Intelligent coating optimization method for quartz crucible surface treatment

Abstract

The invention belongs to the technical field of crucible coating optimization, and relates to an intelligent coating optimization method for quartz crucible surface treatment. According to the method, coating material data, surface treatment environment data, coating process data and performance test data are collected, the performance test data serve as a real result label of a training set, denoising, missing value processing and normalization are carried out on the collected data, and through a Pearson's correlation coefficient and a principal component analysis method, the coating material data, the surface treatment environment data, the coating process data and the performance test data are obtained. According to the method, a group of key features are screened out, data dimensions are reduced, main information influencing the coating performance is reserved, a deep neural network model is trained, the coating performance is predicted through the model, and coating process parameters are optimized by adopting a particle swarm optimization algorithm based on the predicted coating performance, so that the coating performance is optimized. According to the method, the optimal parameter combination can be effectively searched, blindness and high cost of a traditional optimization method are avoided, a closed-loop intelligent coating optimization process is formed, and the optimization efficiency and performance of the coating process are remarkably improved.

Description

Technical Field

[0001] The present application belongs to the technical field of crucible coating optimization, and more specifically, relates to an intelligent coating optimization method for surface treatment of a quartz crucible. Background Art

[0002] At present, the production and manufacturing of quartz crucibles mainly adopts the high-temperature melting method, that is, high-purity quartz sand is melted at high temperature, and then a crucible with a certain shape and size is made through a specific molding process, such as centrifugal casting, hot die casting, etc.; during the manufacturing process, in order to improve the performance and life of the crucible, it is usually necessary to coat the surface of the crucible.

[0003] The specific implementation steps of the coating process are as follows:

[0004] Cleaning and pretreatment: First, clean the surface of the quartz crucible to remove oil, dust and other impurities on the surface, and then dry it.

[0005] Coating material selection: Select appropriate coating materials, such as ceramics, metals, oxides, etc., according to the use environment and performance requirements of the crucible.

[0006] Coating construction: Use spraying, dipping, brushing and other methods to evenly coat the coating material on the crucible surface. Parameters such as coating thickness, uniformity, temperature gradient, spraying rate and angle have an important influence on coating performance.

[0007] Drying and curing: After coating construction, the crucible is dried and cured to make the coating material closely bonded to the crucible surface and to increase the service life of the coating.

[0008] However, the current optimization of coating processes relies on engineers' experience and experimental data, which lacks system and objectivity. Since the optimization of coating processes is often based on personal experience, it is highly subjective and lacks objective evaluation criteria. Therefore, it is difficult to ensure that the results of each optimization can meet the best performance requirements. Summary of the invention

[0009] The present invention provides an intelligent coating optimization method for quartz crucible surface treatment, which is intended to solve the technical problem that the current coating optimization based on personal experience causes each optimization result to be difficult to meet the optimal performance requirements.

[0010] An intelligent coating optimization method for surface treatment of a quartz crucible, characterized in that it comprises the following steps:

[0011] Step 1: Collect coating material data, surface treatment environment data, coating process data and performance test data during the surface treatment of the quartz crucible, and use the performance test data as the true result label of the training set data;

[0012] Step 2: De-noise the collected data, identify and fill missing values, and finally perform data normalization;

[0013] Step 3: Based on the preprocessed data and the feature extracted data, the Pearson correlation coefficient and principal component analysis method are used to perform feature screening to obtain a set of key features;

[0014] Step 4: Train the deep neural network based on the obtained key features to obtain a trained deep neural network, and obtain coating performance based on the trained deep neural network;

[0015] Step 5: Based on the predicted coating performance, the particle swarm optimization algorithm is used to optimize the coating process parameters to obtain the optimal coating process parameter combination.

[0016] The intelligent coating optimization method for quartz crucible surface treatment proposed in the present invention aims at the problem that the optimization results caused by the traditional optimization of coating process based on personal experience are difficult to meet the optimal performance requirements. First, by collecting coating material data, surface treatment environment data, coating process data and performance test data, and taking the performance test data as the real result label of the training set, then, the collected data is denoised, missing value processed and normalized, so as to improve the data quality and lay the foundation for subsequent model training; then, a group of key features are screened out by Pearson correlation coefficient and principal component analysis, which reduces the data dimension and retains the main information affecting the coating performance; using these key features, the deep neural network model is trained, and the coating performance is predicted by the model, realizing the intelligent transformation from data to performance prediction; finally, based on the predicted coating performance, the particle swarm optimization algorithm is used to optimize the coating process parameters. This method can effectively search for the optimal parameter combination and avoid the blindness and high cost of the traditional optimization method; in summary, by integrating data collection, preprocessing, feature screening, deep learning model training and intelligent optimization algorithm, a closed-loop intelligent coating optimization process is formed, which significantly improves the optimization efficiency and performance of the coating process and solves the technical problems in the traditional method.

