Multi-axis polishing batch uniformity control apparatus, method, device, and storage medium

Abstract

The application relates to the technical field of intelligent machining control, and discloses a multi-axis polishing batch consistency control device, method, equipment and storage medium.The device comprises a matrix construction module, which is used for collecting multi-axis parallel polishing historical data and constructing a collaborative feature matrix containing coaxial historical states and different-axis real-time states; a main prediction module, which is used for constructing a deep neural network model, taking the collaborative feature matrix as input, learning a nonlinear mapping relationship between process parameters and polishing removal amounts, and outputting main prediction values of polishing removal amounts of each point of each workpiece; a residual correction module, which is used for constructing a Gaussian process regression model, modeling and predicting the residual of the main prediction values, and outputting residual prediction values; and a fusion module, which is used for adding the main prediction values and the residual prediction values to obtain final polishing removal amount prediction values.Through simultaneously incorporating single-axis cumulative effects and multi-axis collaborative states into modeling, the batch consistency problem caused by the fact that traditional methods ignore the differences between inter-axis coupling and tool wear is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent machining control technology, specifically to a multi-axis polishing batch consistency control device, method, equipment, and storage medium. Background Technology

[0002] In existing technologies, multi-axis parallel polishing is often used to improve polishing efficiency, where multiple grinding wheels simultaneously process multiple workpieces. The core requirement of this process is uniform material removal at all points within a single workpiece, and highly consistent material removal rates across multiple workpieces. However, due to differences in grinding wheel wear, drift in multi-axis dynamic characteristics, and the complex nonlinear relationship between pressure-number of revolutions and material removal rate, it is difficult for the removal rates of multiple workpieces and multiple points on a single workpiece to simultaneously reach the target value, a common challenge in the industry.

[0003] Currently, industrial sites rely heavily on manual "trial and error" to adjust parameters. This method is highly dependent on the operator's experience, has low efficiency and poor stability, and is difficult to cope with the nonlinear characteristics of complex curved surfaces with multi-parameter coupling.

[0004] To overcome the limitations of empirical methods, some studies have employed statistical techniques such as response surface methodology and Taguchi methods to establish the mapping relationship between process parameters and removal amounts. However, these polynomial models have limited expressive power, making it difficult to accurately describe the nonlinear dynamic characteristics of multiaxial polishing, and they are also sensitive to noise and have weak generalization ability.

[0005] In recent years, machine learning algorithms such as neural networks have been introduced into polishing process modeling, and their nonlinear fitting ability is superior to statistical methods. However, existing methods still have obvious shortcomings: single models are prone to overfitting or underfitting, and prediction uncertainty is difficult to quantify; most only establish positive prediction models and lack the ability to solve inversely from quality requirements to parameter settings; they are sensitive to small samples and noise, and cannot effectively quantify the impact of uncertainties such as tool wear and material fluctuations.

[0006] Therefore, there is an urgent need for an intelligent process optimization scheme that can accurately predict the amount of material removed by multi-axis parallel polishing and achieve batch consistency control. Summary of the Invention

[0007] This invention provides a batch consistency control device, method, equipment, and storage medium for multi-axis polishing, to solve the problem of difficulty in ensuring the consistency of batch removal amount in multi-axis parallel polishing.

[0008] In a first aspect, the present invention provides a multi-axis polishing batch consistency control device, the device comprising: The matrix construction module is used to collect historical data of multi-axis parallel polishing and construct a collaborative feature matrix that includes the historical state of the same axis and the real-time state of the opposite axis. The main prediction module is used to construct a deep neural network model, which takes the collaborative feature matrix as input, learns the nonlinear mapping relationship between process parameters and polishing removal amount, and outputs the main predicted value of polishing removal amount at each point of each workpiece. The residual correction module is used to construct a Gaussian process regression model, model and predict the residuals of the master predicted value, and output the residual predicted value. The fusion module is used to add the master prediction value and the residual prediction value to obtain the final polishing removal amount prediction value.

[0009] In one optional embodiment, the device further includes: The optimization module is used to construct a composite loss function and solve for the globally optimal process parameters through constrained Bayesian optimization.

[0010] In one optional implementation, the main prediction module includes: A multi-branch feature extraction network includes at least one branch for extracting spatial distribution features of process parameters and at least one branch for extracting temporal dependency features of process parameters; The feature fusion unit is used to fuse the features output by the multi-branch feature extraction network. The regression mapping unit is used to map the fused features to the master predicted value of the polishing removal amount.

[0011] In one optional implementation, the residual correction module includes: The residual calculation unit is used to obtain the prediction value of the main prediction module on the training set, compare it with the actual removal amount, and calculate the prediction residual. The kernel function construction and fusion unit is used to construct an integrated kernel containing at least two dedicated kernel functions. The dedicated kernel functions are used to quantify the residuals caused by pressure direction mismatch and the residuals caused by multi-axis coordination mismatch, respectively. The weights of each kernel function are adaptively determined through cross-validation to obtain a weighted integrated kernel function. The model training unit is used to construct a Gaussian process regression model with the predicted residual as the training target and the weighted ensemble kernel function as the covariance function, and to train the Gaussian process regression model through hyperparameter optimization. The residual prediction unit is used to input the collaborative feature matrix into the trained Gaussian process regression model and output the residual prediction value.

