Optical cold processing polishing process parameter optimization method based on digital twinning

Abstract

The present application relates to the technical field of optical processing technology, and discloses an optical cold processing polishing process parameter optimization method based on digital twinning, which comprises the following steps: acquiring optical cold processing polishing data sets to construct a hybrid gate cycle unit network, learning a removal efficiency correction and damage depth prediction field, combining a conditional variational autoencoder and meta-learning optimization through hidden space fusion to obtain a process digital twin, and outputting a prediction state and a removal state probability hidden variable; taking the prediction state as an assimilation initial field, analyzing actual removal function parameters and surface shape error distribution fields, performing recursive Bayesian assimilation and optimal reconstruction, calculating a comprehensive deviation degree score to determine a deviation level, and generating a real-time processing state; taking the process digital twin as a virtual processing environment, constructing a process parameter decision maker to deduce an optimal process parameter sequence output in combination with a bat algorithm, converting the optimal process parameter sequence into a process parameter control instruction for execution, calculating a comprehensive control deviation to iteratively optimize, and improving processing efficiency and processing state timeliness.

Description

Technical Field

[0001] This invention relates to the field of optical processing technology, and more specifically to a method for optimizing optical cold processing polishing process parameters based on digital twins. Background Technology

[0002] Optical cold polishing is a crucial step in the manufacturing of optical components. Polishing process parameters directly determine the surface accuracy, surface roughness, and processing efficiency of these components. Traditional optical polishing relies heavily on manual experience to set process parameters, failing to establish a real-time mapping between the process and the final result. This makes it difficult to adapt to the personalized processing needs of optical components made of different materials and with different surface shapes. In actual production, environmental disturbances and tool wear can cause deviations from the actual process state, easily leading to problems such as excessive surface errors and low processing efficiency. Existing polishing process parameter optimization methods are mostly offline static optimizations, lacking dynamic perception and real-time control of the processing, and unable to correct polishing process parameters based on real-time conditions in a closed-loop manner.

[0003] Chinese Patent CN120068675B discloses a method for determining polishing process parameters based on image quality indicators using a neural network. The method includes: determining the structural parameters of an optical system according to specific requirements; marking optical components requiring process decisions within the optical system and recording historical processing data; constructing the optical system, placing the marked optical components within it, adjusting to obtain actual image quality indicators, and collecting the degree of degradation compared to the ideal system image quality; repeating the above steps to obtain a dataset; using the image quality degradation dataset as input and the component process parameters as output, training a modular convolutional neural network to obtain a predictive network model; when an optical system component manufacturing task arrives, inputting the system type and the image quality degradation tolerance into the predictive network model to obtain polishing process parameters that meet the image quality degradation tolerance. The advantages are: rapid decision-making, low computation time, and high ease of understanding.

[0004] Although existing methods for determining polishing process parameters based on image quality indicators, by using image quality degradation data as input and process parameters as output to train modular convolutional neural networks, have achieved rapid decision-making of polishing process parameters from image quality indicators, the existing optical cold processing polishing process parameters rely on human experience and are difficult to adjust dynamically in real time, resulting in lag in processing status perception, low processing efficiency and unstable quality. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a method for optimizing optical cold-working polishing process parameters based on digital twins. By constructing a standardized optical cold-working polishing dataset and integrating it with classical optical polishing mechanisms to build a hybrid gated recurrent unit network, a process digital twin capable of real-time output of predicted states and removal of latent state probability variables is generated. This solves the problem of delayed processing state perception caused by traditional offline detection. Through recursive Bayesian assimilation, real-time sensor observations are optimally reconstructed into overall surface shape error and subsurface damage depth. Combined with a comprehensive deviation score and a deterioration trend coefficient, early perception of processing risks is achieved, avoiding decision-making delays caused by state lag. By encapsulating the process digital twin as a virtual processing environment, a process parameter decision-maker is constructed using reinforcement learning. Parallel cold and hot start multi-step deductions are executed to output the optimal process parameter sequence. Then, based on comprehensive control deviations, the process digital twin and the process parameter decision-maker are iteratively updated online. This overcomes the shortcomings of low processing efficiency and unstable quality in existing methods, achieving real-time perception, adaptive control, and continuous optimization of the polishing process.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for optimizing optical cold-working polishing process parameters based on digital twins includes:

[0008] Collect sensor data, process parameters, surface shape error and subsurface damage data, and preprocess to generate an optical cold processing polishing dataset;

[0009] A hybrid gated recurrent unit network is constructed to learn the removal efficiency correction and damage depth prediction field. Through latent space fusion and combined with conditional variational autoencoder and meta-learning optimization, a process digital twin is obtained, and the predicted state and removal state probability latent variables are output.

[0010] Using the predicted state as the initial field for assimilation, the actual removal function parameters and surface error distribution field are analyzed, recursive Bayesian assimilation is performed, the overall surface error and subsurface damage depth are optimally reconstructed, the comprehensive deviation score is calculated to determine the deviation level, and the real-time processing state is generated.

[0011] The process digital twin is encapsulated as a virtual machining environment. A process parameter decision-maker is constructed and simulated in parallel within the virtual machining environment. The output of the optimal process parameter sequence is converted into process parameter control instructions. After execution, the comprehensive control deviation is calculated for iterative optimization.

[0012] Specifically, the steps for constructing a hybrid gated recurrent unit network include:

[0013] A hybrid gated recurrent unit network is constructed, including an input layer, a gated recurrent feature learning layer, a cross-modal coupling layer, a probabilistic latent variable extraction layer, and an output layer;

[0014] The input layer is used to receive the optical cold polishing dataset and extract the process parameter vector and multi-source sensing feature vector;

[0015] The gated loop feature learning layer is used to execute two parallel gated loop units. The first gated loop unit receives pressure time-series data and speed time-series data, learns the time-varying residual of removal efficiency as process parameters change, and calculates the removal efficiency correction coefficient by combining the Preston equation.

[0016] The second gated recurrent unit introduces a cross-modal attention mechanism, which obtains a context vector by weighted summation of attention, and connects it with the residual of the current hidden state to obtain an enhanced hidden state.

[0017] The abrasive passivation features and thermal accumulation features are extracted and concatenated with the enhanced hidden state to obtain a joint feature vector, which is used to control the update gate and reset gate of the second gated loop unit and outputs the damage depth prediction field.

[0018] Specifically, the steps for constructing a hybrid gated recurrent unit network also include:

[0019] A cross-modal coupling layer is used to receive the removal efficiency correction coefficient, generate the actual removal function by combining the current process parameters, and then stack the actual removal function and the damage depth prediction field in the workpiece grid coordinates to form a joint state diagram. The joint latent space representation vector is obtained by encoder mapping.

[0020] The probabilistic latent variable extraction layer takes the joint latent space representation vector as the input to be encoded and the process parameter vector and multi-source sensor feature vector as the process conditions. The mean and log-variance of the probability distribution of the state latent variables are removed by the output of the conditional variational autoencoder.

[0021] The state probability latent variables are removed by reparameterized sampling. The decoder reconstructs the joint latent space representation vector and multi-source sensing feature vector based on the removed state probability latent variables and process conditions.

[0022] The output layer is used to integrate and output the core state of the process digital twin.

[0023] Specifically, the steps for outputting the predicted state and removing the latent variables of the state probability include:

[0024] Using the optical cold polishing dataset as the training set, a meta-learning algorithm is used to calculate the joint loss function through an inner loop, and the task parameters are updated through the Adam optimizer.

[0025] The outer loop updates the initial parameters of the hybrid gated recurrent unit network based on the gradient aggregation of the updated task parameters, thus obtaining the meta-initial parameters;

[0026] During deployment, the initial parameters of the target process are fine-tuned using labeled samples to obtain a digital twin of the process. During real-time inference, the current process parameter vector and multi-source sensor feature vector are input, and the predicted state and the removed state probability latent variables are output through the trained hybrid gated recurrent unit network.

[0027] Specifically, the steps for analyzing the actual removal function parameters and surface error distribution field include:

[0028] Extract the actual removal function parameters from the predicted state to define the amount of material removed per unit time at each grid point;

[0029] The surface shape error distribution field of the current workpiece surface is extracted from the predicted state. The deviation is calibrated by combining the surface shape error data detected offline by the white light interferometer. The calibrated surface shape error distribution field is then interpolated into the workpiece grid coordinate system to form the initial surface shape error state vector.