[0017] Preferably, for data with multiple measurement points for one feature data in one sample, the steps of performing feature screening by the Pearson correlation coefficient are as follows:

[0018] Calculate the Pearson correlation coefficient between the feature and the target variable based on the weight factor and the smoothing factor:

[0019]

[0020] Where: r xy (j) represents the smoothed Pearson correlation coefficient between the jth feature variable and the target variable; represents the average value of the jth characteristic variable of the i-th sample; represents the average value of the jth feature variable in all samples; y i Represents the value of the target variable of the i-th sample; represents the average value of the target variable of all samples; w ij represents the average weight factor of the jth feature variable of the i-th sample; α ij represents the average smoothing factor of the jth characteristic variable of the i-th sample;

[0021] The calculation formula of the smoothing factor is as follows:

[0022]

[0023] Where: α ik Represents the smoothing factor of the kth measurement value of the jth feature of the ith sample; residual ik Represents the residual of the kth measurement value of the jth feature of the i-th sample relative to the mean value of the jth feature variable of the sample; median(|residuals i |) represents the median of the absolute value of the residual error of the characteristic variable measurement value of the i-th sample; e represents the base of the natural logarithm; M represents the number of measurements of the j-th characteristic variable in the i-th sample;

[0024] The calculation formula of the weight factor is as follows:

[0025]

[0026] Where: w ik represents the weight factor of the kth measurement value of the jth feature of the i-th sample; x ik The feature data value representing the kth measurement value of the jth feature of the i-th sample; variance(x ik ) represents the feature data variance of the kth measurement value of the jth feature of the ith sample; mean variance(x j ) represents the average value of the data variance of the jth feature of all samples;

[0027] Based on the absolute values ​​of the calculated Pearson correlation coefficients, they are sorted from high to low, and the features corresponding to the first k Pearson correlation coefficients are selected as the input of the principal component analysis.

[0028] Preferably, when there is only one measurement value for a feature data in a sample, the Pearson correlation coefficient is calculated as follows:

[0029]

[0030] Where: r xy represents the Pearson correlation coefficient between feature x and target variable y; i represents the characteristic value of the i-th sample; y i Represents the target variable value of the i-th sample; Represents the mean of all samples of feature x; represents the mean of all samples of the target variable y; N represents the number of samples;

[0031] Based on the absolute values ​​of the calculated Pearson correlation coefficients, they are sorted from high to low, and the features corresponding to the first k Pearson correlation coefficients are selected as the input of the principal component analysis.

[0032] Preferably, the deep neural network includes an input layer, a feature cross layer, an intermediate layer and an output layer;

[0033] The input layer is used to receive the feature vector input into the deep neural network;

[0034] The feature cross layer generates new features by inner product cross and product cross, and then outputs them to the middle layer after nonlinear conversion through the ReLU activation function;

[0035] The intermediate layer receives the features output by the feature cross layer, and performs feature extraction based on multiple fully connected layers, and adds a residual connection in each fully connected layer, wherein the activation function adopts a ReLU activation function to introduce nonlinearity, and a regularization term is introduced in the intermediate layer to prevent overfitting;

[0036] Output layer: A fully connected layer containing one neuron is used to output the final predicted coating performance.

[0037] Preferably, the steps of performing data processing at the feature cross layer are as follows:

[0038] Assume that the input feature vector is X = [x1, x2, ..., x n ], where x i represents the i-th feature, and n represents the total number of features;

[0039] Inner product cross: for feature x i and x j Perform inner product crossover to generate new cross features:

[0040]

[0041] Where: z ij Represents feature x i and x j The features generated after inner product crossover are crossovered. Each pair of features is crossovered with inner product;

[0042] Element-wise product cross: Perform element-wise product operations on all features to obtain cross features:

[0043] z=X⊙X=[x1·x2,x1·x3,…,x n-1 ·x n ];

[0044] Where: ⊙ represents the element-wise product operation;

[0045] Generate a cross feature matrix: Concatenate the results of the inner product cross and the element-wise product cross to generate a new feature matrix Z:

[0046] Z=[z 12 ,z 13 ,…,z ij ,z];

[0047] Where: Z represents the generated cross feature vector, which contains the inner product cross and element-by-element product cross of all feature pairs;

[0048] The ReLU activation function is used to perform nonlinear transformation on the generated feature matrix Z and then output it to the middle layer.

[0049] Preferably, the loss function of the deep neural network is as follows:

[0050]

[0051] Where: L represents the value of the loss function; N represents the number of samples; y i Represents the true label value of the i-th sample; represents the predicted value of the i-th sample; λ represents the regularization coefficient; M represents the number of all weight parameters in the model; W j represents the jth weight parameter.