[0012] In one optional implementation, the kernel function construction and fusion unit includes: The first kernel function construction sub-unit is used to construct the first dedicated kernel function. The first dedicated kernel function is used to perform nonlinear transformation on the pressure direction angle to quantify the residual caused by the mismatch between the pressure direction and the surface normal. The second kernel function construction subunit is used to construct the second dedicated kernel function. The second dedicated kernel function is used to calculate the difference in process state vectors between different machining axes and combine the batch machining information to quantify the residuals caused by multi-axis coordination misalignment. The kernel function weighting subunit is used to evaluate the performance of each kernel function through cross-validation, and to adaptively weight and combine each kernel function based on the performance score to output a weighted ensemble kernel function.

[0013] In one optional implementation, the optimization module includes: The target setting unit is used to obtain the target polishing removal amount; The loss function construction unit is used to construct a composite loss function, which includes an accuracy loss term, an inter-workpiece consistency loss term, an intra-workpiece uniformity loss term, and a processing efficiency loss term. The parameter optimization unit is used to perform iterative search using a positive prediction model as the evaluator and the composite loss function as the optimization objective, employing a constrained optimization algorithm to output the globally optimal combination of process parameters that minimizes the composite loss function.

[0014] In one optional embodiment, the device further includes: The closed-loop compensation module is used to apply the globally optimal process parameter combination to the actual processing, and to update the main prediction module, the residual correction module, the fusion module and the optimization module according to the deviation between the actual polishing removal amount and the target polishing removal amount.

[0015] Secondly, the present invention provides a method for batch consistency control in multi-axis polishing, the method comprising: Collect historical data of multi-axis parallel polishing and construct a collaborative feature matrix that includes the historical state of the coaxial axis and the real-time state of the heteroaxial axis. A deep neural network model is constructed, and the nonlinear mapping relationship between process parameters and polishing removal amount is learned by taking the collaborative feature matrix as input. The main predicted value of polishing removal amount at each point of each workpiece is output. Construct a Gaussian process regression model to model and predict the residuals of the main predicted values, and output the predicted residual values; The master prediction value is added to the residual prediction value to obtain the final polishing removal amount prediction value.

[0016] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the multi-axis polishing batch consistency control method described in the second aspect or any corresponding embodiment thereof.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the multi-axis polishing batch consistency control method described in the second aspect or any corresponding embodiment thereof.

[0018] This application, by constructing a collaborative feature matrix that includes both coaxial historical states and interaxial real-time states, is the first to simultaneously incorporate single-axis cumulative effects and multi-axis collaborative states into modeling, fundamentally solving the batch consistency problem caused by neglecting inter-axis coupling and tool wear differences in traditional methods. A deep neural network master predictor learns the nonlinear mapping between process parameters and removal amounts, accurately capturing the main trend; a Gaussian process regression model models and compensates for residuals, significantly improving prediction accuracy and reliability. The fusion of master predictions and residual predictions achieves precise forward prediction. This scheme obtains high-precision polishing removal amount predictions without relying on manual trial and error or complex inverse solutions, providing a reliable foundation for subsequent batch consistency control, effectively reducing fluctuations in removal amounts between and within workpieces during multi-axis parallel polishing, and improving product yield and processing efficiency. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a first schematic diagram of a multi-axis polishing batch consistency control device according to an embodiment of the present invention; Figure 2 This is a second schematic diagram of a multi-axis polishing batch consistency control device according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the residual Gaussian forward prediction and Bayesian inverse optimization model architecture according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating a multi-axis polishing batch consistency control method according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0023] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] This application provides a multi-axis polishing batch consistency control device, which constructs a multi-dimensional collaborative feature system oriented towards "axis-workpiece-bar-area-point". Through the collaborative work of matrix construction module 1, main prediction module 2, residual correction module 3, fusion module 4, optimization module 5, and closed-loop compensation module 6, an intelligent control closed loop with batch consistency as the goal is formed. The following is in conjunction with the appendix... Figure 1 The specific implementation methods of this application will be described in detail. It should be noted that the "multi-axis polishing batch consistency control device" described in this embodiment is applicable to scenarios where any number of machining spindles are polished in parallel. For ease of description, the following uses five-axis parallel polishing (i.e., simultaneously clamping five grinding wheels to process five workpieces) as a specific embodiment, where each workpiece contains 99 polishing points. It is understood that the technical solution of this application is also applicable to four-axis, six-axis, or other numbers of multi-axis polishing equipment, and its implementation principle is completely consistent with the five-axis embodiment.

[0025] like Figure 1 As shown, the multi-axis polishing batch consistency control device includes the following functional modules: Matrix construction module 1 is used to collect historical data of multi-axis parallel polishing and construct a collaborative feature matrix that includes historical states on the same axis and real-time states on different axes. Specifically, its implementation is as follows: First, historical data on five-axis parallel polishing was collected. This historical data includes the Y-axis pressure of each machining spindle. Z-axis pressure , number of positive revolutions (Number of clockwise rotations), Number of counterclockwise rotations (Number of counter-clockwise machining revolutions). It should be noted that five-axis polishing processes typically perform forward polishing first, followed by reverse polishing, to achieve uniform removal. In addition, historical data includes the grinding wheel number (1-4), machining area code (arc surface 1-6, bottom edge, bottom center, R-angle), feed mode (straight edge or corner), and the actual polishing removal amount at each point.