[0030] The damage depth prediction field in the predicted state is concatenated with the initial surface error state vector in the same workpiece mesh coordinate system to construct the initial assimilation state vector.

[0031] Specifically, the steps for performing recursive Bayesian assimilation include:

[0032] The mean and logarithmic variance of the state probability latent variables are extracted to construct the error covariance matrix, and the initial square root covariance matrix is ​​obtained by Cholesky decomposition.

[0033] Force, vibration, and acoustic emission signals are acquired in real time, and after adaptive threshold wavelet denoising, multidimensional statistical features are calculated using a sliding time window to form a multimodal observation vector.

[0034] Using the initial assimilation state vector and the initial square root of covariance matrix as initial conditions, perform recursive Bayesian assimilation;

[0035] Sampling points and corresponding mean weights and covariance weights are generated according to the square root unscented transformation rule. Each sampling point is mapped to a priori sampling point through a preset state transition function. The priori state mean is obtained by weighting. The bias matrix is ​​constructed and the prior covariance square root is obtained by QR decomposition and Cholesky rank-one update.

[0036] Each prior sampling point is mapped to a prior observation sampling point through a preset nonlinear observation function, and the weighted a prior observation mean is obtained. The observation bias matrix is ​​constructed and then subjected to QR decomposition and Cholesky rank-one update to obtain the square root of the observation covariance.

[0037] Specifically, the steps for optimal reconstruction include:

[0038] The cross-covariance matrix between the state and the observation is calculated using the weighted deviation between each prior sampling point and the mean of the prior state, and the weighted deviation between each prior observation sampling point and the mean of the prior observation. The Kalman gain is then calculated by combining the square root of the observation covariance.

[0039] The observation residual is obtained by calculating the difference between the multimodal observation vector and the prior observation mean. The posterior state estimate is obtained by combining the Kalman gain matrix and the prior state mean.

[0040] The multimodal observation vectors acquired in real time are assimilated into the state field to perform optimal reconstruction of the overall surface shape error and subsurface damage depth.

[0041] Specifically, the steps for determining the deviation level include:

[0042] Health theory feature values ​​are extracted from the reconstructed multi-source sensing feature vector to replace invalid feature values ​​in the multimodal observation vector, forming a corrected observation vector that re-participates in Kalman gain calculation and state update;

[0043] The surface shape error and subsurface damage depth are extracted from the posterior state estimate. The root mean square error of the surface shape wavefront is calculated, and the ratio of the damaged area exceeding the standard is calculated. Then, the surface shape deviation and damage deviation are calculated. The comprehensive deviation score is calculated by combining the exponential weighted moving average method. ;

[0044] Set deviation threshold , , Determine if it deviates from the standard level;

[0045] like If it is, then it is judged as a level one deviation; if If so, it is judged as a level two deviation;

[0046] like If so, it is judged as a level three deviation; if If so, it is considered normal.

[0047] Specifically, the steps for generating real-time processing status include:

[0048] The comprehensive deviation scores calculated for each processing cycle are arranged in chronological order to form a comprehensive deviation sequence. Linear regression is performed using the least squares method to fit the deviation change line and extract the slope as the deterioration trend coefficient to detect the risk of deterioration in advance.

[0049] If the deterioration trend coefficient is greater than 0 and the comprehensive deviation score is within... and The predicted time for the interval to reach level one deviation is... and The difference is divided by the deterioration trend coefficient;

[0050] If the predicted time is less than or equal to 5 processing cycles, a Level 1 warning signal will be issued; if the predicted time is less than or equal to 2 processing cycles, a Level 2 warning signal will be issued.

[0051] The final real-time processing status is generated.

[0052] Specifically, the steps for outputting the optimal process parameter sequence include:

[0053] Based on the real-time processing status, the state transition function and nonlinear observation function of the process digital twin are encapsulated into a virtual processing environment;

[0054] After receiving an action, the virtual processing environment calls the state transition function to predict the surface error distribution field and the damage depth prediction field, calculates the next state, and outputs an instant reward and a termination flag.

[0055] A process parameter decision-maker consisting of a stochastic policy network and a Q-value network is constructed. The initial parameters of the stochastic policy network are pre-trained through meta-reinforcement learning to initialize the experience replay buffer and construct a prior action generator to output action sequences.

[0056] At each processing decision moment, N parallel inference trajectories are executed. During hot start, the state sampling action sequence is obtained based on the random policy network. During cold start, Gaussian exploration noise is superimposed on the action sequence.

[0057] Robust returns are calculated by K samplings after removing latent variables of state probabilities. The process parameter sequences corresponding to each inferred trajectory are used as the initial population. After iterative optimization by the bat algorithm, the optimal process parameter sequence is output.

[0058] Specifically, the steps for iterative optimization include:

[0059] The optimal process parameter sequence is converted into process parameter control instructions and sent to the equipment CNC system and actuators. During the processing, multi-source sensor data and actual process parameter feedback values ​​are collected. After the process is completed, the state observation values ​​are obtained, forming real experience tuples and storing them in the experience playback buffer pool.

[0060] The root mean square deviation and maximum depth deviation are calculated from the collected feedback data, and the weighted sum is used to calculate the overall control deviation.

[0061] If the root mean square deviation is greater than the preset first deviation threshold or the maximum depth deviation is greater than the preset second deviation threshold, the process digital twin is fine-tuned online using real experience tuples as samples, the parameters of the state transition function and nonlinear observation function are updated, and the comprehensive control deviation is fed back to the process parameter decision-maker for iterative optimization.

[0062] After the update, real experience is sampled from the experience replay buffer, and the gradient of the stochastic policy network and Q-value network is updated in combination with virtual experience. The gradient update step size is adjusted by using comprehensive control bias.

[0063] The beneficial effects of this invention are:

[0064] 1. This invention fundamentally solves the problems of lag in state perception and difficulty in real-time observation of surface shape and subsurface damage in traditional optical cold processing by multi-source sensor data fusion, hybrid gated cyclic unit network modeling, and square root unscented Kalman filtering recursive assimilation. Using the predicted state of the process digital twin as the initial assimilation field, it analyzes the actual removal function parameters and surface shape error distribution field in real time, combining force, vibration, and acoustic emission signals to achieve optimal state reconstruction across the entire field. When sensors malfunction, it generates a corrected observation vector by removing latent variables of the state probability and using a conditional variational autoencoder, ensuring continuous and uninterrupted assimilation. Based on the optimal reconstruction result, it calculates a comprehensive deviation score and determines the deviation level, triggering trend warnings to identify processing deterioration risks in advance. This improves the real-time performance, reliability, and prediction accuracy of state perception, providing stable and reliable real-time processing state support for process decisions and avoiding over-polishing, increased damage, and quality control failure caused by perception lag.

[0065] 2. This invention achieves risk-free multi-track processing by encapsulating a process digital twin into a virtual manufacturing environment.

