An adaptive parameter optimization and adjustment system for visual detection

Abstract

This invention relates to the field of semiconductor manufacturing technology and discloses an adaptive parameter optimization and adjustment system for visual inspection, comprising: an image acquisition and feature extraction module; a parameter discretization module; a differential privacy protection module; and a parameter aggregation and optimization module. This invention utilizes a quantum annealing-inspired parameter space discretization technique to map continuous detection parameters to finite discrete energy states. This ensures that differential privacy noise only affects the state transition probability and does not directly affect the parameter values, thereby overcoming the problem of traditional differential privacy protection adding noise to the continuous space, which leads to a severe decrease in detection accuracy. It also solves the contradiction between privacy protection and accuracy maintenance in visual inspection parameter optimization in a distributed environment. Through homomorphically encrypted federated aggregation, each factory can share optimization information without exposing the original parameter gradients.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing, and more specifically, to an adaptive parameter optimization and adjustment system for visual inspection. Background Technology

[0002] In modern semiconductor manufacturing, multiple factories often need to collaborate on wafer quality inspection to improve accuracy and efficiency. Polarization spectroscopy, due to its high sensitivity to material stress states and optical properties, is widely used for inspecting wafer stress defects and transparent protective layers. However, in a distributed manufacturing environment, factories face serious privacy risks when sharing experience in optimizing inspection parameters.

[0003] Specifically, polarization parameters such as degree of polarization and polarization angle distribution directly reflect the physical properties of materials and may reveal unique material formulations and process parameters of each factory. Traditional differential privacy protection methods introduce noise into the continuous parameter space, which severely affects the accuracy of polarization detection because polarization measurements are extremely sensitive to minute changes. Furthermore, the detection equipment and environmental conditions vary from factory to factory, necessitating personalized parameter adaptation while protecting privacy. Summary of the Invention

[0004] This invention provides an adaptive parameter optimization and adjustment system for visual inspection, which solves the technical problem of serious privacy leakage risks faced by factories in a distributed manufacturing environment when sharing inspection parameter optimization experience.

[0005] This invention provides an adaptive parameter optimization and adjustment system for visual detection, comprising:

[0006] Image acquisition and feature extraction module: The wafer is imaged from multiple angles using a polarized light illumination system. Wavelet transform denoising and CLAHE contrast enhancement are used for image preprocessing. Stokes parameter vectors and polarization features are calculated. Finally, feature vectors are generated through wavelet multi-scale analysis.

[0007] Parameter discretization module: Standardizes and preprocesses the detection parameters, constructs an energy function based on Hamiltonian, uses simulated annealing algorithm to discretize the continuous parameter space into a finite set of energy states, and calculates the state transition probability matrix;

[0008] Differential privacy protection module: Calculates the gradient of the performance loss function based on local polarization spectral data, and achieves differential privacy protection through gradient clipping and Laplacian noise addition, while optimizing privacy budget parameters to balance privacy protection and detection performance;

[0009] The parameter aggregation and optimization module uses homomorphic encryption technology to securely aggregate the noisy state transition matrices of each factory, calculates factory weights based on data volume, quality, and historical contributions, obtains the globally optimal parameter configuration through Markov chain iteration, and finally performs local parameter adaptation.

[0010] Furthermore, the specific steps performed in the image acquisition and feature extraction module are as follows:

[0011] S110, Image Acquisition: Each factory uses a polarized light irradiation system to irradiate the wafers and acquire polarized spectral image sequences;

[0012] S120, Image preprocessing: An adaptive threshold denoising algorithm based on wavelet transform is adopted. By setting the noise variance estimate, the threshold of each wavelet sub-band is calculated, and soft thresholding is performed on the wavelet coefficients; the contrast enhancement adopts an adaptive histogram equalization algorithm; the pixel intensity normalization adopts the Z-score normalization method.

[0013] S130, Feature Extraction: Calculate the degree of polarization and polarization angle based on Stokes parameters, and generate a polarization spectral feature vector containing the spatial distribution of the degree of polarization and polarization angle;

[0014] S140, Multi-scale Analysis: Wavelet transform is used to decompose the polarization spectral feature map into multiple scales, and polarization feature components at different scales are output.

[0015] Furthermore, the Stokes parameters include the total light intensity, the difference between the horizontal and vertical polarization components, the difference between the 45-degree and 135-degree polarization components, and the difference between the right-hand and left-hand circular polarization components.

[0016] Furthermore, the specific steps performed in the parameter discretization module are as follows:

[0017] S210, Parameter standardization preprocessing: After standardization preprocessing, the visual detection parameters are mapped to discrete energy state representations;

[0018] S220, Constructing the Hamiltonian: Constructing a parameter-optimized Hamiltonian, with all coefficients scaled uniformly to ensure dimensional consistency;

[0019] S230, Energy State Discretization: Using the simulated annealing algorithm, the continuous parameter space is discretized into... There are 1 energy state, each corresponding to a specific set of detection parameter configurations; the transition probabilities between energy states are calculated to generate... The state transition probability matrix;

[0020] S240, Energy Function Calculation: Quantize the continuous parameter vector into discrete grid points, calculate the energy function for each grid point, and construct the state transition probability based on the Boltzmann distribution, so that the system tends to transition to a lower energy state;

[0021] S250, Time-Aware Loss Function: The detection performance loss function needs to consider the dynamic changes in the time dimension;

[0022] S260, Adaptive Adjustment of Regularization Coefficient: Initialization of Regularization Coefficient Value, regularization coefficient The value is then adaptively adjusted based on the parameter dimension.