[0026] Then, physical derivation calculations are performed on the original process parameters: Calculate the resultant force ; Calculate the pressure direction angle ; Calculate the total number of laps .

[0027] Next, the data from the five machining spindles are correlated. For ease of description, when focusing on a particular machining spindle (e.g., axis 1), that axis is defined as "coaxial," and the other four spindles (axis 2, 3, 4, and 5) are defined as "discordant axes." The same applies to the other axes.

[0028] It should be noted that in five-axis parallel polishing, all process parameters (such as pressure, feed rate, and number of revolutions) of the five polishing rods are changed simultaneously. For example, when one polishing rod is pressurized, it is necessary to simultaneously consider whether the other workpieces have reached the critical pressure value. If the critical value has been reached, other parameters (such as feed rate and number of revolutions) need to be changed to adjust the final removal amount. The ultimate goal is to ensure that the removal amount of all five processed products simultaneously meets the tolerance requirements.

[0029] Based on the above process characteristics, matrix construction module 1 generates the following two types of collaborative feature vectors: (1) Coaxial historical state vector For the current machining point, extract the process state statistics of several consecutive points preceding the same spindle (e.g., the first 3 points), including: mean resultant force, mean pressure direction angle, mean total number of revolutions, and resultant force variation trend (e.g., first-order difference). This vector is used to characterize the cumulative machining state of the same grinding wheel bar, reflecting progressive effects such as tool wear and thermo-coupling, enabling the model to perceive the wear degree of a single axis and achieve personalized compensation.

[0030] (2) Real-time state vectors on different axes For the current machining moment, extract the process status statistics of other machining spindles (i.e., different axes) in the same batch at the current point, including: mean resultant force, standard deviation of resultant force, and mean total number of revolutions. This vector is used to characterize the cooperative and competitive state of multi-axis parallel machining. Among them, the mean resultant force is used to monitor the overall system load and prevent single-axis overload; the standard deviation of resultant force serves as a leading indicator of batch consistency, and its increase warns of load imbalance among axes. The introduction of this vector ensures that parameters that synchronize the process status of each axis can be recommended during reverse optimization, thereby achieving the batch consistency goal of simultaneously meeting the tolerance requirements for the removal amount of five workpieces.

[0031] Finally, matrix construction module 1 standardizes all the original data and derived features and cleans outliers, then divides them into training and test sets in a 7:3 ratio. This module outputs a fully processed collaborative feature matrix for use by the main prediction module 2.

[0032] The main prediction module 2 is used to construct a deep neural network model. Taking the collaborative feature matrix as input, it learns the nonlinear mapping relationship between process parameters and polishing removal amount, and outputs the main predicted value of polishing removal amount for each point on each workpiece. Specifically, main prediction module 2 is used to construct a positive prediction model with the global optimization objective of batch workpiece removal amount consistency. Specifically, this module implements a deep neural network main predictor oriented towards multi-dimensional collaboration between "axis-workpiece-bar-region-point," with its input features deeply coupled to the process context, and its output being the main predicted value of polishing removal amount for 99 points on each of the five workpieces.

[0033] The input feature matrix of the main prediction module 2 is provided by the matrix construction module 1, and this matrix already contains the physical decomposition and combination characterization of key process parameters: Key features: Not only includes the original Y-axis pressure Z-axis pressure The calculation included the number of forward and reverse polishing passes, as well as derived features such as the resultant force (characterizing the total mechanical strength of the contact between the grinding wheel and the curved surface), the pressure direction angle (characterizing the distribution relationship between normal and tangential forces), and the total number of polishing passes (the sum of the number of forward and reverse polishing passes, depicting the asymmetric cumulative effect of different polishing path strategies on material removal). The total number of polishing passes was the sum of the number of forward and reverse polishing passes. When constructing the collaborative feature matrix, the number of forward and reverse polishing passes were used as independent feature inputs, while the total number of polishing passes was also used as a feature input to depict the asymmetric cumulative effect of different polishing path strategies on material removal.

[0034] Process context features: grinding wheel bar number (1-4), machining area code (arc surface 1-6, bottom edge, bottom center, R-angle), feed mode (straight edge or corner). These features are used to distinguish the material removal patterns under different geometric constraints.

[0035] Cooperative status characteristics: Real-time process status of the other four machining spindles (different axes) in the same batch, such as average pressure and average number of revolutions. The introduction of this feature enables the model to perceive the cooperative and competitive relationships in multi-axis parallel machining, laying the foundation for subsequent batch consistency control.

[0036] The main prediction module 2 employs a multi-stream parallel neural network architecture. It includes a dedicated convolutional neural network branch to learn the local sensitivity and saturation characteristics of the "pressure-revolution" process spectrum's influence on the removal amount in different processing areas (such as straight edges and rounded corners). A gated recurrent unit network branch is also included to model the cumulative impact of temporal adjustment strategies for pressure and revolution parameters on the removal amount distribution along the processing path. This model directly outputs the predicted polishing removal amount at ninety-nine points across five workpieces.

[0037] After the above network architecture is constructed, the main prediction module 2 is trained. The main prediction module 2 is trained using supervised learning. During training, the Adam optimizer is used, with mean squared error as the loss function, and the network parameters are updated through backpropagation. After training, this module can quickly output the corresponding main prediction value based on the input collaborative feature matrix, providing a foundation for subsequent residual correction and inverse optimization.