[0052] Preferably, step 5 comprises the following steps:

[0053] Initialize the particle swarm: Assume that the coating process parameters have d dimensions, define each particle in the particle swarm as a d-dimensional vector, representing a set of process parameter groups; initialize the position and velocity of each particle in the particle swarm:

[0054] x i =(x i1 ,x i2 ,…,x id ),v i =(v i1 ,v i2 ,…,v id );

[0055] Where: i represents the particle index; d represents the dimension of the process parameter; xi represents the position of the i-th particle; v i represents the velocity of the ith particle; x id represents the position of the dth dimension of the ith particle; v id represents the velocity of the dth dimension of the ith particle;

[0056] Calculate the fitness of particles: Predict the coating performance of each particle based on the trained deep neural network, and calculate the fitness of each particle based on the predicted coating performance:

[0057] f(x i )=-Predicted Performance(x i );

[0058] Where: Predicted Performance(x i ) represents the coating performance of each particle predicted by the deep neural network;

[0059] Update the individual best position and the global best position:

[0060] For each particle, record the optimal position p with the lowest historical fitness i , if the current fitness value is less than the historical optimal position p i The fitness value of the individual is updated:

[0061] p i =x i ,if f(x i ) <f(p i );

[0062] Record the global best position g among all particles, that is, the position of the particle with the smallest fitness:

[0063]

[0064] Update the particle's velocity and position:

[0065]

[0066] Where: represents the updated velocity of particle i in the dth dimension; represents the velocity of particle i in the dth dimension before the update; represents the position of particle i in d dimension before updating; represents the updated position of particle i in d dimension; ω represents the inertia weight; c1 and c2 represent acceleration constants, which control the degree to which the particle moves to the individual optimal position and the global optimal position respectively; r1 and r2 represent random numbers uniformly distributed between [0,1]; p idrepresents the individual optimal position of particle i in dimension d; g d The d-th dimension value representing the global best position;

[0067] Convergence determination: Set a maximum number of iterations. When the particle swarm has reached the maximum number of iterations or the improvement of the global optimal solution is less than the set threshold, stop the iteration; and output the optimal coating process parameter combination.

[0068] The beneficial effects of the present invention include:

[0069] The intelligent coating optimization method for quartz crucible surface treatment proposed in the present invention aims at the problem that the optimization results caused by the traditional optimization of coating process based on personal experience are difficult to meet the optimal performance requirements. First, by collecting coating material data, surface treatment environment data, coating process data and performance test data, and taking the performance test data as the real result label of the training set, then, the collected data is denoised, missing value processed and normalized, so as to improve the data quality and lay the foundation for subsequent model training; then, a group of key features are screened out by Pearson correlation coefficient and principal component analysis, which reduces the data dimension and retains the main information affecting the coating performance; using these key features, the deep neural network model is trained, and the coating performance is predicted by the model, realizing the intelligent transformation from data to performance prediction; finally, based on the predicted coating performance, the particle swarm optimization algorithm is used to optimize the coating process parameters. This method can effectively search for the optimal parameter combination and avoid the blindness and high cost of the traditional optimization method; in summary, by integrating data collection, preprocessing, feature screening, deep learning model training and intelligent optimization algorithm, a closed-loop intelligent coating optimization process is formed, which significantly improves the optimization efficiency and performance of the coating process and solves the technical problems in the traditional method. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0071] Figure 1 An overall step block diagram provided for an embodiment of the present invention.

[0072] Figure 2 A schematic diagram of the execution flow of a particle swarm optimization algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0074] See also Figure 1 As shown, an intelligent coating optimization method for surface treatment of a quartz crucible is characterized by comprising the following steps:

[0075] Step 1: Collect coating material data, surface treatment environment data, coating process data and performance test data during the surface treatment of the quartz crucible, and use the performance test data as the true result label of the training set data;

[0076] The coating material data includes the type, composition, physical and chemical properties of the coating material; wherein the physical and chemical properties include density, hardness, thermal conductivity, surface roughness and chemical stability of the coating material, wherein the chemical stability coating includes corrosion resistance and oxidation resistance under various environmental conditions;

[0077] The surface treatment environment data includes temperature, humidity and air pressure during the coating application process;

[0078] The coating process data includes spraying rate, spraying angle, spraying time, coating thickness, coating roughness, spraying rate and spraying angle;

[0079] Therefore, we test the temperature, humidity, spray thickness and coating roughness by setting multiple measuring points;

[0080] The performance test data is antioxidant test data;

[0081] Step 2: De-noise the collected data, identify and fill missing values, and finally perform data normalization;

[0082] Denoising:

[0083] For each data point, calculate its Z-score, which represents the multiple of the standard deviation of the data point from the mean; if the Z-score of a data point is greater than 3 or less than -3, it is considered an outlier.

[0084] Specific implementation:

[0085] Assume that the mean of the spraying rate is 100 and the standard deviation is 10. If a spraying rate data point is 500, its Z-score is: Z = (500–100) / 10 = 40; since the absolute value of the Z-score is greater than 3, it means that the data point is an outlier and can be deleted or replaced with other reasonable values.

[0086] For other numerical data such as coating roughness and coating thickness, the same Z-score method is used to detect and remove outliers.

[0087] Filling missing values ​​of numerical data:

[0088] For numerical data such as spraying rate, spraying angle, coating thickness, etc., if there are missing values, we will fill them with the mean of the feature.

[0089] Example: Suppose there is a missing value for spray rate at a certain location and the mean of the column is 150. If the value for that data point is missing, it is filled with 150.

[0090] Similar mean filling methods are used for other numerical data such as coating thickness and coating roughness.

[0091] Filling missing values ​​for categorical data:

[0092] For categorical data, such as coating material type, spraying environment, etc., we will use the mode of the column to fill in. The mode is the category with the highest frequency in the column.