[0023] Furthermore, the specific steps executed in the differential privacy protection module are as follows:

[0024] S310, Gradient Calculation and Projection: Based on the local polarization spectral feature vector and the current parameter configuration, calculate the gradient of the detection performance loss function with respect to the parameters;

[0025] S320, Privacy Budget Parameter Optimization: The selection of privacy budget parameters directly affects the balance between the strength of differential privacy protection and data availability;

[0026] S330, Gradient clipping: Limits the gradient norm to a preset threshold and outputs the clipped gradient;

[0027] S340, Gradient Projection Implementation: Project the gradient vector onto the discrete energy state set space, calculate the projection coefficients, normalize the projection coefficients to obtain the transfer tendency, and implement this through Laplace noise perturbation. - Differential privacy protection;

[0028] S350, Energy State Space Projection: Principal component analysis is used to reduce the dimensionality of historical parameter configurations, revealing the main directions of change in the parameter space. The top-order parameters that explain 80% of the variance are selected. The principal components are used as basis vectors.

[0029] Furthermore, the privacy budget parameter is determined using the following adaptive method:

[0030] Establish a privacy-utility evaluation function;

[0031] By using binary search or grid search methods, the value is re-evaluated and adjusted every 20 rounds of optimization iterations, with the rate of change limited to no more than 30% of the previous value, and the maximum value is found in a predetermined candidate set.

[0032] To address changes in data distribution and model accuracy over time, the system reassesses and adjusts privacy budget parameters at fixed intervals.

[0033] Based on the noisy transition tendency vector, update the state transition probability matrix and generate a noisy state transition matrix;

[0034] The transfer tendency vector refers to the degree of transfer preference for each discrete energy state calculated based on the current parameter gradient information. Each component of the vector represents the relative probability weight of transferring to the corresponding energy state.

[0035] Furthermore, the specific steps executed in the parameter aggregation and optimization module are as follows:

[0036] S410, Homomorphic Encryption and Weighted Aggregation: Homomorphic encryption is performed on the noisy state transition matrix to generate ciphertext;

[0037] S420, Factory Weight Calculation: The weight of each factory is calculated using an adaptive multi-factor weighting method;

[0038] S430, Weight Constraints: The weighted aggregation algorithm must satisfy the parameter constraints.

[0039] S440, Homomorphic Encryption Implementation: The specific implementation of homomorphic encryption adopts a lattice-based fully homomorphic encryption scheme, which supports addition and multiplication operations in the ciphertext field and is used for weighted aggregation operations of matrices;

[0040] S450, parameter inverse mapping: The optimized discrete parameter configuration is inversely mapped back to the continuous parameter space to restore the specific visual detection parameter values;

[0041] S460, Adaptive Annealing Scheduling: Design an adaptive annealing scheduler that dynamically adjusts the state transition probability based on the annealing temperature and outputs a parameter configuration sequence that updates iteratively.

[0042] Furthermore, the specific steps for calculating factory weights are as follows:

[0043] Based on the data volume factor, the number of effective samples reflecting the contribution of each factory;

[0044] Based on the data quality factor, the quality level of data from each factory is reflected and evaluated through the signal-to-noise ratio and contrast of polarization spectral images;

[0045] Based on historical contribution factors, reflecting the historical optimization contribution of each factory;

[0046] The final weights are obtained by weighted averaging of the three factors.

[0047] Furthermore, the weight constraints include:

[0048] Weight normalization constraint ensures that each factory has a basic contribution, while preventing a single factory from completely dominating the optimization process.

[0049] Balance the parameter constraints to ensure that all three weighting factors have an appropriate influence;

[0050] Weight smoothing constraints prevent optimization instability caused by sudden weight changes;

[0051] Abnormal factories are excluded from the constraints. If the data quality index is lower than 50% of the group average, or the mean square error of its historical parameter updates exceeds 3 times the group average, its weight is reduced to the minimum value and the weights of other factories are renormalized to prevent abnormal data from polluting the global optimization results.

[0052] Furthermore, the specific steps for implementing homomorphic encryption are as follows:

[0053] The system generates public and private key pairs, and the public key is distributed to each factory by the central server.

[0054] Each factory uses a public key to encrypt the elements of the state transition matrix one by one, allowing for approximate calculations of floating-point numbers;

[0055] After receiving the encryption matrices from each factory, the central server directly calculates the weighted sum in the ciphertext field without needing to decrypt the intermediate results;

[0056] Use private key Decrypt the aggregation results.

[0057] The beneficial effects of this invention are as follows:

[0058] This invention uses a parameter space discretization technique inspired by quantum annealing to map continuous detection parameters to a finite number of discrete energy states. This allows differential privacy noise to affect only the state transition probability and not directly the parameter values, thus overcoming the problem of traditional differential privacy protection that adds noise to the continuous space, which leads to a serious decrease in detection accuracy. It also solves the contradiction between privacy protection and accuracy maintenance in visual detection parameter optimization in a distributed environment.

[0059] Meanwhile, through homomorphic encrypted federated aggregation, each plant can share optimization information without exposing the original parameter gradients, which not only protects the process secrets of each plant, but also achieves the effect of collaborative optimization. Attached Figure Description

[0060] Figure 1 This is a structural block diagram of an adaptive parameter optimization and adjustment system for visual detection proposed in this invention.

[0061] In the diagram: 101, Image Acquisition and Feature Extraction Module; 102, Parameter Discretization Module; 103, Differential Privacy Protection Module; 104, Parameter Aggregation and Optimization Module. Detailed Implementation

[0062] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0063] like Figure 1 As shown, an adaptive parameter optimization and adjustment system for visual detection includes the following modules:

[0064] Image acquisition and feature extraction module 101: The wafer is imaged from multiple angles using a polarized light illumination system. Wavelet transform denoising and CLAHE contrast enhancement are used for image preprocessing. Stokes parameter vector and polarization features are calculated. Finally, feature vectors are generated through wavelet multi-scale analysis.

[0065] The following steps are performed using this module:

[0066] S100 acquires polarization spectral image data from multiple factories and generates polarization spectral feature vectors;

[0067] In one embodiment of the present invention, the following steps are further included:

[0068] S110, Image Acquisition: Each factory uses a polarized light irradiation system to irradiate the wafers and acquire polarized spectral image sequences.