[0066] Parallel simulation, balancing damage suppression, processing efficiency, and parameter smoothness; employing a cold-hot start combined with a stochastic strategy network and robust reward screening, the optimal process parameter sequence is output after iterative optimization using the bat algorithm. This optimal process parameter sequence is converted into process parameter control instructions. After execution, if the root mean square deviation exceeds a preset first deviation threshold or the maximum depth deviation exceeds a preset second deviation threshold, the process digital twin is fine-tuned online using real experience tuples as samples, while simultaneously iteratively optimizing the stochastic strategy network and Q-value network. This invention eliminates the need for repeated manual adjustments, adapts to different workpieces and working conditions, ensures surface accuracy while suppressing subsurface damage, improves processing efficiency and quality stability, and solves the problems of traditional processes relying on experience, low efficiency, and large quality fluctuations. Attached Figure Description

[0067] Figure 1 A schematic diagram of a method for optimizing optical cold-working polishing process parameters based on digital twins;

[0068] Figure 2 This is a flowchart of constructing a hybrid gated recurrent unit network in this invention;

[0069] Figure 3 This is a flowchart for determining the deviation level in this invention;

[0070] Figure 4This is a flowchart illustrating the process of outputting the optimal process parameter sequence in this invention. Detailed Implementation

[0071] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0072] Example

[0073] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a method for optimizing optical cold-working polishing process parameters based on digital twins, including the following steps:

[0074] Force sensors, torque sensors, triaxial vibration sensors, acoustic emission sensors, flow meters, and infrared temperature sensors deployed on the polishing spindle, workpiece axis, and polishing fluid system are used to synchronously acquire sensor data, including force, torque, vibration, acoustic emission, flow rate, and temperature, at a sampling frequency greater than or equal to 1 kHz. Process parameters such as spindle speed, polishing pressure setpoint, and dwell time are acquired from the equipment's CNC system. After each processing step, the workpiece is inspected offline using a white light interferometer to collect surface shape error and subsurface damage data. The multi-source process data is preprocessed, including adaptive threshold wavelet denoising and short-time window statistical feature extraction on the sensor data at edge intelligent nodes, aligning the timestamps of the multi-source data, forming a process parameter vector from the process parameters of each cycle and the multi-source sensor features, spatially matching the process vector with the surface shape error and subsurface damage data according to the workpiece grid coordinates, and using deep discriminative correlation analysis to maximize the fusion of multimodal feature mutual information, ultimately generating a standardized optical cold processing polishing dataset. The multi-source process data includes sensor data, process parameters, surface shape error, and surface damage data.

[0075] Based on the optical cold polishing dataset, a hybrid gated recurrent unit network is constructed by combining the classic mechanism of optical polishing. The first gated recurrent unit is used to learn the time-varying residual of material removal efficiency. By receiving the optical cold polishing dataset, it uses the gating mechanism to capture the hysteresis response of removal efficiency to pressure and speed fluctuations, and outputs the removal efficiency correction. The second gated recurrent unit is used to deduce the dynamic evolution of subsurface damage depth. It senses the coupling effect of abrasive passivation and thermal accumulation through cross-modal interaction, and performs attention alignment between the damage evolution state and the sensing features in the latent space, outputting the damage depth prediction field of the workpiece subsurface. The actual removal function formed based on the removal efficiency correction is fused with the damage depth prediction field in the latent space, and a conditional variational autoencoder is embedded to extract the latent variables of the removal state probability. The initial parameters of the hybrid gated recurrent unit network are optimized by the meta-learning algorithm to obtain the process digital twin. Finally, the predicted state and the latent variables of the removal state probability of the process digital twin are output.

[0076] Using the predicted state of the process digital twin as the initial assimilation field, the actual removal function parameters and surface error distribution field are extracted from the initial assimilation field. The multidimensional statistical features of real-time force, vibration, and acoustic emission are used as multimodal observation vectors. Recursive Bayesian assimilation is performed using square root unscented Kalman filtering to optimally reconstruct the overall surface shape error and subsurface damage depth that cannot be directly measured. When the sensor malfunctions, such as drift or short-term loss, the probability sampling inference of the variational autoencoder is triggered by the latent variable of the removal state probability of the process digital twin to generate a reasonable corrected observation vector, thereby maintaining the continuity of assimilation. Based on the optimal reconstruction result, the root mean square error of the surface wavefront and the ratio of the area exceeding the damage standard are calculated. The continuous comprehensive deviation score is calculated through an exponentially weighted sliding window to determine the deviation level, generating a real-time processing state that includes the deviation level and the subsurface damage state, realizing early perception of the risk of processing deterioration.

[0077] Based on real-time machining status, the process digital twin is encapsulated as a virtual machining environment. A process parameter decision-maker is constructed based on reinforcement learning. The state space includes deviation level, subsurface damage exceeding the standard, machining time, and cumulative removal amount. The action space consists of continuously adjustable polishing pressure, spindle speed, tool oscillation frequency, and polishing fluid flow rate parameters. The reward function is a weighted synthesis of subsurface damage suppression, machining efficiency improvement, and parameter adjustment smoothing, ensuring that equipment impact is avoided while guaranteeing machining quality. The process parameter decision-maker performs cold and hot start multi-step simulations in parallel in the virtual machining environment. Using the process parameter sequence corresponding to each simulation trajectory as the initial population, the optimal process parameter sequence is output after iterative optimization using the bat algorithm. The optimal process parameter sequence is converted into process parameter control commands. After execution, feedback data is collected to calculate the comprehensive control deviation, and the process digital twin is updated online using the comprehensive control deviation. At the same time, the feedback is fed back to the process parameter decision-maker for iterative optimization.

[0078] Specifically, the steps for outputting the predicted state of the process digital twin and removing latent variables of state probability include:

[0079] Based on the optical cold polishing dataset, a hybrid gated recurrent unit network was constructed by combining the classical mechanism of optical polishing. The network includes an input layer, a gated recurrent feature learning layer, a cross-modal coupling layer, a probabilistic latent variable extraction layer, and an output layer. The classical mechanism of optical polishing includes the linear relationship between removal efficiency and pressure and relative velocity described by the Preston equation, as well as the influence of abrasive passivation and thermal accumulation on removal efficiency and the evolution of subsurface damage depth.

[0080] The input layer receives the optical cold processing polishing dataset and extracts process parameter vectors and multi-source sensing feature vectors through feature extraction.

[0081] The gated loop feature learning layer is used to execute two parallel gated loop units. The first gated loop unit receives the pressure time-series data and rotation speed time-series data from the optical cold polishing dataset. It uses the gating mechanism to capture the hysteresis response of the removal efficiency to pressure and rotation speed fluctuations. By updating the gate and resetting the gate, it dynamically adjusts the fusion of the forgotten historical information and the current input, and learns the time-varying residual of the removal efficiency as the process parameters change. The hidden state of the first gated loop unit at the final moment is mapped to the time-varying residual through a fully connected layer. The time-varying residual is normalized with the theoretical removal efficiency calculated based on the classical Preston equation to obtain the dimensionless removal efficiency correction coefficient.

[0082] The dynamic adjustment process includes: obtaining the pressure change rate by performing differential calculations on the values ​​of adjacent time points of pressure time series data, obtaining the speed change rate by performing differential calculations on the values ​​of adjacent time points of speed time series data, setting a pressure change threshold based on the allowable pressure fluctuation range of the process, with a value of 5%~10% of the rated pressure, and setting a speed change threshold based on the allowable speed fluctuation range of the process, with a value of 5%~10% of the rated speed. The rated pressure and rated speed are determined according to the standard specifications of the optical cold processing polishing process. When the pressure change rate is greater than the pressure change threshold or the speed change rate is greater than the speed change threshold, the reset gate output value is close to 0, quickly forgetting historical information that does not match the current process state. When the pressure change rate is less than or equal to the pressure change threshold and the speed change rate is less than or equal to the speed change threshold, the reset gate output value is close to 1, retaining historical removal efficiency information to improve prediction stability. When the pressure change rate is greater than half of the pressure change threshold but less than or equal to the pressure change threshold or the speed change rate is greater than half of the speed change threshold but less than or equal to the speed change threshold, the update gate output value is between [0,1], gradually integrating new process information to capture the hysteresis response of removal efficiency.

[0083] The second gated loop unit introduces a cross-modal attention mechanism in the latent state evolution. Using the current latent state as the query pair, it calculates attention weights in the multi-modal sensing feature subspaces such as force, vibration, and acoustic emission, and sums them to obtain a context vector. The context vector is then residually connected to the current latent state to obtain an enhanced latent state that integrates multi-modal sensing information. Abrasive passivation features and thermal accumulation features are extracted from the multi-source sensing features. The enhanced latent state is then concatenated with the abrasive passivation features and thermal accumulation features to obtain a joint feature vector, which serves as the control input for the update gate and reset gate of the second gated loop unit. Through cross-modal interactive sensing and gated state updates, attention alignment between the damage evolution state and multi-source sensing features is performed in the latent space, and finally, the damage depth prediction field of the workpiece subsurface is output.