[0038] Residual correction module 3 is used to construct a Gaussian process regression model, model and predict the residuals of the master predicted value, and output the predicted residual values. Specifically, this module first calculates the predicted residuals between the master predicted value and the actual removal amount. Then, a hierarchical adaptive multi-kernel ensemble system is constructed, in which two special kernel functions are designed: an anisotropic RBF kernel to quantify the residuals caused by the mismatch between the pressure direction and the surface normal, and a consistency deviation kernel to quantify the residuals caused by multi-axis coordination misalignment. In addition, the ensemble system also includes several basic kernel functions. The performance of each kernel function is evaluated through cross-validation, and the kernel functions are automatically weighted and combined based on the performance scores to obtain a weighted ensemble kernel function. Using this weighted ensemble kernel as the covariance function and the predicted residuals as the training objective, the Gaussian process regression model is trained. After training, the co-feature matrix is ​​input into the model, and the predicted residual values ​​are output.

[0039] Fusion module 4 is used to add the master predicted value and the residual predicted value to obtain the final predicted value of polishing removal amount. Specifically, this module adds the master predicted value and the residual predicted value point by point. That is, for each point on each workpiece, the master predicted value of the deep neural network is added to the residual value predicted by the Gaussian process regression model, and the sum is used as the final predicted value of polishing removal amount for that point. This fusion operation can be expressed as:

[0040] in, This is the predicted value for the final polishing removal amount. Main forecast value, These are the residual prediction values. The fused prediction values ​​include both main trend information and residual fluctuation information, resulting in higher prediction accuracy and reliability.

[0041] After obtaining the final predicted polishing removal amount, the fusion module 4 evaluates the performance of the forward prediction model on a pre-defined test set, calculating metrics such as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) to verify the model's prediction accuracy for unknown data. These metrics quantify the model's prediction bias and goodness of fit, providing a reliable model foundation for the subsequent inverse optimization module.

[0042] In one optional implementation, the device further includes the following functional modules: Optimization module 5 is used to construct a composite loss function and solve for the globally optimal process parameters through constrained Bayesian optimization. Specifically, this module first obtains the user-defined target removal amount and constructs a composite loss function containing multiple competing loss terms. Then, using the constrained Bayesian optimization framework, it uses the positive prediction model output by fusion module 4 as the objective function evaluator and leverages the constraint expectation to improve the acquisition function and guide the parameter search process. Several sets of process parameter combinations are randomly initialized in the parameter space, and the Gaussian process surrogate model is iteratively updated until a preset stopping condition is met (e.g., the change in the objective function value is less than a threshold after multiple consecutive iterations or the maximum number of iterations is reached). Finally, the globally optimal process parameter combination that minimizes the composite loss function is output.

[0043] In one alternative implementation, such as Figure 2 As shown, the main prediction module 2 includes the following functional units: The multi-branch feature extraction network 21 includes at least one branch for extracting spatial distribution features of process parameters and at least one branch for extracting temporal dependency features of process parameters. Specifically, the multi-branch feature extraction network 21 adopts a multi-stream parallel architecture to extract multi-dimensional features from the collaborative feature matrix. In this embodiment, the network specifically includes four parallel branches: The first CNN (Convolutional Neural Network) branch uses 16 convolutional kernels of size 2 with the tanh activation function to extract local sensitivity features of processing parameters. Local sensitivity refers to the instantaneous impact of small changes in process parameters (pressure, number of revolutions) on the removal amount at a single point. Mathematically, it is the partial derivative of the removal amount with respect to the process parameters, or the local gradient. The learning objective of this branch is to achieve regional differentiated control (fine-tuning in highly sensitive areas and appropriately increasing parameters in less sensitive areas) and to provide gradient information for inverse optimization.

[0044] The second CNN branch uses 16 convolutional kernels of size 3 with the tanh activation function to capture cross-regional differences and saturation characteristics of processing parameters. Cross-regional differences refer to the systematic differences in the impact of the same process parameter on the removal amount in different geometric feature regions (curved surfaces, bottom surfaces, R-angles) (e.g., R-angles are pressure-sensitive, bottom surfaces are sensitive to the number of turns). The size 3 convolutional kernels can simultaneously perceive region encoding, process parameters, and characteristics of adjacent regions, achieving adaptive region prediction and ensuring smooth transitions at region boundaries. Saturation characteristics refer to the nonlinear phenomenon where the removal amount increases at a certain pressure level. Learning this characteristic aims to avoid ineffective pressurization, identify "ineffective pressurization zones," guide optimization towards other parameters, optimize efficiency gains, and ensure that the pursuit of high pressure does not lead to saturation.

[0045] The GRU (Gated Recurrent Unit) branch contains 8 gated recurrent units with a dropout rate of 0.1. It is used to learn point sequence dependencies, that is, the short-term temporal effects between points along the same workpiece along the machining path.

[0046] The LSTM (Long Short-Term Memory) branch contains 8 long short-term memory units with a dropout rate of 0.1, used to learn the process transfer characteristics between regions. Process transfer refers to the smooth transition and evolution of process rules as the grinding wheel moves between different regions, including knowledge transfer, cumulative effect transmission, and policy continuation. Utilizing the long short-term memory capability of LSTM, the long-term impact of the processing state of preceding regions on subsequent regions can be remembered, achieving consistent prediction across the entire path and knowledge sharing between regions.