[0069] The original polarization spectrum image is preprocessed, including noise reduction, contrast enhancement, and pixel intensity normalization, to make image data acquired by different devices comparable.

[0070] S120, Image preprocessing: An adaptive threshold denoising algorithm based on wavelet transform is used, by setting a noise variance estimate. (in (For the first-level wavelet detail coefficients), calculate the threshold for each wavelet subband. ( For the first The number of coefficients in each sub-band is counted, and soft thresholding is applied to the wavelet coefficients.

[0071] Contrast enhancement employs the Adaptive Histogram Equalization (CLAHE) algorithm, segmenting the image into... The image is divided into small blocks, and the cumulative distribution function is calculated for each block, with the contrast gain capped at 4 times the standard deviation of the original image.

[0072] Pixel intensity normalization uses the Z-score normalization method, i.e. ,in, These are the original pixel values. To standardize pixel values, and These are the global mean and standard deviation of the image, respectively, and are finally linearly mapped to the range [0, 1].

[0073] This preprocessing process ensures the uniformity and comparability of images acquired by different devices, while preserving key information about polarization characteristics.

[0074] S130, Feature Extraction: Calculating the Stokes parameter vector ,in Indicates total light intensity. This represents the difference between the horizontal and vertical polarization components. This represents the difference between the 45-degree and 135-degree polarization components. This represents the difference between the right-handed and left-handed circular polarization components.

[0075] Calculation of polarization degree based on Stokes parameters and polarization angle The Stokes parameters are all dimensionless normalized values, which makes the degree of polarization... Within the range [0, 1], the polarization angle exist Within the range, a polarization spectral feature vector containing the spatial distribution of polarization degree and polarization angle is generated.

[0076] Furthermore, the polarization angle needs to be specified. exist The range is defined because polarization spectroscopy primarily focuses on the principal axis direction of linear polarization. Using an arctan function with a coefficient of 0.5 ensures a unique representation of the polarization angle within a semicircular range. For cases requiring consideration of the complete polarization direction, this can be achieved by checking... and The sign is used to extend the polarization angle to Scope, i.e. , where arctan2 is the arctangent function in the fourth quadrant.

[0077] Acquiring polarization spectral image data refers to using a multi-angle polarized light illumination system to image a wafer at different polarization angles, thereby obtaining a series of image data containing different polarization information.

[0078] In this embodiment of the application, in order to identify stress modes at different scales, after generating the polarization spectral feature vector, the following steps are also included:

[0079] S140, Multi-scale analysis: Wavelet transform is used to decompose the polarization spectral feature map into multiple scales, and polarization feature components at different scales are output.

[0080] The method of multi-scale decomposition of polarization spectral feature map using wavelet transform takes the polarization spectral feature map as input and outputs a set of wavelet coefficients of multiple scale layers. Each scale layer contains approximation coefficients and detail coefficients, where the approximation coefficients reflect low-frequency polarization features and the detail coefficients reflect high-frequency polarization features.

[0081] Parameter discretization module 102: performs standardized preprocessing on the detection parameters, constructs an energy function based on Hamiltonian, uses simulated annealing algorithm to discretize the continuous parameter space into a finite set of energy states, and calculates the state transition probability matrix;

[0082] The following steps are performed using this module:

[0083] S200 discretizes continuous visual detection parameters into a finite set of energy states, generating a state transition probability matrix;

[0084] In one embodiment of the present invention, the following steps are specifically included:

[0085] S210, Parameter Normalization Preprocessing: After normalization preprocessing, visual detection parameters (including filter parameters, thresholds, weights, etc.) are mapped to discrete energy state representations.

[0086] Normalization is performed on different types of parameters: for filter kernel parameters, the following is used: Norm normalization Where p is the original parameter vector; for the threshold parameter, a minimum-maximum scaling is applied to the [0, 1] interval. ,in and These are the minimum and maximum values ​​of the parameter, respectively; the weight parameters are standardized using standard deviation. ,in The mean, This is the standard deviation. This ensures that parameters with different dimensions and numerical ranges can be optimized in a unified discrete space.

[0087] S220, Constructing the Hamiltonian: Constructing a parameter-optimized Hamiltonian. ,in Indicates the first The state of each parameter (mapped to) (binary variables) Represents the coefficient of the single-parameter energy term. This represents the coupling coefficients between parameters. All coefficients are uniformly scaled to ensure dimensional consistency.

[0088] S230, Energy State Discretization: Using the simulated annealing algorithm, the continuous parameter space is discretized into... There are 1 energy state, each corresponding to a specific set of detection parameter configurations; calculate the transition probability between energy states. ,in This is the energy difference (dimensionless value). The annealing temperature (also a dimensionless normalized value) produces The state transition probability matrix.

[0089] in, Let i be the energy difference between energy states i and j. Let be the probability of transitioning from energy state i to j;

[0090] It should be noted that an energy state refers to a state point in a discrete space that maps a set of parameter configurations in a continuous parameter space. Each state point has a specific energy value, which reflects the optimization objective function value of the corresponding parameter configuration.

[0091] Discretize the continuous parameter space as The simulated annealing algorithm for each energy state takes the parameter-optimized Hamiltonian and the initial temperature as inputs, and outputs a discretized... A set of energy states, each containing a corresponding parameter configuration vector and energy value.

[0092] The quantum annealing-inspired parameter space discretization method, which discretizes visual detection parameters into a finite set of energy states, is implemented through the following sub-steps:

[0093] S240, Energy Function Calculation: Calculating a Continuous Parameter Vector Quantized into discrete grid points, the energy function is calculated for each grid point. ,in To detect the performance loss function, , Is it using parameters? For the Prediction results for each sample It is the first The true label of each sample It is the total number of samples; For regularization terms, Combining L1 and L2 regularization to balance sparsity and smoothness The L2 regularization coefficient;

[0094] State transition probabilities are constructed based on Boltzmann distribution, making the system tend to transition to lower energy states.