[0093] Example: Suppose there are multiple categories (such as "A", "B", "C") in the coating material type column, and the missing value is in a row. If the mode of the column is "A", then fill the missing value with "A".

[0094] Similarly, other categorical data (such as surface treatment environment) also use the majority filling method.

[0095] Z-score standardization: For each numerical feature, calculate its mean and standard deviation, then subtract the mean from each data point and divide it by the standard deviation to obtain the standardized data.

[0096] Categorical data encoding: For categorical data, the unique hot encoding method is used to convert all categorical features into binary vector form:

[0097] The Coating Material Type column contains multiple categories (such as "A", "B", "C"). We will create a binary variable for each category.

[0098] For example, for the "Coating Material Type" column, its categories "Material A", "Material B" and "Material C" will be converted into three binary features: "A" → [1,0,0]; "B" → [0,1,0]; "C" → [0,0,1]; other categorical data such as spraying environment also use the same one-hot encoding method. In this way, the original category information is converted into a numerical vector, which is convenient for subsequent model training.

[0099] It should be noted that the spraying program is currently controlled based on pre-set fixed parameters. The parameters in the spraying process will not change frequently and continuously and will not rely on time continuity. Therefore, time series data is not considered in this embodiment.

[0100] Step 3: Based on the preprocessed data and the feature extracted data, the Pearson correlation coefficient and principal component analysis method are used to perform feature screening to obtain a set of key features;

[0101] For data with multiple measurement points for one feature data in a sample, the steps of feature screening by the Pearson correlation coefficient are as follows:

[0102] Calculate the Pearson correlation coefficient between the feature and the target variable based on the weight factor and the smoothing factor:

[0103]

[0104] Where: r xy (j) represents the smoothed Pearson correlation coefficient between the jth feature variable and the target variable; represents the average value of the jth characteristic variable of the i-th sample; represents the average value of the jth feature variable in all samples; y i Represents the value of the target variable of the i-th sample; represents the average value of the target variable of all samples; w ij represents the average weight factor of the jth feature variable of the i-th sample; α ij represents the average smoothing factor of the jth characteristic variable of the i-th sample;

[0105] The calculation formula of the smoothing factor is as follows:

[0106]

[0107] Where: α ik Represents the smoothing factor of the kth measurement value of the jth feature of the ith sample; residual ik Represents the residual of the kth measurement value of the jth feature of the i-th sample relative to the mean value of the jth feature variable of the sample; median(|residuals i |) represents the median of the absolute value of the residual error of the characteristic variable measurement value of the i-th sample; e represents the base of the natural logarithm; M represents the number of measurements of the j-th characteristic variable in the i-th sample;

[0108] The reason for introducing the smoothing factor in this embodiment is that in the analysis of coating uniformity (i.e., roughness), there may be some fluctuations due to accidental factors, which are manifested as residuals in the data; through the above formula, we assign a lower smoothing factor to the data points with larger residuals, thereby reducing their influence in the analysis; for example, if the coating uniformity of a certain area is abnormal due to external interference (such as temperature change), this abnormality will be reflected in the residual, and the smoothing factor will be reduced accordingly, reducing the interference of the data point on the final analysis result.

[0109] The calculation formula of the weight factor is as follows:

[0110]

[0111] Where: w ik represents the weight factor of the kth measurement value of the jth feature of the i-th sample; x ik The feature data value representing the kth measurement value of the jth feature of the i-th sample; variance(x ik ) represents the feature data variance of the kth measurement value of the jth feature of the ith sample; mean variance(x j ) represents the average value of the data variance of the jth feature of all samples;

[0112] The weight factor is introduced in this embodiment for the following reasons: for example, in coating thickness data, some measurement points may have greater fluctuations due to the particularity of their locations (such as near edges or seams). These measurement points have larger variances, and therefore should be given lower weights in the correlation analysis to avoid their excessive influence on the overall trend. The above method ensures that data points with larger variances have lower weights, while data points with smaller variances have higher weights.

[0113] In this embodiment, the inherent characteristics of the data are taken into account. The weight factor focuses on the relative size of the fluctuation of the data point itself, while the smoothing factor focuses on the consistency of the data point with the overall trend. By combining these two factors, the key factors in the surface treatment process of the quartz crucible can be analyzed more accurately and the coating process parameters can be optimized. This method helps to improve the explanatory power of the data, thereby achieving better results in coating performance prediction and process optimization.

[0114] Based on the absolute values ​​of the calculated Pearson correlation coefficients, they are sorted from high to low, and the features corresponding to the first k Pearson correlation coefficients are selected as the input of the principal component analysis.