[0047] The feature fusion unit 22 is used to fuse the features output by the multi-branch feature extraction network. Specifically, the feature fusion unit 22 concatenates the feature vectors output by the first CNN branch, the second CNN branch, the GRU branch, and the LSTM branch in the multi-branch feature extraction network 21 along the feature dimension to form a fused feature vector.

[0048] The regression mapping unit 23 is used to map the fused features into the main predicted values ​​of polishing removal amount. Specifically, the regression mapping unit 23 receives the fused feature vector output by the feature fusion unit 22, passes it sequentially through an LSTM layer containing 3 units, two GRU layers containing 3 units and 16 units respectively, and finally maps it into the main predicted values ​​of polishing removal amount via a fully connected layer. In this embodiment, the output dimension is 5×99, corresponding to the predicted removal amount values ​​for 99 points on each of the five workpieces. .

[0049] After determining the network structure of the regression mapping unit, the main prediction module 2 is trained. The main prediction module 2 is trained using supervised learning. The Adam optimizer is used, with mean squared error as the loss function, a learning rate of 0.001, a batch size of 32, and 200 training epochs. After training, this module can output the main prediction value based on the input collaborative feature matrix, providing a foundation for subsequent residual correction and inverse optimization.

[0050] This application constructs a multi-dimensional predictor of "axis-workpiece-bar-region-point". For the first time, it uses the grinding wheel bar number and wear state, the geometric features of the processing area (arc 1-6, bottom surface, R angle), the feed mode (straight edge / corner), and the real-time status of other axes in the same batch as collaborative inputs. It combines multi-scale CNN and recurrent neural network to decouple and predict the complex parameter coupling effects in multi-axis parallel polishing. It fundamentally solves the problem that traditional single models are insufficient in modeling the complex nonlinear relationship between the "pressure-number of revolutions" process spectrum and the material removal rate.

[0051] In one alternative implementation, the residual correction module 3 employs a physically interpretable hierarchical integrated kernel Gaussian process to directly correlate process parameters with residual sources: 1) using a kernel function with periodic components to model the periodic fluctuation residuals of the removal amount caused by the "forward-reverse" cycle; 2) using an anisotropic kernel related to the pressure direction angle to characterize the systematic residuals caused by the mismatch between the pressure direction and the surface normal in different regions (such as the arc surface and the bottom surface); 3) designing a consistency deviation kernel to quantify the cross-workpiece consistency residuals caused by the mismatch of the "pressure-cycle" parameter sets of each axis in multi-axis parallel machining. Further, such as... Figure 2 As shown, the residual correction module 3 includes the following functional units: The residual calculation unit 31 is used to obtain the predicted value of the main prediction module on the training set and compare it with the actual removal amount to calculate the prediction residual. Specifically, for each sample in the training set, the difference between the main prediction value of the deep neural network and the actual removal amount is calculated to obtain the prediction residual, i.e. ,in, To predict residuals, This represents the actual amount removed. The primary forecast value.

[0052] The kernel function construction and fusion unit 32 is used to construct an integrated kernel containing at least two dedicated kernel functions. These dedicated kernel functions are used to quantify the residuals caused by pressure direction mismatch and the residuals caused by multi-axis coordination misalignment, respectively. The weights of each kernel function are adaptively determined through cross-validation to obtain a weighted integrated kernel function. Specifically, this unit further includes: The first kernel function construction subunit 321 is used to construct a first dedicated kernel function. This dedicated kernel function performs a nonlinear transformation on the pressure direction angle to quantify the residual caused by the mismatch between the pressure direction and the surface normal. In this embodiment, the first dedicated kernel function is an anisotropic RBF kernel (Radial Basis Function). Based on the physical fact that the decomposition of pressure truly affects the removal amount in the polishing process, the pressure direction angle... Using transformation This is to characterize the physical property of "sensitive to changes in tangential force at small angles and approaching saturation at large angles". The anisotropic RBF core takes the form of... ,in and These are feature vectors of points under two different processing conditions. It is the signal variance of the anisotropic RBF core, used to control the overall amplitude of the output value. and It is the eigenvector of the eigenvector. Values ​​for each dimension (excluding the pressure direction angle and other process parameters). It is the first The length scale of each feature dimension is used to control the sensitivity of process parameter values ​​other than the pressure direction angle. It is a special length scale for the pressure direction angle dimension, used to control the sensitivity to changes in pressure direction. This kernel function reflects that the residual similarity between two processing states is determined by the differences in conventional processing parameters and the differences in special treatment of pressure direction.

[0053] The second kernel function construction subunit 322 is used to construct a second dedicated kernel function. This second dedicated kernel function is used to calculate the difference in process state vectors between different machining axes and, in conjunction with batch machining information, to quantify the residuals caused by multi-axis coordination misalignment. In this embodiment, the second dedicated kernel function is a consistency deviation kernel. This kernel function acts directly on the machining axis rather than on the point, and its form is as follows: ,in It is the signal variance of the consistency kernel, used to control the overall fluctuation level of the consistency residuals. For the first Axis and the first The Euclidean distance between the shaft process state vectors. It is the length scale of the consistency kernel, reflecting the sensitivity of inter-axis synchronization. This is an indicator variable for processing in the same batch (1 for the same batch, 0 otherwise). By calculating the differences in process conditions between axes, this kernel function can quantify the degree of multi-axis coordination misalignment and predict the resulting systematic residual correlation between workpieces.