[0095] S250, Time-Aware Loss Function: The detection performance loss function needs to consider the dynamic changes over time. Specifically, the loss function is modified as follows: ,in The time window length (set to the most recent 30 production batches). For time points The number of samples, and They are time points respectively The The predicted results and true labels for each sample. The time decay weighting function is... The decay coefficient (set to 0.7) gives more weight to recent samples.

[0096] Time-aware loss function design can adapt to the slow drift of process parameters and material properties, making the optimized parameters more sensitive to the latest production status and improving the system's adaptability.

[0097] Furthermore, the regularization coefficient in the energy function and The quality of parameter optimization results has a significant impact and needs to be set appropriately according to the specific task characteristics.

[0098] In this embodiment, these coefficients are determined using an automatic tuning method based on cross-validation.

[0099] S260, Adaptive adjustment of regularization coefficient: Initialization Value set to ,in For the initial parameter configuration, this setting ensures that the initial contribution of the regularization term is approximately 1% of the loss function, guaranteeing that the optimization process primarily focuses on performance objectives. The value is determined by the parameter dimension. Perform adaptive adjustments: ,in The desired number of sparse parameters (usually set to 20%-40% of the total number of parameters) is used to avoid excessively penalizing non-zero parameters while maintaining appropriate sparsity.

[0100] During the optimization process, adjustments are made dynamically based on the performance of the validation set after each preset validation cycle. Value: Performance degrades after three consecutive verifications. ,but Reduce to 80% of the original; if performance improves after three consecutive verifications, then The value is increased to 120% of the original value, thereby achieving adaptive control of the regularization strength.

[0101] Differential privacy protection module 103: Calculates the gradient of the performance loss function based on local polarization spectral data, and achieves differential privacy protection by gradient clipping and Laplacian noise addition, while optimizing the privacy budget parameters to balance privacy protection and detection performance;

[0102] The following steps are performed using this module:

[0103] S300 calculates the gradient based on local polarization spectral data and applies differential privacy to generate a noisy state transition matrix;

[0104] In one embodiment of the present invention, the following steps are specifically included:

[0105] S310, Gradient Calculation and Projection: Calculating the detection performance loss function based on the local polarization spectral feature vector and the current parameter configuration. gradient with respect to parameters .

[0106] Projecting the gradient vector onto the discrete energy state set space generates a state transition tendency vector. ,in Indicates to the first The tendency of energy state transition.

[0107] To apply differential privacy protection, Laplace noise is added to the transition bias vector: ,in The scale parameter is represented as The Laplace distribution random sampling function, Sensitivity (the maximum possible change in the gradient). Budget for privacy.

[0108] S320, Privacy Budget Parameter Optimization: Privacy Budget Parameters The choice of [aspect name] directly affects the balance between the strength of differential privacy protection and data availability. In this embodiment, The range of values ​​is limited to Within this range, smaller values ​​offer stronger privacy protection but introduce more noise, while larger values ​​retain more data usefulness but offer relatively weaker privacy protection.

[0109] Specific The value is determined using the following adaptive method:

[0110] Establish a privacy-utility evaluation function:

[0111] ;

[0112] in The model accuracy measured on the validation set (normalized to) scope), For privacy breach risk assessment function (defined as) , The risk factor is set to 2. To balance parameters (set according to factory privacy protection requirements) (within the range)

[0113] Using binary search or grid search methods, the value is re-evaluated and adjusted every 20 rounds of optimization iterations, with the rate of change limited to no more than 30% of the previous value, within a predetermined range. Finding the maximum value in the candidate set The value;

[0114] To address changes in data distribution and model accuracy over time, the system re-evaluates and adjusts at fixed intervals (e.g., 20 optimization iterations). Values ​​are used to maintain a dynamic balance between privacy protection and utility.

[0115] In addition, the system also has settings The rate of change limit, i.e., the percentage of each adjustment. The value should not exceed 30% of the previous value to avoid abrupt changes in privacy strength. This is based on a noisy transfer tendency vector. Update the state transition probability matrix and generate a noisy state transition matrix. .

[0116] It should be noted that the transfer tendency vector refers to the degree of transfer preference for each discrete energy state calculated based on the current parameter gradient information. Each component of this vector represents the relative probability weight of transferring to the corresponding energy state.

[0117] S330, Applying Gradient Clipping: Limiting the gradient norm to a preset threshold. within, that is Output the clipped gradient;

[0118] Furthermore, gradient clipping threshold The choice of threshold is crucial for the effectiveness of differential privacy protection. In this embodiment, the threshold is determined based on the statistical characteristics of historical gradients. Specifically, the threshold is selected by collecting the distribution of gradient norms across multiple iterations. It is the 95th quantile of the gradient norm distribution, i.e. ,in Indicates the first The gradient of the next iteration. This represents the number of historical iterations.

[0119] This approach ensures that most normal gradients are not over-pruned, while effectively limiting the impact of abnormally large gradients on the differential privacy mechanism, keeping the amount of added noise within a reasonable range.

[0120] The aforementioned method for calculating gradients based on local polarization spectral data and applying differential privacy to generate a noisy state transition matrix is ​​implemented through the following steps:

[0121] S340, Gradient Projection Implementation: Transforming the gradient vector... Project onto the discrete energy state set space and calculate the projection coefficients. ,in For the first The direction vector of each energy state;

[0122] Normalized projection coefficients show a tendency to shift. ,in For temperature parameters, It is a function that converts the projection coefficients into a probability distribution;

[0123] Achieved through Laplace noise perturbation - Differential privacy protection.

[0124] S350, Energy State Space Projection: Dimensionality reduction of historical parameter configurations is achieved through Principal Component Analysis (PCA) to obtain the main directions of change in the parameter space, selecting the top-order parameters that explain 80% of the variance. The principal components are used as basis vectors;

[0125] For each discrete energy state Calculate the difference vector between its parameter configuration and the parameter configuration of its adjacent energy states. ,in For the reference state parameter configuration, the current state or the globally optimal state is usually selected, and the difference vector is projected onto the principal component space to obtain the direction vector. ,in This is the PCA principal component matrix, and Normalize is the vector normalization operation.