[0115] When there is only one measurement value for a feature data in a sample, the Pearson correlation coefficient is calculated as follows:

[0116]

[0117] Where: r xy represents the Pearson correlation coefficient between feature x and target variable y; i represents the characteristic value of the i-th sample; y i Represents the target variable value of the i-th sample; Represents the mean of all samples of feature x; represents the mean of all samples of the target variable y; N represents the number of samples;

[0118] Based on the calculated absolute values ​​of the Pearson correlation coefficients, they are sorted from high to low, and the features corresponding to the first k Pearson correlation coefficients are selected as the input of the principal component analysis; the principal component analysis will not be described in detail in this embodiment, where principal component analysis is a currently known method for data dimensionality reduction and feature extraction; PCA (principal component analysis) was first proposed by Karl Pearson in 1901, and was further developed and promoted by other scholars; therefore, on the basis of giving a specific implementation method, how to perform principal component analysis based on the feature data after the output of the Pearson correlation coefficient is a conventional technical means in this field, so it will not be described in detail in this embodiment.

[0119] Step 4: Train the deep neural network based on the obtained key features to obtain a trained deep neural network, and obtain coating performance based on the trained deep neural network;

[0120] The deep neural network includes an input layer, a feature cross layer, an intermediate layer and an output layer;

[0121] The input layer is used to receive the feature vector input into the deep neural network;

[0122] The feature cross layer generates new features by inner product cross and product cross, and then outputs them to the middle layer after nonlinear conversion through the ReLU activation function;

[0123] The steps of data processing in the feature cross layer are as follows:

[0124] Assume that the input feature vector is X = [x1, x2, ..., x n ], where x i represents the i-th feature, and n represents the total number of features;

[0125] Inner product cross: for feature x i and x j Perform inner product crossover to generate new cross features:

[0126]

[0127] Where: zij Represents feature x i and x j The features generated after inner product crossover are crossovered. Each pair of features is crossovered with inner product;

[0128] Element-wise product cross: Perform element-wise product operations on all features to obtain cross features:

[0129] z=X⊙X=[x1·x2,x1·x3,…,x n-1 ·x n ];

[0130] Where: ⊙ represents the element-wise product operation;

[0131] Generate a cross feature matrix: Concatenate the results of the inner product cross and the element-wise product cross to generate a new feature matrix Z:

[0132] Z=[z 12 ,z 13 ,…,z ij ,z];

[0133] Where: Z represents the generated cross feature vector, which contains the inner product cross and element-by-element product cross of all feature pairs;

[0134] The ReLU activation function is used to perform nonlinear transformation on the generated feature matrix Z and then output it to the middle layer.

[0135] The intermediate layer receives the features output by the feature cross layer, and performs feature extraction based on multiple fully connected layers, and adds a residual connection in each fully connected layer, wherein the activation function adopts a ReLU activation function to introduce nonlinearity, and a regularization term is introduced in the intermediate layer to prevent overfitting;

[0136] Multiple fully connected layers: The middle layer is composed of multiple fully connected layers (MLP) to learn deeper feature relationships;

[0137] Residual connections: To avoid the gradient vanishing or gradient exploding problem in deep networks, residual connections (skip connections) are used; in each layer, the original input is directly added to the output of the layer, thereby achieving effective information transmission;

[0138] ReLU activation function: The output of each layer will pass through the ReLU activation function to increase the nonlinear mapping ability of the network;

[0139] The application of regularization is as follows:

[0140] Suppose we use two fully connected layers, the first and second layers, with weight matrices W1 and W2, and biases b1 and b2;

[0141] The weight matrix for the first layer is: Where m1 represents the number of parameters in the first layer weight matrix W1; λ1 represents the regularization coefficient;

[0142] The weight matrix for the second layer is: Where m2 represents the number of parameters in the second layer weight matrix W2; λ2 represents the regularization coefficient;

[0143] Output layer: A fully connected layer containing one neuron is used to output the final predicted coating performance;

[0144] For the weight matrix of the output layer: In the output layer W out L2 regularization is also performed:

[0145]

[0146] Where: m out Represents the output layer weight matrix W out The number of parameters in ; out represents the regularization coefficient of the output layer;

[0147] The loss function of the deep neural network is as follows:

[0148]

[0149] Where: L represents the value of the loss function; N represents the number of samples; y i Represents the true label value of the i-th sample; represents the predicted value of the i-th sample; λ represents the regularization coefficient; M represents the number of all weight parameters in the model; W j represents the jth weight parameter; represents the L2 regularization term.

[0150] In summary, the complete form of the loss function is as follows:

[0151]

[0152] therefore:

[0153]

[0154] Based on the above, it can be known that the final expression of the loss function L is related to the structure of the intermediate layer, and based on the above exemplary technical solution, the change law of the loss function L has been clearly stated, so it will not be repeated in this embodiment.

[0155] In this embodiment, the coating performance is antioxidant; of course, it is not a limitation of the present application, and the coating hardness can also be predicted. The difference is that the real result labels used are different, that is, our target variables are different; and other schemes are the same. Therefore, based on the specific technical scheme for predicting coating performance given in the above technical scheme, replacing it with the prediction of coating hardness is a conventional alternative in this field, so it will not be repeated in this embodiment.