[0054] The kernel function weighting subunit 323 is used to evaluate the performance of each kernel function through cross-validation and adaptively weight and combine the kernel functions based on the performance scores to output a weighted integrated kernel function. In this embodiment, this subunit first constructs an integrated system containing multiple basic kernels and the two special-purpose kernels mentioned above. The basic kernels include the RBF kernel (length scale of 1.0), the Matern kernel (nu=1.5, length scale of 1.0), the rational quadratic kernel (length scale of 1.0, α=1.0), and the exponential sine square kernel (length scale of 1.0, periodicity=1.0); the special-purpose kernels are the aforementioned anisotropic RBF kernel and the consistency deviation kernel.

[0055] Then, the performance of each kernel function was evaluated using 5-fold cross-validation, with negative log-likelihood (NLL) as the evaluation metric. The performance scores of each kernel function were converted into normalized weights using the softmax function.

[0056] in, For the first i Normalized weights of each kernel function, For the first i Performance scores for each kernel function Assign a performance score to each kernel function. Finally, construct the weighted integrated kernel:

[0057] in, For the first i A basic core or a dedicated core, The kernel is a white noise kernel with a noise level of 0.1. This weighted ensemble kernel will be used for training the subsequent Gaussian process regression model.

[0058] Model training unit 33 is used to construct a Gaussian process regression model with the predicted residuals as the training objective and a weighted ensemble kernel function as the covariance function, and to train the Gaussian process regression model through hyperparameter optimization. Specifically, this unit uses the weighted ensemble kernel function as the covariance function, defines the Gaussian process prior distribution, and uses the predicted residuals on the training set as observations to optimize the model hyperparameters by maximizing the marginal log-likelihood or minimizing the negative log-likelihood. The hyperparameters include, but are not limited to, the length scale of each kernel function, signal variance, and noise level. After optimization, a trained Gaussian process regression model is obtained, capable of predicting the residuals at any input point.

[0059] The residual prediction unit 34 is used to input the collaborative feature matrix into the trained Gaussian process regression model and output residual prediction values. Specifically, the residual prediction unit 34 inputs the collaborative feature matrix output by the matrix construction module 1 into the Gaussian process regression model trained by the model training unit 33, and calculates the prediction mean and prediction variance using the posterior distribution of the Gaussian process. The prediction mean is output as the residual prediction value for use by the fusion module 4; the prediction variance can be used to quantify the uncertainty of the prediction. In this embodiment, the residual prediction unit 34 outputs a residual prediction value matrix (e.g., 5×99 dimensions) corresponding to the output dimension of the main prediction module 2, which is then added to the main prediction value point by point.

[0060] This application's hierarchical integrated kernel Gaussian process residual corrector innovatively links the process mechanism directly with the kernel function design. It uses a periodic kernel to quantify the cyclic fluctuations of the "forward-reverse" cycles, uses an anisotropic kernel to characterize the mismatch between the pressure direction and the surface normal, and pioneers a "consistency deviation kernel" specifically to quantify the residuals generated by multi-axis collaborative misalignment. This "mechanism-data" dual-drive mode significantly improves the accuracy and interpretability of residual correction.

[0061] In one alternative implementation, such as Figure 2 As shown, optimization module 5 includes the following functional units: The target setting unit 51 is used to obtain the target polishing removal amount. Specifically, the target removal amount can be expressed as... This corresponds to the expected removal amount or tolerance range for 99 points across five workpieces.

[0062] Loss function construction unit 52 is used to construct a composite loss function, which includes an accuracy loss term, an inter-workpiece consistency loss term, an intra-workpiece uniformity loss term, and a processing efficiency loss term. Specifically, given the target removal amount, a multi-objective constrained optimization problem is constructed, with the objective of minimizing the composite loss function. This composite loss function contains four competing loss terms, and its form is as follows:

[0063] in, To account for accuracy loss and ensure that the overall removal amount hits the target value, it is usually defined as the sum of the squares or absolute values ​​of the differences between the predicted removal amount and the target removal amount at each location. , It is the first One workpiece ( ), No. Points ( The predicted removal amount; The consistency loss between workpieces is used to minimize the difference in the average removal amount among the five workpieces, and can be defined as, for example, as the variance of the average removal amount of the five workpieces. ,in, It is the first Average removal amount per workpiece It is the total average removal amount of all workpieces; This refers to the uniformity loss within a workpiece, used to minimize the fluctuation in the removal amount at ninety-nine points within a single workpiece. For example, it can be defined as the sum of the range or standard deviation of the removal amount at each point within the workpiece. ; For efficiency losses, this is used to encourage the use of larger pressures and fewer revolutions, for example, defined as a negative value of the pressure or a linear function of the number of revolutions. , It is a coefficient that optimizes the tendency to regulate pressure. It is a coefficient that adjusts the number of rotations to optimize performance. The number of workpieces is 5. The number of points per workpiece is 99. It is the processing of the first The first workpiece, the first The combined pressure at each point It is the upper limit of the maximum resultant force specified in the process. It is the processing of the first The first workpiece, the first The total number of laps at each point It is the maximum total number of processing turns specified in the process; , , , The weighting coefficients for each loss term can be preset according to process requirements or adjusted experimentally. Minimize the composite loss function. The goal is to achieve the objective of having all five workpieces pass inspection at once, with highly consistent quality and optimal processing efficiency. In one alternative approach, the consistency loss between workpieces can be calculated based on the resultant force standard deviation in the real-time state of the different axes. This proactively guides the optimization algorithm to select parameters that can synchronize the process states of each axis, thereby incorporating the different axis standard deviation into the optimization objective.