[0126] Calculate the projection coefficients of the gradient vector in each direction. The negative sign ensures that the gradient direction is the energy descent direction.

[0127] Temperature parameters Defined as a function of the standard deviation of the projection coefficients: ,in Let the standard deviation of all projection coefficients be . Minimum temperature (set to 0.05).

[0128] Parameter aggregation and optimization module 104: It uses homomorphic encryption technology to securely aggregate the noisy state transition matrices of each factory, calculates the factory weights based on data volume, quality and historical contribution, obtains the globally optimal parameter configuration through Markov chain iteration, and finally performs local parameter adaptation.

[0129] The following steps are performed using this module:

[0130] S400 aggregates the noisy state transition matrices of multiple plants and outputs global optimization parameter configurations;

[0131] In one embodiment of the present invention, the following steps are specifically included:

[0132] S410, Homomorphic Encryption and Weighted Aggregation: Transforming Noisy State Transition Matrices Perform homomorphic encryption to generate ciphertext. .

[0133] The central server performs weighted aggregation operations in the encrypted field:

[0134] ;

[0135] in For the first The weight of each factory, satisfying .

[0136] S420, Factory Weight Calculation: Weight of Each Factory The calculation employs an adaptive multi-factor weighting method to ensure a scientific and reasonable weight allocation. The specific implementation includes three key factors:

[0137] Based on data volume factor The effective sample size reflecting the contribution of each factory is calculated using the following formula: ,in For the first Number of valid samples per factory;

[0138] Based on data quality factors This reflects the quality level of data from each factory, and is evaluated through the signal-to-noise ratio and contrast of polarization spectral images. The calculation formula is as follows: ,in For the first Average signal-to-noise ratio of each factory image Average contrast;

[0139] Based on historical contribution factors This reflects the historical optimization contribution of each factory, and the calculation formula is as follows: ,in For the first Mean squared error of historical parameter updates for each factory This is the regularization coefficient (set to 5).

[0140] The final weights are obtained by weighted averaging of the three factors:

[0141] ;

[0142] in , This is the balance coefficient.

[0143] This method ensures that factories with large amounts of high-quality data and good historical contributions receive higher weights, making global parameter optimization more effective.

[0144] S430, Weight Constraints: The weighted aggregation algorithm needs to satisfy a series of parameter constraints to ensure the stability and effectiveness of the optimization process. These constraints include:

[0145] (1) Weight normalization constraint: and ,in The number of factories participating in federal optimization. The minimum weight threshold (set to 0.05) is used to ensure that each factory has a basic contribution, while preventing a single factory from completely dominating the optimization process.

[0146] (2) Balance parameter constraints: and This ensures that all three weighting factors have appropriate influence.

[0147] (3) Weight smoothing constraint: ,in Indicates the first In the first iteration The weight of each factory, Set the maximum allowable change (to 0.15) to avoid optimization instability caused by sudden weight changes;

[0148] (4) Abnormal Factory Exclusion Constraints: If the data quality indicators of a certain factory are abnormal Less than 50% of the population average, or the mean squared error of its historical parameter updates. If the value exceeds three times the group average, its weight is reduced to the minimum. The weights of other factories are then renormalized to prevent outlier data from contaminating the global optimization results. These constraints work together to ensure the robustness and fairness of the weighted aggregation process, effectively balancing the contributions of each factory in the federated optimization.

[0149] S440, Homomorphic Encryption Implementation: The specific implementation of homomorphic encryption adopts a lattice-based fully homomorphic encryption (FHE) scheme. This scheme supports addition and multiplication operations in the ciphertext field, and is particularly suitable for weighted aggregation operations on matrices. The specific implementation steps include:

[0150] The system generates public and private key pairs. The public key is distributed to each factory by the central server;

[0151] Each factory uses public keys For the state transition matrix The elements are encrypted one by one, that is... It adopts the CKKS (Cheon-Kim-Kim-Song) scheme, which allows for approximate calculations of floating-point numbers and supports batch encryption to improve efficiency;

[0152] After receiving the encryption matrices from each factory, the central server directly calculates the weighted sum in the ciphertext field. No need to decrypt intermediate results;

[0153] Use private key Decrypt the aggregation result:

[0154] .

[0155] To ensure numerical accuracy, the parameters of the CKKS scheme are configured as follows: polynomial degree selection The modulus size is 128 bits, ensuring 40 significant decimal places of precision and 128-bit security strength, which is sufficient to meet the precision requirements of the state transition matrix elements. Decryption yields the global state transition matrix. .

[0156] Based on the global state transition matrix The Markov chain Monte Carlo method is used to identify steady-state distributions. The parameter configuration corresponding to the energy state with the highest probability in the steady-state distribution is the global optimal parameter configuration.

[0157] S450, Parameter Inverse Mapping: This involves mapping the optimized discrete parameter configuration back to the continuous parameter space, restoring it to specific visual detection parameter values. For binarized parameter state variables... The parameters are converted into actual parameter values ​​according to their mapping rules; for standardized parameters, destandardization is performed (e.g., multiplying by the original standard deviation and adding the original mean); for normalized parameters, denormalization is performed to restore them to their original numerical range and physical dimensions.

[0158] Furthermore, the specific implementation method for mapping parameters back from discrete energy states to continuous parameter space is as follows:

[0159] Identifying the optimal energy state from the steady-state distribution of a Markov chain. and its corresponding discrete parameter configuration ;

[0160] Specific inverse mapping processing is used for different types of parameters:

[0161] For parameters of the binary encoding, the sign function is used to restore them:

[0162] ;

[0163] in It is a binary state variable. and These are the minimum and maximum valid values ​​of the parameter, respectively.