[0084] A cross-modal coupling layer receives the removal efficiency correction coefficient output by the first gated recurrent unit and generates the actual removal function by combining it with the current process parameters. The expression of the actual removal function is: Removal amount = Preston constant × Current pressure × Current relative velocity × Removal efficiency correction coefficient. The Preston constant is obtained from the material parameter table in the optical cold polishing dataset, and the current pressure and current relative velocity are obtained from the process parameter vector output by the input layer. The actual removal function and the damage depth prediction field are hierarchically superimposed in the workpiece grid coordinates to form a joint state map. An encoder composed of convolutional and fully connected layers maps the joint state map into a low-dimensional joint latent space representation vector. During training, mutual information maximization regularization is introduced to ensure that the joint latent space representation vector retains the key discriminative information of the removal and damage coupling. The workpiece grid coordinates are determined according to the workpiece geometry and the discrete grid node coordinates set by the finite element partitioning rules. The finite element partitioning rules are based on the maximum diameter of the workpiece and the effective diameter of the polishing tool's removal function, using a uniform grid partitioning method. The grid spacing is between one-quarter and one-half of the effective diameter of the polishing tool's removal function, ensuring sufficient sampling of the spatial distribution of the removal function.

[0085] The probabilistic latent variable extraction layer is used to employ a conditional variational autoencoder. The joint latent space representation vector is used as the input to be encoded, and the process parameter vector and multi-source sensor feature vector are used as process conditions. The mean and log-variance of the probability distribution of the removed state latent variables are output by the conditional variational autoencoder. The removed state probabilistic latent variables are obtained by reparameterization sampling. The decoder reconstructs the joint latent space representation vector and multi-source sensor feature vector based on the removed state probabilistic latent variables and process conditions, and obtains the reconstructed joint latent space representation vector and multi-source sensor feature vector.

[0086] The output layer is used to integrate and output the core states of the process digital twin, including the removal of state probability latent variables, the subsurface damage depth prediction field, the actual removal function, and the reconstructed joint latent space representation vector and multi-source sensing feature vector, providing probabilistic information for process monitoring and optimization.

[0087] Using a dataset of optical cold-working polishing processes with various conditions as the training set, a meta-learning algorithm is employed to optimize the initial parameters of a hybrid gated recurrent unit network. The specific optimization process of the meta-learning algorithm includes: the inner loop calculates the joint loss function, which includes the first-level error of removing the efficiency correction coefficient, the second-level error of the damage depth prediction field, the reconstruction loss and KL divergence of the conditional variational autoencoder, and a mutual information maximization regularization term; the task parameters are updated using the Adam optimizer, which includes all trainable weights and bias parameters in the gated recurrent feature learning layer, the cross-modal coupling layer, and the probabilistic latent variable extraction layer; the outer loop updates the parameters based on the gradient aggregation of the updated task parameters in each task. The initial parameters of the hybrid gated recurrent unit network are used to obtain meta-initial parameters that can quickly adapt to new processes. During deployment, the meta-initial parameters are fine-tuned using 5 to 20 sets of labeled samples from the target process to obtain a process digital twin. During real-time inference, the current process parameter vector and multi-source sensor feature vector are input, and the trained hybrid gated recurrent unit network outputs the predicted state of the process digital twin and the removed state probability latent variables. The predicted state includes the damage depth prediction field. The fine-tuning process involves calculating the joint loss function using 5 to 20 sets of labeled samples from the target process, and using gradient descent to iteratively update the trainable parameters based on the meta-initial parameters until the loss converges.

[0088] For example, taking the polishing process of a fused silica lens with a diameter of 100mm as an example, a hybrid gated recurrent unit network is constructed based on the optical cold working polishing dataset and combined with the classical mechanism of optical polishing.

[0089] The input layer receives the optical cold processing polishing dataset and extracts the process parameter vector and multi-source sensing feature vector. The process parameter vector includes pressure of 0.35 MPa, rotation speed of 150 r / min, and polishing slurry concentration of 12%. The multi-source sensing feature vector includes force of 0.45 kN, vibration amplitude of 0.23 m / s², and root mean square value of acoustic emission of 0.67 V.

[0090] The gated loop feature learning layer executes two parallel gated loop units. The first gated loop unit receives pressure timing [0.32, 0.31, 0.33, 0.52, 0.51] in MPa and rotational speed timing [120, 119, 121, 180, 179] in r / min. The pressure change rates are calculated as 0.01, 0.02, 0.19, and 0.01 (MPa / s) using adjacent differentials, and the rotational speed change rates are calculated as 1, 2, 59, and 1 (r / min / s). The rated pressure is 0.35 MPa, so the pressure change threshold is taken as the rated pressure. 10% of the rated speed is 0.035 MPa; the rated speed is 150 r / min, so the speed change threshold is 10% of the rated speed, which is 15 r / min; taking a pressure change rate of 0.19 as an example, when the pressure change rate of 0.19 is greater than the pressure change threshold of 0.035 and the corresponding speed change rate of 59 is greater than the speed change threshold of 15, the reset gate output value is close to 0, and the history is quickly forgotten; the hidden state of the first gated loop unit at the final moment is mapped to a time-varying residual of 0.12 through a fully connected layer, and the time-varying residual of 0.12 is normalized to the theoretical removal efficiency of 0.85 μm / min calculated based on the classical Preston equation, and the removal efficiency correction coefficient is 1.14;

[0091] The second gated recurrent unit uses the current hidden state as the query, calculates attention weights for the multimodal sensing feature subspace, and obtains a context vector by weighted summation. After residual connection, it obtains an enhanced hidden state that integrates multimodal sensing information. It extracts abrasive passivation features and thermal accumulation features from the multi-source sensing features, and concatenates the enhanced hidden state, abrasive passivation features, and thermal accumulation features to obtain a joint feature vector, which serves as the control input for the update gate and reset gate of the second gated recurrent unit. The maximum value of the final output damage depth prediction field is 2.3 μm.

[0092] A cross-modal coupling layer is used to receive the removal efficiency correction coefficient of 1.14 and generate the actual removal function by combining it with the current process parameters: removal amount = Preston constant × current pressure × current relative velocity × 1.14; where the Preston constant is 3.2e-6 mm³ / (N·m); the actual removal function and the damage depth prediction field are hierarchically superimposed in the workpiece grid coordinates to form a joint state diagram. An encoder composed of convolutional layers and fully connected layers is used to map the joint state diagram into a 64-dimensional joint latent space representation vector, and the mutual information maximization regularization term is set to 0.01;

[0093] The probabilistic latent variable extraction layer takes the joint latent space representation vector as the input to be encoded and the process parameter vector and multi-source sensor feature vector as the process conditions. The conditional variational autoencoder outputs the mean of the probability distribution of the removed state latent variables, which is 0.32 and the log-variance is -1.85. After reparameterization and sampling, the removed state probabilistic latent variables are 0.28. The joint latent space representation vector and multi-source sensor feature vector are obtained through the decoder reconstruction.

[0094] The output layer integrates the core states of the output process digital twin, including the removal of the latent state probability variable 0.28, the maximum value of the subsurface damage depth prediction field 2.3μm, the actual removal function, the reconstructed joint latent space representation vector, and the multi-source sensing feature vector, providing probabilistic information for process monitoring and optimization.

[0095] An optical cold polishing dataset containing 50-100 samples under 5 different process conditions was used. The joint loss was calculated in the inner loop, including: a first-order error of 0.03 for removal efficiency correction coefficient, a second-order error of 0.15 for damage depth prediction field, a reconstruction loss of 0.08 for the conditional variational autoencoder, a KL divergence of 0.02, and a mutual information maximization regularization term of 0.01. The summation yielded a joint loss of 0.29. Adam optimizer parameters were set as follows: inner loop learning rate of 0.001, outer loop learning rate of 0.001... The learning rate is 0.0001, the first moment decay rate is 0.9, the second moment decay rate is 0.999, and the weight decay rate is 0.0001. The inner loop stops when the number of iterations in each task reaches 50 rounds or the joint loss function decreases by less than 0.001 for 5 consecutive rounds. The outer loop stops when the outer loop iterates for 100 rounds. The outer loop updates the initial parameters of the hybrid gated recurrent unit network based on the gradient aggregation of the task parameters updated by the inner loop on each task, thus obtaining the meta-initial parameters that can quickly adapt to the new process.