[0064] The parameter optimization unit 53 uses a forward prediction model as the evaluator and a composite loss function as the optimization objective. It employs a constrained optimization algorithm for iterative searching, outputting the globally optimal combination of process parameters that minimizes the composite loss function. Specifically, this unit first randomly initializes 10 sets of parameter combinations in the parameter space and calculates the corresponding composite loss function values ​​for each. Then, using a constrained Bayesian optimization framework, with the forward prediction model as the evaluator, it uses the constraint expectation improvement acquisition function to guide the selection of the next set of parameters, updating the Gaussian process surrogate model in each iteration. The search stops when the objective function value changes by less than 0.001 for 10 consecutive iterations, or when the number of iterations reaches the preset maximum of 200 iterations. The final output is the globally optimal combination of process parameters that minimizes the composite loss function, including the Y-axis pressure, Z-axis pressure, number of forward rotations, and number of reverse rotations for each machining spindle.

[0065] The Bayesian reverse optimization module constructs a multi-objective constrained optimization problem with the consistency of batch workpiece removal amount as the core. Its composite loss function simultaneously minimizes accuracy loss, differences between workpieces, and fluctuations within workpieces, while maximizing processing efficiency gains. The final output is a global collaborative process scheme that enables all five workpieces to pass the test at once, with highly consistent quality and optimal processing efficiency, realizing a paradigm leap from "single-point parameter compensation" to "batch collaborative intelligent control".

[0066] For the residual Gaussian forward prediction and Bayesian inverse optimization model architecture, see [link / reference]. Figure 3 This system constructs a complete intelligent decision-making closed loop of "forward accurate prediction - reverse collaborative optimization". Targeting the specific process characteristics of five-axis parallel polishing, this closed-loop system deeply integrates multi-source process features, physical mechanism understanding, and data-driven learning, achieving end-to-end intelligent mapping from process parameters to mass production quality targets.

[0067] In one alternative implementation, such as Figure 2 As shown, the device also includes the following functional modules: The closed-loop compensation module 6 is used to apply the globally optimal combination of process parameters to the actual processing, and to update the main prediction module, residual correction module, fusion module and optimization module according to the deviation between the actual polishing removal amount and the target polishing removal amount.

[0068] Specifically, this module first sends the globally optimal combination of process parameters obtained through Bayesian optimization to the five-axis polishing machine control system to execute the actual polishing process. After processing, the actual polishing removal amount at each point on each workpiece is measured and compared with the target removal amount to calculate the deviation.

[0069] If the actual polishing removal amount does not meet the expected requirements (e.g., the deviation exceeds the preset tolerance range), all data from this processing (including process parameters and corresponding actual removal amounts) are added to the training set, triggering the main prediction module 2, residual correction module 3, and fusion module 4 to retrain and update the forward prediction model. Simultaneously, the optimization module 5 is triggered to re-execute Bayesian inverse optimization based on the updated forward prediction model, generating a new optimal combination of process parameters. This process is repeated to achieve dynamic compensation and continuous optimization of the parameters.

[0070] In this embodiment, the aforementioned functions are integrated into a parameter compensation software module. This software module receives polishing dimension detection results in real time, automatically invokes a forward prediction model and a Bayesian optimization algorithm, generates a parameter compensation scheme, and sends it to the machine tool control system for execution, forming an intelligent closed-loop control system of "perception-decision-execution". Through the collaborative work of the above modules, this device achieves a leap from traditional single-point parameter compensation to multi-axis batch collaborative intelligent control, enabling five machining spindles with different wear states to output five workpieces with highly consistent removal amounts while ensuring their respective machining accuracy.

[0071] like Figure 4 The diagram shows a flowchart of a multi-axis polishing batch consistency control method according to an embodiment of the present invention. It should be noted that the steps shown in the flowchart can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that presented here. The process includes the following steps: Step S101: Collect historical data of multi-axis parallel polishing and construct a collaborative feature matrix that includes the historical state of the coaxial axis and the real-time state of the heteroaxial axis.

[0072] Step S102: Construct a deep neural network model, using the collaborative feature matrix as input, to learn the nonlinear mapping relationship between process parameters and polishing removal amount, and output the main predicted value of polishing removal amount at each point of each workpiece.

[0073] Step S103: Construct a Gaussian process regression model, model and predict the residuals of the master predictor, and output the predicted residual values.

[0074] Step S104: Add the master prediction value and the residual prediction value to obtain the final polishing removal amount prediction value.

[0075] The multi-axis polishing batch consistency control method provided in this invention can execute the multi-axis polishing batch consistency control device provided in any embodiment of this invention, and has the corresponding operating steps and beneficial effects of the control device. Further functional descriptions of the above steps are the same as in the corresponding embodiments described above, and will not be repeated here.