[0164] For the standardized weight parameters, apply the inverse Z-score transform:

[0165] ;

[0166] in and It is the first The original standard deviation and mean of each parameter before standardization;

[0167] for Normalized filter kernel parameters are then denormalized:

[0168] ;

[0169] in It is the scaling factor of the original filter kernel;

[0170] For the threshold parameter of minimum-maximum scaling, perform an inverse linear transform:

[0171] ;

[0172] To ensure the physical validity of the inverse mapping parameters, additional constraints are applied to specific parameter types, such as ensuring that the sum of the convolution kernel elements is 1 and that the bias term is within a valid numerical range. Finally, progressive simulation validation (first on a small sample, then on the full dataset) ensures that the inverse mapping parameters can work correctly in a real detection system.

[0173] S460, Adaptive Annealing Scheduling: Design of an Adaptive Annealing Scheduling System ,in The initial temperature, For annealing rate, The iteration number is used to dynamically adjust the state transition probability based on the annealing temperature, and the output is a sequence of parameter configurations that are updated with each iteration.

[0174] Adaptive annealing scheduling parameters and Making the right choice is crucial for optimizing efficiency.

[0175] In this embodiment, the initial temperature It should be configured to allow the system sufficient state exploration capability in the initial stage, determined based on the statistical properties of the energy function: ,in Let the standard deviation of the energy values ​​for all energy states be denoted as . This represents the total number of energy states. This setting ensures that there are approximately [number missing] energy states in the initial stage. The probability of receiving energy increases The state transition.

[0176] Annealing rate Then optimize the number of iterations as expected. set up: ,in The desired final temperature (usually set to) (0.1%-1%), to ensure that the temperature can be reduced to the target value within the expected number of iterations, thus ensuring algorithm convergence and avoiding local optimum traps caused by excessively rapid cooling.

[0177] S470, Markov Chain Convergence Control: A Monte Carlo Method for Identifying Steady-State Distributions of Markov Chains Using the Global State Transition Matrix As input, a Markov chain is generated through iterative sampling, and the output is a steady-state distribution. ,in Indicates that the system is at the th The steady-state probability of each energy state. Specifically, the steady-state distribution is calculated using a power-law iteration method, i.e., repeated execution. ,until (in (As the convergence threshold), the final result is This is the steady-state distribution of a Markov chain.

[0178] Furthermore, to ensure the reliability and efficiency of the Markov chain iterative process, robust termination conditions need to be established. Specifically, the iterative process terminates when any of the following conditions are met:

[0179] Convergence condition: ,in It is the L1 norm (the sum of the absolute values ​​of all elements). This is the convergence threshold for the current iteration;

[0180] Maximum number of iterations limit: ,in Set the maximum number of iterations to a preset value (10000 to prevent infinite loops).

[0181] Steady-state dominance condition: and ,in For the first The highest steady-state probability in the next iteration. The threshold for the maximum probability change in continuous iterations (set as follows): ), The steady-state probability threshold (set to 0.6) indicates that a certain energy state has already become dominant.

[0182] (4) Oscillation detection conditions: If continuous The next iteration ( A cyclic oscillation mode was found in (set to 50), that is ( The oscillation detection threshold is set to [value]. If the average value of the state distribution in the cyclic sequence is taken as the steady-state distribution, then the system simultaneously tracks the changes in the global optimal energy state during the iteration process so as to output the best result if the iteration terminates prematurely.

[0183] Furthermore, the convergence threshold of the Markov chain The selection of the threshold should ensure a balance between the computational accuracy and computational efficiency of the steady-state distribution. In this embodiment, an adaptive convergence threshold determination method is adopted, with initial settings... The threshold is dynamically adjusted as the number of iterations increases.

[0184] ;

[0185] in This represents the current iteration number. This represents the floor function.

[0186] This design allows for faster convergence in the initial stages, while increasing accuracy requirements as the system approaches steady state. Simultaneously, a minimum threshold is set. Ensure the stability of numerical calculations and avoid infinite iterations caused by floating-point precision limitations.

[0187] S480, Local Parameter Adaptation: Each factory fine-tunes the parameters based on the local equipment calibration parameters within the neighborhood of the globally optimized parameters to generate final detection parameters adapted to the local environment.

[0188] Specifically, the fine-tuning process includes:

[0189] The global optimization parameters are applied to the local polarization spectroscopy detection system to collect the detection results of local standard samples;

[0190] Calculate the local loss function value and construct a small-scale grid search region in the parameter space; then, evaluate the performance indicators at the grid points and select the optimal local parameter configuration.

[0191] The optimal local parameters are applied to the actual detection system, and the output execution parameters suitable for the local environment characteristics are generated, including specific filter parameter settings, detection threshold configuration and algorithm weight values.

[0192] Based on the actions performed by each module in the aforementioned adaptive parameter optimization and adjustment system for visual detection, the following example is proposed:

[0193] This example describes how, during the integrated circuit manufacturing process, five wafer foundries located in different regions (denoted as foundries A, B, C, D, and E) jointly optimize the polarization spectral detection parameters of the silicon nitride passivation layer on a 12-inch silicon wafer.

[0194] Each factory uses the same model but different instances of polarization spectroscopy detection equipment. They need to jointly optimize the detection parameters to improve the detection rate of silicon nitride layer defects on wafers, while protecting the privacy of their respective process parameters.

[0195] Polarization spectral feature vector generation: Polarization spectral images of wafers are collected from each factory. High-resolution images of each wafer are obtained at four polarization angles (0°, 45°, 90°, 135°). After preprocessing, Stokes parameters and polarization features are calculated.

[0196] According to the wavelet transform multi-scale decomposition described in S110, the analysis results show that at scale 3 (corresponding to an 8×8 pixel area), the standard deviation of the high-frequency coefficients of the defective samples is 230 times higher than that of the normal samples, making it the most significant feature identifier. The polarization spectral feature vector of each factory has a dimension of 512. After wavelet transform, four scale layers are generated, each containing approximation coefficients and detail coefficients, forming a complete polarization feature representation, as shown in Table 1.