[0096] During deployment, 12 tagged samples were provided for the target fused silica process, with pressures ranging from 0.30 to 0.55 MPa and rotation speeds from 110 to 190 r / min. Based on a joint loss of 0.29, gradient descent was used with a learning rate of 0.001. All trainable parameters were iteratively updated on the basis of the initial parameters until the loss converged after 38 steps, thus obtaining a digital twin of the process.

[0097] During real-time inference, the current process parameter vector and multi-source sensor feature vector are input. Through the trained hybrid gated recurrent unit network, the predicted state and removed state probability latent variables of the process digital twin are output, including the maximum value of the damage depth prediction field of 2.3μm and the removed state probability latent variable of 0.28, so as to realize real-time perception of processing status and risk prediction.

[0098] Specifically, the steps for generating a real-time processing state that includes deviation level and subsurface damage state include:

[0099] The actual removal function parameters are extracted from the predicted state of the process digital twin, including but not limited to the Preston constant, current applied pressure, current relative velocity and removal efficiency correction coefficient, to define the amount of material removed per unit time at each grid point on the workpiece surface at the current moment;

[0100] The surface shape error distribution field of the current workpiece surface is extracted from the predicted state. Combined with the surface shape error data of the white light interferometer offline detection after the previous process, the least square surface matching method is used to perform deviation calibration to eliminate the systematic deviation between digital twin prediction and offline detection. The calibrated surface shape error distribution field is interpolated to the same workpiece grid coordinate system as the predicted state to form the assimilated initial surface shape error state vector.

[0101] The predicted field of subsurface damage depth in the predicted state is mapped to the same workpiece mesh coordinate system and concatenated with the initial surface error state vector to construct a complete initial assimilation state vector. The initial assimilation state vector includes the surface error value of each mesh point and the corresponding subsurface damage depth value of each mesh point. The initial assimilation state vector also implies the distribution information of the material removal state in the whole field. The change in surface error directly corresponds to the amount of material removed per unit time.

[0102] Based on the fact that the latent variables of the removed state probabilities follow a multivariate normal distribution, the mean and logarithmic variance corresponding to each state component are extracted. The exponent of the logarithmic variance is used as the initial uncertainty of the corresponding surface shape error and subsurface damage depth. A diagonal matrix is ​​constructed as the initial error covariance matrix, and the diagonal elements are taken as the exponential logarithmic variance values ​​of each state component. The initial error covariance matrix is ​​decomposed by Cholesky to obtain the initial square root covariance matrix, thus completing the square root initialization of the state covariance.

[0103] The force, vibration, and acoustic emission signals on the polishing spindle and workpiece shaft are collected in real time by sensors. These three types of real-time signals have strong dynamic response characteristics to the removal state and subsurface damage evolution. At the edge node, adaptive threshold wavelet denoising is performed on each type of real-time signal. The specific denoising process includes: multi-scale wavelet decomposition of the real-time signal, adaptive calculation of the threshold at each decomposition scale according to the current noise level, soft thresholding of detail coefficients, and reconstruction of the denoised signal through inverse wavelet transform to filter out high-frequency electromagnetic interference and mechanical background noise.

[0104] A fixed-length time window is slid across the denoised signal. The fixed length is determined based on the sensor sampling frequency and the dynamic response time of the process, and is set to 1 second, corresponding to 1000 sampling points at a sampling frequency of 1 kHz. Multidimensional statistical features within each time window are calculated to form a multimodal observation vector for the current moment, including force features, vibration features, and acoustic emission features. The force features include the mean, standard deviation, and peak factor. The vibration features include the root mean square value of the triaxial vibration signal, the power spectrum energy of a specific frequency band calculated by fast Fourier transform, and kurtosis, which are used to sense abrasive breakage and abrupt changes in friction state. The specific frequency band range is set according to the abrasive breakage and subsurface damage characteristic frequency bands determined by offline experiments. The acoustic emission features include the root mean square value of the signal, ring count, and center frequency, which are highly sensitive to the generation of microcracks and the propagation of subsurface damage.

[0105] The initial assimilation state vector is used as the initial state estimate at the current time step. The initial covariance square root is extracted from the initial covariance square root matrix and used as the corresponding state covariance square root. When the assimilation is recursively applied to the Kth step, the initial state is the posterior estimate of the previous time step; where K≥1.

[0106] Recursive Bayesian assimilation is performed using square root unscented Kalman filtering. The specific process includes: generating a set of deterministic sampling points based on the posterior state estimate and the square root of the state covariance from the previous time step, according to the square root unscented transformation rule. Each sampling point corresponds to a mean weight for calculating the mean and a covariance weight for calculating the covariance. The number of deterministic sampling points in the set is 2L+1, where L is the dimension of the initial assimilation state vector.

[0107] Each sampling point is sequentially input into the state transition function to obtain the corresponding prior sampling point. The state transition function is defined by the optical cold processing removal mechanism, and the calculation rules of the state transition function are as follows: For the surface shape error value in the initial assimilation state vector, the predicted surface shape error is obtained by subtracting the product of the actual removal function and the residence time from the surface shape error value at the previous moment; For the subsurface damage depth value in the initial assimilation state vector, the process parameters and state at the current moment are used as input, and the process digital twin is called to perform a forward prediction step to obtain the subsurface damage depth prediction field at the next moment; By mapping the state of each sampling point using the calculation rules of the state transition function, the corresponding prior sampling point is obtained.

[0108] The values ​​of each prior sampling point are weighted and summed according to their mean weights to obtain the prior state mean. A deviation matrix is ​​constructed by calculating the weighted deviation between each prior sampling point and the prior state mean. The deviation matrix is ​​then decomposed using QR decomposition to obtain the square root of the prior covariance. The square root of the process noise is then used for Cholesky rank-one update to obtain the complete square root of the prior covariance. The square root of the process noise is the Cholesky decomposition factor of the process noise covariance matrix, which is determined based on the historical statistical variance of the process modeling error.

[0109] Based on prior physical knowledge of polishing contact mechanics and acoustic emission generation mechanisms, a nonlinear observation function is derived and constructed. This function consists of a set of explicit physical mapping relationships. For example, the mean value of the force characteristics is proportional to the local removal amount of surface shape error and the contact area; the root mean square value of acoustic emission has a nonlinear monotonic mapping relationship with the square root of the subsurface damage depth; and the vibration power spectrum energy is correlated with the abrasive particle breakage intensity. The specific coefficients of each mapping relationship are determined through offline calibration experiments and stored as fixed parameters in the system. Each prior sampling point is sequentially input into the nonlinear observation function to obtain the corresponding prior observation points. For each sample point, a weighted sum is calculated based on the mean weight of all prior observation sample points to obtain the prior observation mean. An observation bias matrix is ​​constructed by calculating the weighted deviation between each prior observation sample point and the prior observation mean, and QR decomposition is performed to obtain the square root of the observation covariance. The square root of the observation noise is then used for Cholesky rank-one update to obtain the complete square root of the observation covariance. The square root of the observation noise is the Cholesky decomposition factor of the observation noise covariance matrix, which is determined based on the measurement noise variance obtained from sensor measurements and the statistical characteristics of historical static tests.

[0110] The cross-covariance matrix between the state and the observation is calculated using the weighted deviation between each prior sampling point and the mean of the prior state, and the weighted deviation between each prior observation sampling point and the mean of the prior observation.

[0111] Using the square root of the observation covariance and the cross-covariance matrix, the Kalman gain is calculated by solving a system of equations, such that the Kalman gain multiplied by the square root of the observation covariance equals the cross-covariance matrix of the state and the observation. The difference between the multimodal observation vector and the prior observation mean obtained at the previous time step is calculated to obtain the observation residual. The product of the Kalman gain matrix and the observation residual is calculated, and the product is added to the prior state mean to obtain the posterior state estimate. By assimilating the real-time acquired multimodal observation vector into the state field, optimal reconstruction of the overall surface shape error and subsurface damage depth is achieved.

[0112] By using Kalman gain and the square root of observation covariance to update the prior square root of covariance under Cholesky, the updated posterior square root of covariance is obtained, which reduces the uncertainty of state estimation.