[0156] Step 5: Based on the predicted coating performance, the particle swarm optimization algorithm is used to optimize the coating process parameters to obtain the optimal coating process parameter combination;

[0157] See also Figure 2 As shown, step 5 includes the following steps:

[0158] Initialize the particle swarm: Assume that the coating process parameters have d dimensions, define each particle in the particle swarm as a d-dimensional vector, representing a set of process parameter groups; initialize the position and velocity of each particle in the particle swarm:

[0159] x i =(x i1 ,x i2 ,…,x id ),v i =(v i1 ,v i2 ,…,v id );

[0160] Where: i represents the particle index; d represents the dimension of the process parameter; x i represents the position of the i-th particle; v i represents the velocity of the ith particle; x id represents the position of the dth dimension of the ith particle; v id represents the velocity of the dth dimension of the ith particle;

[0161] Calculate the fitness of particles: Predict the coating performance of each particle based on the trained deep neural network, and calculate the fitness of each particle based on the predicted coating performance:

[0162] f(x i )=-Predicted Performance(x i );

[0163] Where: Predicted Performance(x i ) represents the coating performance of each particle predicted by the deep neural network;

[0164] Update the individual best position and the global best position:

[0165] For each particle, record the optimal position p with the lowest historical fitness i , if the current fitness value is less than the historical optimal position p i The fitness value of the individual is updated:

[0166] p i =x i ,if f(x i ) <f(p i );

[0167] Record the global best position g among all particles, that is, the position of the particle with the smallest fitness:

[0168]

[0169] Update the particle's velocity and position:

[0170]

[0171] Where: represents the updated velocity of particle i in the dth dimension; represents the velocity of particle i in the dth dimension before the update; represents the position of particle i in d dimension before updating; represents the updated position of particle i in d dimension; ω represents the inertia weight; c1 and c2 represent acceleration constants, which control the degree to which the particle moves to the individual optimal position and the global optimal position respectively; r1 and r2 represent random numbers uniformly distributed between [0,1]; p id represents the individual optimal position of particle i in dimension d; g d The d-th dimension value representing the global best position;

[0172] Convergence determination: Set a maximum number of iterations. When the particle swarm has reached the maximum number of iterations or the improvement of the global optimal solution is less than the set threshold, stop the iteration; and output the optimal coating process parameter combination.

[0173] Exemplarily, the coating process parameters include 4 dimensions, and the data of other dimensions remain unchanged, namely: x1: spraying temperature (unit: °C), x2: spraying humidity (unit: %), x3: coating thickness (unit: μm) and x4: spraying rate (unit: mm / s); therefore, the dimension d of the process parameter space is 4;

[0174] Initialize particle position and velocity: Each particle in the particle swarm is a d-dimensional vector, representing a set of process parameter groups; initialize the position x of each particle in the particle swarm i =(x i1 ,x i2 ,x i3 ,xi4 ) and speed v i =(v i1 ,v i2 ,v i3 ,v i4 ), initialization strategy: initial position x i The parameters can be randomly set within the upper and lower limits of each process parameter. For example, assuming that the spray temperature ranges from 50°C to 150°C, the spray humidity ranges from 40% to 80%, the coating thickness ranges from 10μm to 100μm, and the spray rate ranges from 5mm / s to 20mm / s.

[0175] Initial velocity v i It can be randomly set within a reasonable range, usually within a small interval to avoid large jumps. For example, the speed can be initialized to ±10% of the process parameter range.

[0176] x i1 represents the value of particle i in the spray temperature dimension; for example, x i1 =120 means the current spraying temperature of the particle is 120℃;

[0177] x i2 Represents the value of particle i in the spraying humidity dimension; for example, x i2 =60 means that the current spraying humidity corresponding to the particle is 60%;

[0178] x i3 represents the value of particle i in the coating thickness dimension; for example, x i3 =50 means that the current coating thickness of the particle is 50 μm;

[0179] x i4 Represents the value of particle i in the spraying rate dimension; for example, x i4 =15 means that the current spraying rate of the particle is 15mm / s;

[0180] x i1 represents the velocity of particle i in the spray temperature dimension, for example, x i1 =2, which means that the position of particle i in the spray temperature dimension will increase by 2°C in each iteration cycle;

[0181] v i2 represents the velocity of particle i in the spraying humidity dimension, for example, v i2 = -3, which means that the position of particle i in the spraying wetness dimension will decrease by 3% in each iteration cycle;

[0182] v i3 represents the velocity of particle i in the coating thickness dimension, for example, v i3=5, which means that the position of particle i in the coating thickness dimension will increase by 5 μm in each iteration cycle;

[0183] v i4 Represents the velocity of particle i in the spraying rate dimension, for example, v i4 =-1, which means that the position of particle i in the spraying rate dimension will decrease by 1 mm / s in each iteration cycle.

[0184] Initialize the best position of the individual (p i ) and the global optimal position (g)

[0185] The individual optimal position p of each particle i It is the parameter position corresponding to the lowest fitness achieved by the particle in history, and can be set as the initial position of the particle at the beginning.

[0186] For example, p1=x1=(120,60,50,15);

[0187] The global best position g is the parameter position corresponding to the lowest fitness value historically achieved among all particles. Initially, the fitness of all particles is predicted by a deep neural network, and the global best position is determined by the parameter position of the best particle.