[0197] Table 1: Characteristics of Polarization Spectrum Data Samples from Each Factory

[0198]

[0199] Visual detection parameter discretization: 68 continuous parameters (including 21 filter kernel parameters, 15 threshold parameters, and 32 weight parameters) are discretized and mapped to 32 discrete energy states, as shown in Table 2;

[0200] Table 2: Examples of Discretization Results for Visual Inspection Parameters

[0201]

[0202] Based on the parameter space discretization method inspired by quantum annealing of S200, this example uses a temperature parameter of 0.8, 1000 simulated annealing iterations, and energy function regularization coefficients λ=0.015 and β=0.45. L1 and L2 regularizations are combined to balance sparsity and smoothness. The calculated average transition probability between energy states is 0.032, forming a complete 32×32 state transition probability matrix.

[0203] Differential privacy and state transition matrix generation: Following the S300 method, each factory calculates the gradient of the detection performance loss function based on the local polarization spectral feature vector and the current parameter configuration. The gradient vector is projected onto the discrete energy state set space to generate a state transition tendency vector. Then, differential privacy protection is applied to generate a noisy state transition matrix, as shown in Table 3, which takes factory A as an example.

[0204] Table 3: Example of a noisy state transition matrix for factory A (partial data)

[0205]

[0206] Following the gradient pruning method described in S310, the gradient norm is restricted to a preset threshold C=3.5, the differential privacy parameter ε is 1.8 (range [0.1, 10]), and the standard deviation of the added Laplacian noise is 0.23. In the adaptive privacy-utility evaluation function, the balance parameter α is set to 0.6, and the risk coefficient k is set to 2.

[0207] Verification tests show that, at this level of privacy, the success rate of parameter recipe inference attacks is less than 3%, while maintaining the effectiveness of parameter optimization.

[0208] Multi-factory data aggregation and parameter optimization: The noisy state transition matrix for each factory in the S400 architecture is encrypted using lattice-based fully homomorphic encryption (FHE). Specifically, the CKKS (Cheon-Kim-Kim-Song) scheme is adopted, which allows for approximate calculations of floating-point numbers, with the polynomial degree set to 2. 15 The modulus size is 128 bits. The central server performs weighted aggregation operations in the ciphertext field, as shown in Table 4;

[0209] Table 4: Example of factory weight calculation (based on adaptive multi-factor weighting method)

[0210]

[0211] Using the Markov chain Monte Carlo method described in S400, the system converged after 2567 iterations (convergence threshold δ = 10). -6 The steady-state distribution shows that the E3 energy state has the highest probability (0.214), and the corresponding parameter configuration is selected as the global optimization result.

[0212] Adaptive annealing scheduling using S410 was adopted, with the initial temperature T0 set to 5.3°C and the annealing rate α to 0.002, and the expected final temperature to be 0.5% of the initial temperature. Subsequently, each plant fine-tuned the parameters within the neighborhood of the global optimization parameters according to the S420 method, generating the final locally adapted application parameters, as shown in Table 5.

[0213] Table 5: Global Optimization Parameters and Local Adaptation Results (Some Key Parameters)

[0214]

[0215] Through local adaptation, each factory successfully adjusted its global optimization parameters to the optimal operating point suitable for its own equipment characteristics. The local fine-tuning process constructed a grid search within a ±5% range in the parameter space, evaluated performance based on local standard samples, and determined the optimal local configuration. Practice has demonstrated that the parameter optimization method applied in this example effectively balances privacy protection and detection performance, improving the defect detection rate by 15.3% (from 78.6% to 93.9%) and reducing the false positive rate by 4.2% (from 6.8% to 2.6%), while ensuring the security of process parameters in each factory and successfully responding to parameter recipe inference attack tests.

[0216] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention, all of which are within the protection scope of the present invention.

Claims

1. An adaptive parameter optimization and adjustment system for visual detection, characterized in that, include: Image acquisition and feature extraction module: The wafer is imaged from multiple angles using a polarized light illumination system. Wavelet transform denoising and CLAHE contrast enhancement are used for image preprocessing. Stokes parameter vectors and polarization features are calculated. Finally, feature vectors are generated through wavelet multi-scale analysis. Parameter discretization module: Standardizes and preprocesses the detection parameters, constructs an energy function based on Hamiltonian, uses simulated annealing algorithm to discretize the continuous parameter space into a finite set of energy states, and calculates the state transition probability matrix; Differential privacy protection module: Calculates the gradient of the performance loss function based on local polarization spectral data, and achieves differential privacy protection through gradient clipping and Laplacian noise addition, while optimizing privacy budget parameters to balance privacy protection and detection performance; The parameter aggregation and optimization module uses homomorphic encryption technology to securely aggregate the noisy state transition matrices of each factory, calculates factory weights based on data volume, quality, and historical contributions, obtains the globally optimal parameter configuration through Markov chain iteration, and finally performs local parameter adaptation.

2. The adaptive parameter optimization and adjustment system for visual detection according to claim 1, characterized in that, The specific steps performed in the image acquisition and feature extraction module are as follows: S110, Image Acquisition: Each factory uses a polarized light irradiation system to irradiate the wafers and acquire polarized spectral image sequences; S120, Image preprocessing: An adaptive threshold denoising algorithm based on wavelet transform is adopted. By setting the noise variance estimate, the threshold of each wavelet sub-band is calculated, and soft thresholding is performed on the wavelet coefficients; the contrast enhancement adopts an adaptive histogram equalization algorithm; the pixel intensity normalization adopts the Z-score normalization method. S130, Feature Extraction: Calculate the degree of polarization and polarization angle based on Stokes parameters, and generate a polarization spectral feature vector containing the spatial distribution of the degree of polarization and polarization angle; S140, Multi-scale Analysis: Wavelet transform is used to decompose the polarization spectral feature map into multiple scales, and polarization feature components at different scales are output.

