix of shape [m,n] is generated , , , ;
[0067] Finally, define the multiple matrix concatenation operation as , get the feature matrix of the input Resnet neural network W :
[0068] ;
[0069] At this time, the processed input feature matrix W The shape is [4,m,n].
[0070] S82. Record the actual tool mark error information obtained by calculation or actual processing as a tool mark error matrix ;
[0071] Among them, the tool mark error information can be calculated using the method in the paper [Tool mark prediction on the surface of large-aperture mirrors via magnetorheological finishing].
[0072] S83. The feature matrix W The input is connected to the tool mark error matrix in the Resnet neural network, and the optimization objective function is set to MAE:
[0073] ;
[0074] In the formula, Num is the total number of samples, is the actual tool mark residual, is the predicted tool mark residual; the optimization algorithm used is the Adam algorithm.
[0075] The structure of the Resnet neural network is as follows Figure 3As shown in the figure, considering that the feature matrix is larger after splicing, a deeper neural network is needed to train the data in the actual training process. Since the general DNN network will degenerate and experience gradient explosion and gradient loss when the number of network layers is large, the present invention adopts the Resnet neural network structure for training. Figure 3 As shown in the figure, the difference from the traditional DNN structure is that the Resnet network adds residual connections, adding the features of the previous layer to the features processed by the current layer. Even if the network in this layer has not learned anything, it can ensure that the network will not degenerate due to the increase in the number of layers. At the same time, saving the feature output of the previous layer can effectively avoid the problem of gradient disappearance / explosion. In the specific implementation case, n samples are divided into training set, test set and validation set according to 8:1:1. The initial parameters of the neural network are set as follows: the step size is set to 0.00002, the number of iterations epoch is set to 100,000 times, the weight decay rate is set to 0.5, and the weight of the optimal validation set is saved after each iteration.
[0076] S9. Obtaining initial optical processing parameters of the optical element whose tool mark error is to be predicted, inputting the initial optical processing parameters into a tool mark error prediction model for the surface of the polished optical element, and obtaining the surface tool mark error of the optical element;
[0077] The initial parameters of optical processing include initial surface shape residual, removal function, processing trajectory and dwell time; the method for obtaining the initial parameters of optical processing is: repeating steps S1 to S3 to obtain the initial surface shape residual, removal function, processing trajectory and dwell time of the optical element whose tool mark error is to be predicted.
[0078] Although the amount of data collected in the early stage of this method is relatively large, once the neural network calculation model is built, the prediction of the tool mark error on the surface of the polished optical element is very fast. Therefore, the neural network-based tool mark error prediction method for the surface of the polished optical element proposed in the present invention has obvious advantages in terms of calculation time consumption and ease of understanding.
[0079] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the disclosure of the present invention can be performed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.
[0080] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A prediction method for tool mark error on polishing component surface based on neural network, It is characterized in that The specific steps include: S1. Measuring the surface shape of the workpiece to be processed by an interferometer to obtain a discretized initial surface shape residual on the workpiece to be processed; S2. Selecting a processing method for the workpiece to be processed and measuring a removal function; S3. Design the machining trajectory according to the size of the workpiece to be machined; S4. Calculating the dwell time distribution of the tool according to the initial surface shape residual, the processing trajectory and the removal function; S5. taking the initial surface shape residual, the processing trajectory, the removal function and the dwell time as input processing parameter data; S6. Calculate the tool mark residual according to the current parameter data or the actual processing to obtain a set of error information data, and use it as the output tool mark error information data; S7. Repeat steps S1 to S6 N times to obtain N sets of processing parameter data and error information data, and establish a processing parameter data set and a tool mark error information data set; S8. Constructing a Resnet neural network; inputting a processing parameter data set and a tool mark error information data set to train the Resnet neural network to obtain a tool mark error prediction model for the surface of a polished optical element; S9. Obtaining initial optical processing parameters of the optical element whose tool mark error is to be predicted, inputting the initial optical processing parameters into a tool mark error prediction model for the surface of the polished optical element, and obtaining the surface tool mark error of the optical element.
2. According to claim 1, a method for predicting tool mark errors on the surface of a polishing element based on a neural network, It is characterized in that The training process in step S8 is specifically as follows: S81. Perform matrix concatenation processing on the feature matrices of the initial surface shape residual, the removal function, the processing trajectory and the dwell time; let the feature matrix of the initial surface shape residual, the feature matrix of the removal function, the feature matrix of the processing trajectory and the feature matrix of the dwell time be , , , , let the feature matrix , , , The shapes are [a,b], [c,d], [e,f], [g,h]; Find the maximum shape of all feature matrices [m,n] where , ; Expand the initial feature matrix by filling the edges with 0 so that its shape becomes [m,n], and then generate a matrix with the shape [m,n] accordingly , , , ;Finally, define the multiple matrix concatenation operation as , get the feature matrix of the input Resnet neural network W : ; The processed input feature matrix W The shape of is [4,m,n]; S82. Record the actual tool mark error information obtained by calculation or actual processing as a tool mark error matrix ; S83. The feature matrix W The input is connected to the tool mark error matrix in the Resnet neural network, and the optimization objective function is set to MAE: ; In the formula, Num is the total number of samples, is the actual tool mark residual, is the predicted tool mark residual.
3. The method for predicting the tool mark error on the surface of a polishing element based on a neural network according to claim 2, Features: The algorithm optimized in step S82 is the Adam algorithm; N in step S7>200.
4. The method for predicting the tool mark error on the surface of a polishing element based on a neural network according to claim 3, It is characterized in that The step S1 specifically includes: using an interferometer to measure the surface shape of the workpiece to be processed, and comparing it with the target surface shape to obtain the initial surface shape residual of the workpiece to be processed, and obtaining discrete points, , M is the number of discrete points on the surface.
5. The method for predicting the tool mark error on the surface of a polishing element based on a neural network according to claim 4, It is characterized in that The step S2 specifically includes the following sub-steps: S21. Selecting a test piece of the same material as the workpiece to be processed, and measuring the initial surface shape of the test piece; S22. The experimental piece is polished at a fixed point and time, and the surface shape of the experimental piece after polishing is measured to obtain the polished surface shape; S23. Subtract the initial surface shape from the polished surface shape, and then divide the result by the polishing time to obtain the removal function.
6. The method for predicting the tool mark error on the surface of a polishing element based on a neural network according to claim 5, It is characterized in that The step S3 specifically includes: setting a polishing trajectory according to the surface sampling size of the workpiece to be processed, obtaining a processing trajectory, and obtaining discrete points, , N is the number of discrete points of the trajectory.