[0113] The physical validity and statistical characteristics of the signals collected by the sensors are continuously monitored. Sensor health diagnosis conditions include: signal exceeding the physical range, sensor drift or short-term loss. Once any sensor health diagnosis condition is triggered, the sensor signal data is determined to be unusable, and the probabilistic latent variable inference mechanism of the process digital twin is invoked.

[0114] Random sampling is performed on the normal distribution of the state probability latent variables output from the process digital twin to obtain latent variable samples. The latent variable samples and the current process parameter vector are input into the decoder of the conditional variational autoencoder. The decoder will reconstruct the complete joint latent space representation vector and the corresponding multi-source sensor feature vector. The health theory feature value corresponding to the fault sensor is extracted from the reconstructed multi-source sensor feature vector. The invalid feature value calculated by the fault sensor in the multimodal observation vector is replaced with the health theory feature value to form a reasonable corrected observation vector. The corrected observation vector is then used to re-enter the Kalman gain calculation and state update stage, so that the assimilation continuity is maintained during the state update without interruption due to the lack of sensor data.

[0115] The surface shape error is extracted from the posterior state estimate of the optimal reconstruction. All workpiece mesh points are traversed, the mean of the squares of the surface shape error is calculated, and the square root of the mean is taken to obtain the root mean square of the surface shape wavefront error at the current moment. The subsurface damage depth is extracted from the posterior state estimate of the optimal reconstruction. All workpiece mesh points are traversed, the number of workpiece mesh points with subsurface damage depth greater than the damage threshold is counted, the total number of mesh points on the workpiece surface is counted, and the number of workpiece mesh points is divided by the total number of mesh points to obtain the damage excess area ratio. The damage threshold is determined according to the upper limit of the subsurface damage depth allowed by the process, and is set to 5 micrometers.

[0116] The surface shape deviation is obtained by calculating the ratio of the root mean square error of the surface shape wavefront to the root mean square threshold. The damage deviation is obtained by dividing the ratio of the area exceeding the damage standard by the ratio of the maximum area exceeding the damage standard. The root mean square threshold is determined according to the surface shape accuracy requirements of the optical element, and the value is λ / 50~λ / 10, where λ is the working wavelength, usually 632.8 nm. The maximum area exceeding the damage standard is determined according to the damage tolerance requirements of the optical element, and the value is 0.05~0.1.

[0117] The continuous comprehensive deviation score is calculated using the exponentially weighted moving average method. The calculation process includes: determining the surface deviation coefficient based on the quality priority and processing requirements of the optical element, with a value of 0.6; since the sum of the surface deviation coefficient and the damage deviation coefficient is 1, the damage deviation coefficient is 0.4; determining the smoothing coefficient based on the dynamic change rate of the process, with a value of 0.1~0.3, to balance the influence of the current state and the historical state; calculating the product of the surface deviation coefficient and the surface deviation degree to obtain the surface deviation weighted degree; calculating the product of the damage deviation coefficient and the damage deviation degree to obtain the damage weighted degree; summing the surface deviation weighted degree and the damage weighted degree to obtain the instantaneous comprehensive deviation degree at the current moment; multiplying the instantaneous comprehensive deviation degree by the smoothing coefficient to obtain the current instantaneous weighted term; multiplying the difference between the value 1 and the smoothing coefficient by the comprehensive deviation degree score at the previous moment to obtain the historical weighted term; and summing the current instantaneous weighted term and the historical weighted term to obtain the continuous comprehensive deviation degree score. ;

[0118] Based on the statistical distribution of historical processing data and differentiated quality control requirements, the deviation threshold is set as follows: , , Determine if the deviation from the level is true. If it is, then it is judged as a level one deviation; if If it is, it is judged as a level two deviation; if If so, it is judged as a level three deviation; if If the deviation is normal, it is considered normal; from Level 1 to Level 3 deviation, the degree of deviation decreases progressively.

[0119] The comprehensive deviation scores calculated for each processing cycle are arranged in chronological order to form a comprehensive deviation sequence at the most recent time T. The least squares method is used to perform linear regression on the comprehensive deviation sequence to fit a straight line of deviation change over time. The slope of the deviation change line is extracted as a deterioration trend coefficient to detect deterioration risks in advance. The value of T is determined based on the dynamic response time of the process, and is usually taken as 10 to 20 time windows.

[0120] If the deterioration trend coefficient is greater than 0 and the comprehensive deviation score is... In and The predicted time to reach the first level of deviation within the interval is equal to... and The difference is divided by the deterioration trend coefficient; if the predicted time is less than or equal to 5 processing cycles, a first-level warning signal is issued; if the predicted time is less than or equal to 2 processing cycles, a second-level warning signal is issued; where, the processing cycle refers to the time to complete a complete polishing path, which is determined according to the workpiece size and process parameters;

[0121] Ultimately, a real-time processing status is generated, including deviation level and subsurface damage state, enabling early detection of processing deterioration risks.

[0122] Specifically, the steps for iterative optimization include:

[0123] Based on real-time processing status, the state transition function and nonlinear observation function of the process digital twin are encapsulated into a virtual processing environment conforming to the reinforcement learning interaction interface. The state space of the environment is defined by the real-time processing status, specifically including deviation level, subsurface damage exceeding the standard, processing time, and cumulative removal amount. Among them, the subsurface damage exceeding the standard is characterized by the ratio of damage exceeding the standard area, the processing time is a normalized value relative to the total planned processing time, and the cumulative removal amount is the ratio of the current removed material volume to the target removal volume. The action space consists of continuously adjustable multi-dimensional process parameters, including polishing pressure, spindle speed, tool oscillation frequency, and polishing fluid flow rate. The value range of the multi-dimensional process parameters is constrained by the physical limits of the equipment and is normalized. Each time the virtual processing environment receives an action, it calls the state transition function of the process digital twin to predict the surface shape error distribution field and the subsurface damage depth prediction field after one processing cycle, estimates the multi-source sensing characteristics based on the nonlinear observation function, and calculates the new deviation level, subsurface damage exceeding the standard, processing time, and cumulative removal amount to form the next state.

[0124] The reward function consists of three weighted components: The first is the subsurface damage suppression amount, which is the difference between the ratio of the area exceeding the damage limit in the previous period and the ratio in the current period. A positive difference indicates that damage has been suppressed and a positive reward is given. The second is the processing efficiency improvement term, defined as the increment of the cumulative removal amount multiplied by an efficiency coefficient. The efficiency coefficient is determined based on the dimensional matching requirements between the processing efficiency target and damage suppression, and is set to 1.0, ensuring that the efficiency improvement and damage suppression rewards are on similar numerical scales. The third is the parameter adjustment smoothing term, whose value is the negative of the Euclidean distance between adjacent action vectors. This penalizes drastically fluctuating process parameters to avoid equipment shock. The quality priority of the parts and the processing technology target are set with three weights, with values ​​of 0.5, 0.3, and 0.2 respectively, which can be adjusted according to process requirements. The virtual processing environment simultaneously outputs an immediate reward and a termination flag indicating whether the simulation has reached the termination condition. The termination condition includes the total processing time reaching the set time value or the deviation level reaching level one and the deterioration trend coefficient being greater than 0. The set time value is determined based on the total planned processing cycle number and is set to the normalized upper limit of the total planned processing cycle number. Each state transition in the virtual processing environment uses Monte Carlo sampling of the output of the process digital twin to remove the latent variables of the state probability in order to characterize the uncertainty of the prediction, so that the simulation trajectory can reflect multiple possible evolutions.