3. The adaptive parameter optimization and adjustment system for visual detection according to claim 2, characterized in that, The Stokes parameters include the total light intensity, the difference between the horizontal and vertical polarization components, the difference between the 45-degree and 135-degree polarization components, and the difference between the right-hand and left-hand circular polarization components.

4. The adaptive parameter optimization and adjustment system for visual detection according to claim 1, characterized in that, The specific steps performed in the parameter discretization module are as follows: S210, Parameter standardization preprocessing: After standardization preprocessing, the visual detection parameters are mapped to discrete energy state representations; S220, Constructing the Hamiltonian: Constructing a parameter-optimized Hamiltonian, with all coefficients scaled uniformly to ensure dimensional consistency; S230, Energy State Discretization: Using the simulated annealing algorithm, the continuous parameter space is discretized into... There are 1 energy state, each corresponding to a specific set of detection parameter configurations; the transition probabilities between energy states are calculated to generate... The state transition probability matrix; S240, Energy Function Calculation: Quantize the continuous parameter vector into discrete grid points, calculate the energy function for each grid point, and construct the state transition probability based on the Boltzmann distribution, so that the system tends to transition to a lower energy state; S250, Time-Aware Loss Function: The detection performance loss function needs to consider the dynamic changes in the time dimension; S260, Adaptive Adjustment of Regularization Coefficient: Initialization of Regularization Coefficient Value, regularization coefficient The value is then adaptively adjusted based on the parameter dimension.

5. The adaptive parameter optimization and adjustment system for visual detection according to claim 1, characterized in that, The specific steps executed in the differential privacy protection module are as follows: S310, Gradient Calculation and Projection: Based on the local polarization spectral feature vector and the current parameter configuration, calculate the gradient of the detection performance loss function with respect to the parameters; S320, Privacy Budget Parameter Optimization: The selection of privacy budget parameters directly affects the balance between the strength of differential privacy protection and data availability; S330, Gradient clipping: Limits the gradient norm to a preset threshold and outputs the clipped gradient; S340, Gradient Projection Implementation: Project the gradient vector onto the discrete energy state set space, calculate the projection coefficients, normalize the projection coefficients to obtain the transfer tendency, and implement this through Laplace noise perturbation. - Differential privacy protection; S350, Energy State Space Projection: Principal component analysis is used to reduce the dimensionality of historical parameter configurations, revealing the main directions of change in the parameter space. The top-order parameters that explain 80% of the variance are selected. The principal components are used as basis vectors.

6. The adaptive parameter optimization and adjustment system for visual detection according to claim 5, characterized in that, The privacy budget parameter is determined using the following adaptive method: Establish a privacy-utility evaluation function; By using binary search or grid search methods, the value is re-evaluated and adjusted every 20 rounds of optimization iterations, with the rate of change limited to no more than 30% of the previous value, and the maximum value is found in a predetermined candidate set. To address changes in data distribution and model accuracy over time, the system reassesses and adjusts privacy budget parameters at fixed intervals. Based on the noisy transition tendency vector, update the state transition probability matrix and generate a noisy state transition matrix; The transfer tendency vector refers to the degree of transfer preference for each discrete energy state calculated based on the current parameter gradient information. Each component of the vector represents the relative probability weight of transferring to the corresponding energy state.

7. The adaptive parameter optimization and adjustment system for visual detection according to claim 1, characterized in that, The specific steps performed in the parameter aggregation and optimization module are as follows: S410, Homomorphic Encryption and Weighted Aggregation: Homomorphic encryption is performed on the noisy state transition matrix to generate ciphertext; S420, Factory Weight Calculation: The weight of each factory is calculated using an adaptive multi-factor weighting method; S430, Weight Constraints: The weighted aggregation algorithm must satisfy the parameter constraints. S440, Homomorphic Encryption Implementation: The specific implementation of homomorphic encryption adopts a lattice-based fully homomorphic encryption scheme, which supports addition and multiplication operations in the ciphertext field and is used for weighted aggregation operations of matrices; S450, parameter inverse mapping: The optimized discrete parameter configuration is inversely mapped back to the continuous parameter space to restore the specific visual detection parameter values; S460, Adaptive Annealing Scheduling: Design an adaptive annealing scheduler that dynamically adjusts the state transition probability based on the annealing temperature and outputs a parameter configuration sequence that updates iteratively.

8. The adaptive parameter optimization and adjustment system for visual detection according to claim 7, characterized in that, The specific steps for calculating factory weights are as follows: Based on the data volume factor, the number of effective samples reflecting the contribution of each factory; Based on the data quality factor, the quality level of data from each factory is reflected and evaluated through the signal-to-noise ratio and contrast of polarization spectral images; Based on historical contribution factors, reflecting the historical optimization contribution of each factory; The final weights are obtained by weighted averaging of the three factors.

9. The adaptive parameter optimization and adjustment system for visual detection according to claim 8, characterized in that, Weight constraints include: Weight normalization constraint ensures that each factory has a basic contribution, while preventing a single factory from completely dominating the optimization process. Balance the parameter constraints to ensure that all three weighting factors have an appropriate influence; Weight smoothing constraints prevent optimization instability caused by sudden weight changes; Abnormal factories are excluded from the constraints. If the data quality index is lower than 50% of the group average, or the mean square error of its historical parameter updates exceeds 3 times the group average, its weight is reduced to the minimum value and the weights of other factories are renormalized to prevent abnormal data from polluting the global optimization results.

10. The adaptive parameter optimization and adjustment system for visual detection according to claim 9, characterized in that, The specific steps for implementing homomorphic encryption are as follows: The system generates public and private key pairs, and the public key is distributed to each factory by the central server. Each factory uses a public key to encrypt the elements of the state transition matrix one by one, allowing for approximate calculations of floating-point numbers; After receiving the encryption matrices from each factory, the central server directly calculates the weighted sum in the ciphertext field without needing to decrypt the intermediate results; Use private key Decrypt the aggregation results.

Images