[0076] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0077] The following is a detailed reference. Figure 5 The diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 501, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 502 or a program loaded from memory 508 into random access memory (RAM) 503. The RAM 503 also stores various programs and data required for the operation of the electronic device. The processor 501, ROM 502, and RAM 503 are interconnected via a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0078] Typically, the following devices can be connected to I / O interface 505: input devices 506 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 507 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 508 including, for example, magnetic tapes, hard disks, etc.; and communication devices 509. Communication device 509 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0079] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 509, or installed from a memory 508, or installed from a ROM 502. When the computer program is executed by the processor 501, it performs the functions defined in the multi-axis polishing batch consistency control method of the embodiments of the present invention.

[0080] Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0081] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the multi-axis polishing batch consistency control method shown in the above embodiments is implemented.

[0082] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A multi-axis polishing batch consistency control device, characterized in that, The device includes: The matrix construction module is used to collect historical data of multi-axis parallel polishing and construct a collaborative feature matrix that includes the historical state of the same axis and the real-time state of the opposite axis. The main prediction module is used to construct a deep neural network model, which takes the collaborative feature matrix as input, learns the nonlinear mapping relationship between process parameters and polishing removal amount, and outputs the main predicted value of polishing removal amount at each point of each workpiece. The residual correction module is used to construct a Gaussian process regression model, model and predict the residuals of the master predicted value, and output the residual predicted value. The fusion module is used to add the master prediction value and the residual prediction value to obtain the final polishing removal amount prediction value.

2. The multi-axis polishing batch consistency control device according to claim 1, characterized in that, The device further includes: The optimization module is used to construct a composite loss function and solve for the globally optimal process parameters through constrained Bayesian optimization.

3. The multi-axis polishing batch consistency control device according to claim 1, characterized in that, The main prediction module includes: A multi-branch feature extraction network includes at least one branch for extracting spatial distribution features of process parameters and at least one branch for extracting temporal dependency features of process parameters; The feature fusion unit is used to fuse the features output by the multi-branch feature extraction network. The regression mapping unit is used to map the fused features to the master predicted value of the polishing removal amount.

4. The multi-axis polishing batch consistency control device according to claim 1, characterized in that, The residual correction module includes: The residual calculation unit is used to obtain the prediction value of the main prediction module on the training set, compare it with the actual removal amount, and calculate the prediction residual. The kernel function construction and fusion unit is used to construct an integrated kernel containing at least two dedicated kernel functions. The dedicated kernel functions are used to quantify the residuals caused by pressure direction mismatch and the residuals caused by multi-axis coordination mismatch, respectively. The weights of each kernel function are adaptively determined through cross-validation to obtain a weighted integrated kernel function. The model training unit is used to construct a Gaussian process regression model with the predicted residual as the training target and the weighted ensemble kernel function as the covariance function, and to train the Gaussian process regression model through hyperparameter optimization. The residual prediction unit is used to input the collaborative feature matrix into the trained Gaussian process regression model and output the residual prediction value.

5. The multi-axis polishing batch consistency control device according to claim 4, characterized in that, The kernel function construction and fusion unit includes: The first kernel function construction sub-unit is used to construct the first dedicated kernel function. The first dedicated kernel function is used to perform nonlinear transformation on the pressure direction angle to quantify the residual caused by the mismatch between the pressure direction and the surface normal. The second kernel function construction subunit is used to construct the second dedicated kernel function. The second dedicated kernel function is used to calculate the difference in process state vectors between different machining axes and combine the batch machining information to quantify the residuals caused by multi-axis coordination misalignment. The kernel function weighting subunit is used to evaluate the performance of each kernel function through cross-validation, and to adaptively weight and combine each kernel function based on the performance score to output a weighted ensemble kernel function.

6. The multi-axis polishing batch consistency control device according to claim 2, characterized in that, The optimization module includes: The target setting unit is used to obtain the target polishing removal amount; The loss function construction unit is used to construct a composite loss function, which includes an accuracy loss term, an inter-workpiece consistency loss term, an intra-workpiece uniformity loss term, and a processing efficiency loss term. The parameter optimization unit is used to perform iterative search using a positive prediction model as the evaluator and the composite loss function as the optimization objective, employing a constrained optimization algorithm to output the globally optimal combination of process parameters that minimizes the composite loss function.

7. The multi-axis polishing batch consistency control device according to claim 6, characterized in that, The device further includes: The closed-loop compensation module is used to apply the globally optimal process parameter combination to the actual processing, and to update the main prediction module, the residual correction module, the fusion module and the optimization module according to the deviation between the actual polishing removal amount and the target polishing removal amount.

8. A method for batch consistency control in multi-axis polishing, characterized in that, The method includes: Collect historical data of multi-axis parallel polishing and construct a collaborative feature matrix that includes the historical state of the coaxial axis and the real-time state of the heteroaxial axis. A deep neural network model is constructed, and the nonlinear mapping relationship between process parameters and polishing removal amount is learned by taking the collaborative feature matrix as input. The main predicted value of polishing removal amount at each point of each workpiece is output. Construct a Gaussian process regression model to model and predict the residuals of the main predicted values, and output the predicted residual values; The master prediction value is added to the residual prediction value to obtain the final polishing removal amount prediction value.

9. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes these computer instructions to perform the multi-axis polishing batch consistency control method as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the multi-axis polishing batch consistency control method as described in claim 8.

Images