[0125] Since the SAC algorithm is applicable to continuous action spaces, it is used as the basic framework for reinforcement learning. The specific process of the SAC algorithm includes: constructing a process parameter decision-maker based on reinforcement learning, which includes a stochastic policy network and two Q-value networks; the stochastic policy network takes the current state as input and outputs the mean and logarithmic standard deviation of the actions, obtaining the specific actions through reparameterized sampling; the Q-value networks take state-action pairs as input and extract features through a network body composed of two stacked fully connected layers, where the state and action are concatenated before the first fully connected layer, each fully connected layer is followed by a ReLU activation function, and a linear layer outputs the Q-value. All networks use fully connected layers stacked together, with a hidden layer dimension of 256×256, and the activation function is ReLU, with randomization. The initial parameters of the policy network are obtained by pre-training a set of virtual machining tasks generated by the process digital twin for different optical components and process conditions through meta-reinforcement learning, so as to endow the process parameter decision-maker with the ability to quickly adapt to new workpieces and process conditions for a hot start. An experience replay buffer is initialized to store real interaction data. To support cold start exploration, a prior action generator is constructed. The prior action generator queries similar machining states from the historical best process parameter library based on the current deviation level and the degree of subsurface damage exceeding the standard, and outputs the corresponding action sequence as the cold start benchmark. The historical best process parameter library is constructed from the parameter sequence corresponding to the period with the normal deviation level and the minimum degree of subsurface damage exceeding the standard in the historical successful machining cases.

[0126] At each processing decision point, N parallel inference trajectories are executed to search for the optimal action sequence. The specific search process includes: 80% of the inference trajectories are hot-start inference, i.e., 0.8N parallel inference trajectories. Based on the currently trained random policy network, the state-sampled action sequence is obtained, and the current random policy network is used to directly infer the inference step size based on the state-sampled action sequence. The inference step size is determined according to the dynamic response time of the process system and the control command update cycle, and is 5 to 10 processing cycles. 20% of the inference trajectories are cold-start inference, i.e., 0.2N parallel inference trajectories. The action sequence output by the prior action generator is used and superimposed with decreasing Gaussian exploration noise, and the inference is performed to the same inference step size. Each inference trajectory is recursively accumulated in the virtual processing environment to obtain the discounted total reward for the next H steps, which is used as the evaluation value of each inference trajectory. To further consider the model uncertainty, each step in each inference trajectory is sampled K times by removing the latent variables of the state probability, and the average of the K accumulated evaluation values ​​is taken as the evaluation value of each inference trajectory. To improve the global optimality of parameter optimization among N parallel simulation trajectories, a bat algorithm is introduced for global optimization. The specific calculation process of the bat algorithm includes: using the process parameter sequence corresponding to each simulation trajectory as the initial population, taking it as the initial bat individual position, and initializing the velocity, frequency, loudness, and pulse emission rate; in each iteration, each bat updates the frequency and velocity according to the current global optimal solution, generates a new process parameter sequence, and simulates in a virtual processing environment and samples the latent variables K times to calculate the robust reward; based on fitness comparison and random acceptance criteria, it is decided whether to update the individual position, while increasing the pulse emission rate and decreasing the loudness according to the random acceptance rule, repeating the iteration until the preset maximum number of iterations is reached or the robust reward converges, and finally selecting the action sequence corresponding to the simulation trajectory with the highest robust reward as the optimal process parameter sequence; in this embodiment, K is set to 5, and the preset maximum number of iterations is set to 20 times based on the time constraint of a single control cycle and the simulation calculation power;

[0127] The optimal process parameter sequence is converted into process parameter control instructions using the CNC system protocol mapping and time synchronization module. These instructions are then sent to the equipment's CNC system and actuators via an industrial real-time communication protocol to execute the process parameter control instructions for one machining cycle. During the machining process, multi-source sensor data and actual process parameter feedback values ​​are collected synchronously. After the machining cycle ends, the current workpiece surface shape error and subsurface damage depth status observation values ​​are obtained, forming a real experience tuple and storing it in the experience playback buffer. After execution, feedback data is collected to calculate the comprehensive control deviation, including the root mean square deviation and the maximum depth deviation. The calculation process includes: obtaining the root mean square deviation by subtracting the actual surface shape error from the predicted surface shape error, and calculating the maximum depth deviation by subtracting the actual subsurface damage depth field from the predicted subsurface damage depth field at each grid point. Based on the relative importance of the surface shape error and subsurface damage on the final optical performance of the optical element, a first deviation weight is set with a value of 0.6. Since the sum of the first deviation weight and the second deviation weight is 1, the second deviation weight is set to 0.4. The comprehensive control deviation is calculated by weighted summation of the root mean square deviation and the maximum depth deviation.

[0128] A first deviation threshold is set based on the target surface shape accuracy requirements of the optical element, with a value of λ / 20, where λ is the working wavelength. A second deviation threshold is set based on the allowable subsurface damage tolerance of the process, with a value of 2 micrometers. If the root mean square deviation is greater than the first deviation threshold or the maximum depth deviation is greater than the second deviation threshold, the process digital twin is fine-tuned online using the most recent real experience tuple as a sample. The parameters of the state transition function and nonlinear observation function in the virtual machining environment are updated, and the comprehensive control deviation is fed back to the process parameter decision-maker as a weighting coefficient for iterative optimization. The feedback data includes, but is not limited to, actual polishing pressure, actual spindle speed, actual tool oscillation frequency, actual polishing fluid flow rate, actual surface shape error, and actual subsurface damage depth field.

[0129] After the process digital twin is updated, real experience is sampled from the experience replay buffer according to priority, and combined with the virtual experience generated by the updated virtual processing environment, the gradient of the stochastic policy network and Q-value network of the SAC algorithm is updated. The gradient update step size is adjusted by the comprehensive control deviation, and the larger the comprehensive control deviation, the larger the step size.

[0130] In summary, this invention aims to solve the problems of lag in processing status perception, low processing efficiency, and unstable quality during optical cold polishing. By collecting sensor data, process parameters, surface error, and subsurface damage data, and generating a standardized optical cold polishing dataset after preprocessing, it solves the problem of perception lag caused by traditional detection relying on offline methods and lacking real-time status. Based on the optical cold polishing dataset, a hybrid gated recurrent unit network is constructed to learn in parallel the time-varying residual and damage depth evolution law of removal efficiency. The latent variables of the removal state probability are extracted through cross-modal coupling and conditional variational autoencoder. Meta-learning is used to optimize the initial parameters to form a process digital twin. Using the predicted state of the process digital twin as the initial field, the actual removal function and surface error distribution field are analyzed, and the squared... Root-unscented Kalman filtering achieves recursive Bayesian assimilation, enabling optimal reconstruction and early perception of surface shape errors and subsurface damage depth. When sensors malfunction, assimilation continuity is maintained through probabilistic latent variable inference, effectively avoiding state interruptions. Addressing low processing efficiency and unstable quality, a process digital twin is encapsulated as a virtual processing environment. A process parameter decision-maker is constructed based on reinforcement learning, executing multi-step cold and hot start-up deductions in parallel to output the optimal process parameter sequence. This optimal process parameter sequence is converted into process parameter control commands, and feedback is collected after execution to calculate the comprehensive control deviation. The process digital twin and process parameter decision-maker are updated online, forming a closed-loop iteration. This invention achieves real-time perception of processing status, adaptive optimization and dynamic adjustment of process parameters, improving processing efficiency and quality consistency.

[0131] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing optical cold-working polishing process parameters based on digital twins, characterized in that, include: Collect sensor data, process parameters, surface shape error and subsurface damage data, and preprocess to generate an optical cold processing polishing dataset; A hybrid gated recurrent unit network is constructed to learn the removal efficiency correction and damage depth prediction field. Through latent space fusion and combined with conditional variational autoencoder and meta-learning optimization, a process digital twin is obtained, and the predicted state and removal state probability latent variables are output. Using the predicted state as the initial field for assimilation, the actual removal function parameters and surface error distribution field are analyzed, recursive Bayesian assimilation is performed, the overall surface error and subsurface damage depth are optimally reconstructed, the comprehensive deviation score is calculated to determine the deviation level, and the real-time processing state is generated. The process digital twin is encapsulated as a virtual machining environment. A process parameter decision-maker is constructed and simulated in parallel within the virtual machining environment. The optimal process parameter sequence is output by combining the Bat algorithm and converted into process parameter control instructions. After execution, the comprehensive control deviation is calculated for iterative optimization.

2. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 1, characterized in that, The specific steps for constructing a hybrid gated recurrent unit network include: A hybrid gated recurrent unit network is constructed, including an input layer, a gated recurrent feature learning layer, a cross-modal coupling layer, a probabilistic latent variable extraction layer, and an output layer; The input layer is used to receive the optical cold polishing dataset and extract the process parameter vector and multi-source sensing feature vector; The gated loop feature learning layer is used to execute two parallel gated loop units. The first gated loop unit receives pressure time-series data and speed time-series data, learns the time-varying residual of removal efficiency as process parameters change, and calculates the removal efficiency correction coefficient by combining the Preston equation. The second gated recurrent unit introduces a cross-modal attention mechanism, which obtains a context vector by weighted summation of attention, and connects it with the residual of the current hidden state to obtain an enhanced hidden state. The abrasive passivation features and thermal accumulation features are extracted and concatenated with the enhanced hidden state to obtain a joint feature vector, which is used to control the update gate and reset gate of the second gated loop unit and outputs the damage depth prediction field.

3. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 2, characterized in that, The specific steps for constructing a hybrid gated recurrent unit network also include: A cross-modal coupling layer is used to receive the removal efficiency correction coefficient, generate the actual removal function by combining the current process parameters, and then stack the actual removal function and the damage depth prediction field in the workpiece grid coordinates to form a joint state diagram. The joint latent space representation vector is obtained by encoder mapping. The probabilistic latent variable extraction layer takes the joint latent space representation vector as the input to be encoded and the process parameter vector and multi-source sensor feature vector as the process conditions. The mean and log-variance of the probability distribution of the state latent variables are removed by the output of the conditional variational autoencoder. The state probability latent variables are removed by reparameterized sampling. The decoder reconstructs the joint latent space representation vector and multi-source sensing feature vector based on the removed state probability latent variables and process conditions. The output layer is used to integrate and output the core state of the process digital twin.

4. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 3, characterized in that, The specific steps for outputting the predicted state and removing the latent variables of the state probability include: Using the optical cold polishing dataset as the training set, a meta-learning algorithm is used to calculate the joint loss function through an inner loop, and the task parameters are updated through the Adam optimizer. The outer loop updates the initial parameters of the hybrid gated recurrent unit network based on the gradient aggregation of the updated task parameters, thus obtaining the meta-initial parameters; During deployment, the initial parameters of the target process are fine-tuned using labeled samples to obtain a digital twin of the process. During real-time inference, the current process parameter vector and multi-source sensor feature vector are input, and the predicted state and the removed state probability latent variables are output through the trained hybrid gated recurrent unit network.

5. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 4, characterized in that, The specific steps for analyzing the actual removal function parameters and surface error distribution field include: Extract the actual removal function parameters from the predicted state to define the amount of material removed per unit time at each grid point; The surface shape error distribution field of the current workpiece surface is extracted from the predicted state. The deviation is calibrated by combining the surface shape error data detected offline by the white light interferometer. The calibrated surface shape error distribution field is then interpolated into the workpiece grid coordinate system to form the initial surface shape error state vector. The damage depth prediction field in the predicted state is concatenated with the initial surface error state vector in the same workpiece mesh coordinate system to construct the initial assimilation state vector.

6. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 5, characterized in that, The specific steps for performing recursive Bayesian assimilation include: The mean and logarithmic variance of the state probability latent variables are extracted to construct the error covariance matrix, and the initial square root covariance matrix is ​​obtained by Cholesky decomposition. Force, vibration, and acoustic emission signals are acquired in real time, and after adaptive threshold wavelet denoising, multidimensional statistical features are calculated using a sliding time window to form a multimodal observation vector. Using the initial assimilation state vector and the initial square root of covariance matrix as initial conditions, perform recursive Bayesian assimilation; Sampling points and corresponding mean weights and covariance weights are generated according to the square root unscented transformation rule. Each sampling point is mapped to a priori sampling point through a preset state transition function. The priori state mean is obtained by weighting. The bias matrix is ​​constructed and the prior covariance square root is obtained by QR decomposition and Cholesky rank-one update. Each prior sampling point is mapped to a prior observation sampling point through a preset nonlinear observation function, and the weighted a prior observation mean is obtained. The observation bias matrix is ​​constructed and then subjected to QR decomposition and Cholesky rank-one update to obtain the square root of the observation covariance.

7. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 6, characterized in that, The specific steps for optimal reconstruction include: The cross-covariance matrix between the state and the observation is calculated using the weighted deviation between each prior sampling point and the mean of the prior state, and the weighted deviation between each prior observation sampling point and the mean of the prior observation. The Kalman gain is then calculated by combining the square root of the observation covariance. The observation residual is obtained by calculating the difference between the multimodal observation vector and the prior observation mean. The posterior state estimate is obtained by combining the Kalman gain matrix and the prior state mean. The multimodal observation vectors acquired in real time are assimilated into the state field to perform optimal reconstruction of the overall surface shape error and subsurface damage depth.

8. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 7, characterized in that, The specific steps for determining the deviation level include: Health theory feature values ​​are extracted from the reconstructed multi-source sensing feature vector to replace invalid feature values ​​in the multimodal observation vector, forming a corrected observation vector that re-participates in Kalman gain calculation and state update; The surface shape error and subsurface damage depth are extracted from the posterior state estimate. The root mean square error of the surface shape wavefront is calculated, and the ratio of the damaged area exceeding the standard is calculated. Then, the surface shape deviation and damage deviation are calculated. The comprehensive deviation score is calculated by combining the exponential weighted moving average method. ; Set deviation threshold , , Determine if it deviates from the standard level; like If it is, then it is judged as a level one deviation; if If so, it is judged as a level two deviation; like If so, it is judged as a level three deviation; if If so, it is considered normal.

9. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 8, characterized in that, The specific steps for generating real-time processing status include: The comprehensive deviation scores calculated for each processing cycle are arranged in chronological order to form a comprehensive deviation sequence. Linear regression is performed using the least squares method to fit the deviation change line and extract the slope as the deterioration trend coefficient to detect the risk of deterioration in advance. If the deterioration trend coefficient is greater than 0 and the comprehensive deviation score is within... and The predicted time for the interval to reach level one deviation is... and The difference is divided by the deterioration trend coefficient; If the predicted time is less than or equal to 5 processing cycles, a Level 1 warning signal will be issued; if the predicted time is less than or equal to 2 processing cycles, a Level 2 warning signal will be issued. The final real-time processing status is generated.

10. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 9, characterized in that, The specific steps for outputting the optimal process parameter sequence include: Based on the real-time processing status, the state transition function and nonlinear observation function of the process digital twin are encapsulated into a virtual processing environment; After receiving an action, the virtual processing environment calls the state transition function to predict the surface error distribution field and the damage depth prediction field, calculates the next state, and outputs an instant reward and a termination flag. A process parameter decision-maker consisting of a stochastic policy network and a Q-value network is constructed. The initial parameters of the stochastic policy network are pre-trained through meta-reinforcement learning to initialize the experience replay buffer and construct a prior action generator to output action sequences. At each processing decision moment, N parallel inference trajectories are executed. During hot start, the state sampling action sequence is obtained based on the random policy network. During cold start, Gaussian exploration noise is superimposed on the action sequence. Robust returns are calculated by K samplings after removing latent variables of state probabilities. The process parameter sequences corresponding to each inferred trajectory are used as the initial population. After iterative optimization by the bat algorithm, the optimal process parameter sequence is output.

11. The method for optimizing optical cold-working polishing process parameters based on digital twins according to claim 10, characterized in that, The specific steps for iterative optimization include: The optimal process parameter sequence is converted into process parameter control instructions and sent to the equipment CNC system and actuators. During the processing, multi-source sensor data and actual process parameter feedback values ​​are collected. After the process is completed, the state observation values ​​are obtained, forming real experience tuples and storing them in the experience playback buffer pool. The root mean square deviation and maximum depth deviation are calculated from the collected feedback data, and the weighted sum is used to calculate the overall control deviation. If the root mean square deviation is greater than the preset first deviation threshold or the maximum depth deviation is greater than the preset second deviation threshold, the process digital twin is fine-tuned online using real experience tuples as samples, the parameters of the state transition function and nonlinear observation function are updated, and the comprehensive control deviation is fed back to the process parameter decision-maker for iterative optimization. After the update, real experience is sampled from the experience replay buffer, and the gradient of the stochastic policy network and Q-value network is updated in combination with virtual experience. The gradient update step size is adjusted by using comprehensive control bias.

Images