[0188] Calculate the fitness of particles: Calculate the fitness of each particle based on the fitness calculation formula;

[0189] Example: Suppose a deep neural network predicts:

[0190] The coating performance of particle 1 is 90% (i.e., the predicted performance value is 0.9), and the fitness f(x1) = -0.9;

[0191] The coating performance of particle 2 is 85% (i.e., the predicted performance value is 0.85), and the fitness f(x2) = -0.85;

[0192] For each particle i, we check whether the current fitness is lower than the historical fitness (i.e., whether the coating performance is better), and if so, update the individual best position p i ,For example:

[0193] If the fitness of particle 1 f(x1) = -0.9 is lower than the historical fitness f(x2) = -0.85; then update p1 = x1;

[0194] The global optimal position g records the position of the particle with the lowest fitness; for example, assuming that particle 1 has the best coating performance (the lowest fitness) among all particles, then g = p1

[0195] Update the position and velocity of particles: Update the position and velocity of particles based on the velocity and position update formula;

[0196] Convergence determination: set the maximum number of iterations (e.g., 1000), or stop iteration when the improvement of the global optimal solution is less than a set threshold (e.g., 0.01);

[0197] At the end of the iteration, the parameter position of the particle with the best fitness in the particle swarm is output, that is, the global optimal position g corresponds to the optimal coating process parameter combination, as shown below:

[0198] The optimal coating process parameter combination: g = (118, 63, 55, 16); spraying temperature: 118 ° C; spraying humidity: 63%; coating thickness: 55 μm; spraying rate: 16 mm / s.

[0199] The intelligent coating optimization method for quartz crucible surface treatment proposed in the present invention aims at the problem that the optimization results caused by the traditional optimization of coating process based on personal experience are difficult to meet the optimal performance requirements. First, by collecting coating material data, surface treatment environment data, coating process data and performance test data, and taking the performance test data as the real result label of the training set, then, the collected data is denoised, missing value processed and normalized, so as to improve the data quality and lay the foundation for subsequent model training; then, a group of key features are screened out by Pearson correlation coefficient and principal component analysis, which reduces the data dimension and retains the main information affecting the coating performance; using these key features, the deep neural network model is trained, and the coating performance is predicted by the model, realizing the intelligent transformation from data to performance prediction; finally, based on the predicted coating performance, the particle swarm optimization algorithm is used to optimize the coating process parameters. This method can effectively search for the optimal parameter combination and avoid the blindness and high cost of the traditional optimization method; in summary, by integrating data collection, preprocessing, feature screening, deep learning model training and intelligent optimization algorithm, a closed-loop intelligent coating optimization process is formed, which significantly improves the optimization efficiency and performance of the coating process and solves the technical problems in the traditional method.

[0200] The above are only preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An intelligent coating optimization method for surface treatment of quartz crucible, characterized in that: The following steps are involved: Step 1: Collect coating material data, surface treatment environment data, coating process data and performance test data during the surface treatment of the quartz crucible, and use the performance test data as the true result label of the training set data; Step 2: De-noise the collected data, identify and fill missing values, and finally perform data normalization; Step 3: Based on the preprocessed data, the Pearson correlation coefficient and principal component analysis method are used to screen the features and obtain a set of key features; Step 4: Train the deep neural network based on the obtained key features to obtain a trained deep neural network, and obtain coating performance based on the trained deep neural network; Step 5: Based on the predicted coating performance, the particle swarm optimization algorithm is used to optimize the coating process parameters to obtain the optimal coating process parameter combination.

2. The intelligent coating optimization method for surface treatment of a quartz crucible according to claim 1, characterized in that: For data with multiple measurement points for one feature data in a sample, the steps of feature screening by the Pearson correlation coefficient are as follows: Calculate the Pearson correlation coefficient between the feature and the target variable based on the weight factor and the smoothing factor: Where: r xy (j) represents the smoothed Pearson correlation coefficient between the jth feature variable and the target variable; represents the average value of the jth characteristic variable of the i-th sample; represents the average value of the jth feature variable in all samples; y i Represents the value of the target variable of the i-th sample; represents the average value of the target variable of all samples; w ij represents the average weight factor of the jth feature variable of the i-th sample; α ij represents the average smoothing factor of the jth characteristic variable of the i-th sample; The calculation formula of the smoothing factor is as follows: Where: α ik Represents the smoothing factor of the kth measurement value of the jth feature of the ith sample; residual ik Represents the residual of the kth measurement value of the jth feature of the i-th sample relative to the mean value of the jth feature variable of the sample; median(|residuals i |) represents the median of the absolute value of the residual error of the characteristic variable measurement value of the i-th sample; e represents the base of the natural logarithm; M represents the number of measurements of the j-th characteristic variable in the i-th sample; The calculation formula of the weight factor is as follows: Where: w ik represents the weight factor of the kth measurement value of the jth feature of the i-th sample; x ik The feature data value representing the kth measurement value of the jth feature of the i-th sample; variance(x ik ) represents the feature data variance of the kth measurement value of the jth feature of the ith sample; mean variance(x j ) represents the average value of the data variance of the jth feature of all samples; Based on the absolute values ​​of the calculated Pearson correlation coefficients, they are sorted from high to low, and the features corresponding to the first k Pearson correlation coefficients are selected as the input of the principal component analysis.

3. The intelligent coating optimization method for surface treatment of a quartz crucible according to claim 1, characterized in that: When there is only one measurement value for a feature data in a sample, the Pearson correlation coefficient is calculated as follows: Where: r xy represents the Pearson correlation coefficient between feature x and target variable y; i represents the characteristic value of the i-th sample; y i Represents the target variable value of the i-th sample; Represents the mean of all samples of feature x; represents the mean of all samples of the target variable y; N represents the number of samples; Based on the absolute values ​​of the calculated Pearson correlation coefficients, they are sorted from high to low, and the features corresponding to the first k Pearson correlation coefficients are selected as the input of the principal component analysis.

4. The intelligent coating optimization method for surface treatment of a quartz crucible according to claim 1, characterized in that: The deep neural network includes an input layer, a feature cross layer, an intermediate layer and an output layer; The input layer is used to receive the feature vector input into the deep neural network; The feature cross layer generates new features by inner product cross and product cross, and then outputs them to the middle layer after nonlinear conversion through the ReLU activation function; The intermediate layer receives the features output by the feature cross layer, and performs feature extraction based on multiple fully connected layers, and adds a residual connection in each fully connected layer, wherein the activation function adopts a ReLU activation function to introduce nonlinearity, and a regularization term is introduced in the intermediate layer to prevent overfitting; Output layer: A fully connected layer containing one neuron is used to output the final predicted coating performance.

5. The intelligent coating optimization method for surface treatment of a quartz crucible according to claim 4, characterized in that: The steps of data processing in the feature cross layer are as follows: Assume that the input feature vector is X = [x1, x2, ..., x n ], where x i represents the i-th feature, and n represents the total number of features; Inner product cross: for feature x i and x j Perform inner product crossover to generate new cross features: Where: z ij Represents feature x i and x j The features generated after inner product crossover are crossovered. Each pair of features is crossovered with inner product; Element-wise product cross: Perform element-wise product operations on all features to obtain cross features: z=X⊙X=[x1·x2,x1·x3,…,x n-1 ·x n ]: Where: ⊙ represents the element-wise product operation; Generate a cross feature matrix: Concatenate the results of the inner product cross and the element-wise product cross to generate a new feature matrix Z: From=[from 12 ,With 13 ,…,With ij ,With]; Where: Z represents the generated cross feature vector, which contains the inner product cross and element-by-element product cross of all feature pairs; The ReLU activation function is used to perform nonlinear transformation on the generated feature matrix Z and then output it to the middle layer.

6. The intelligent coating optimization method for surface treatment of a quartz crucible according to claim 4, characterized in that: The loss function of the deep neural network is as follows: Where: L represents the value of the loss function; N represents the number of samples; y i Represents the true label value of the i-th sample; represents the predicted value of the i-th sample; λ represents the regularization coefficient; M represents the number of all weight parameters in the model; W j represents the jth weight parameter.

7. The intelligent coating optimization method for surface treatment of a quartz crucible according to claim 1, characterized in that: The step 5 comprises the following steps: Initialize the particle swarm: Assume that the coating process parameters have d dimensions, define each particle in the particle swarm as a d-dimensional vector, representing a set of process parameter groups; initialize the position and velocity of each particle in the particle swarm: x i =(x i1 ,x i2 ,…,x id ),v i =(v i1 ,v i2 ,…,v id ); Where: i represents the particle index; d represents the dimension of the process parameter; x i represents the position of the i-th particle; v i represents the velocity of the ith particle; x id represents the position of the dth dimension of the ith particle; v id represents the velocity of the dth dimension of the ith particle; Calculate the fitness of particles: Predict the coating performance of each particle based on the trained deep neural network, and calculate the fitness of each particle based on the predicted coating performance: f(x i )=-Predicted Performance(x i ); Where: Predicted Performance(x i ) represents the coating performance of each particle predicted by the deep neural network; Update the individual best position and the global best position: For each particle, record the optimal position p with the lowest historical fitness i , if the current fitness value is less than the historical optimal position p i The fitness value of the individual is updated: p i =x i ,if f(x i )<f(p i ); Record the global best position g among all particles, that is, the position of the particle with the smallest fitness: Update the particle's velocity and position: Where: represents the updated velocity of particle i in the dth dimension; represents the velocity of particle i in the dth dimension before the update; represents the position of particle i in d dimension before updating; represents the updated position of particle i in d dimension; ω represents the inertia weight; c1 and c2 represent acceleration constants, which control the degree to which the particle moves to the individual optimal position and the global optimal position respectively; r1 and r2 represent random numbers uniformly distributed between [0,1]; p id represents the individual optimal position of particle i in dimension d; g d The d-th dimension value representing the global best position; Convergence determination: Set a maximum number of iterations. When the particle swarm has reached the maximum number of iterations or the improvement of the global optimal solution is less than the set threshold, stop the iteration; and output the optimal coating process parameter combination.

Images