A glass hardness prediction method based on hierarchical prediction neural network

Abstract

The present application discloses a glass hardness prediction method based on a hierarchical prediction neural network, which belongs to the field of data prediction systems and is used for glass research and development. It includes data acquisition, data preprocessing, data set construction, building a hierarchical prediction neural network composed of a Kalmogorov-Arnold network and a feedforward neural network, training a neural network model by gradient descent and back propagation, and evaluating the prediction error by the mean absolute error. After the prediction error of the trained neural network model meets the accuracy requirements, the training is stopped and the corresponding network structure and parameters are saved for predicting glass hardness and guiding glass research and development. In view of the above technical solution, the present application can predict glass hardness through existing data before preparing glass, thereby accelerating glass design and shortening the glass research and development cycle.

Description

Technical Field

[0001] The present application belongs to the field of data prediction systems, and more specifically, relates to a glass hardness prediction method based on a hierarchical prediction neural network. Background Art

[0002] In the traditional glass development process, trial and error methods are mostly used. Not only are the experimental steps cumbersome, but it also takes a long time from research and development to application, and often fails to achieve the expected results. At the same time, there are more and more material characterization techniques, and the corresponding data and dimensions are becoming more and more complex. Traditional experimental analysis methods that simply rely on traditional manpower sometimes cannot dig out the deep connection between material characteristics and performance, although material calculation simulation methods such as first principle methods (such as the method for predicting glass system performance disclosed in Chinese patent publication number CN110364231A), molecular dynamics, Monte Carlo technology, phase field theory, and finite element analysis can perform calculations and predictions on the structure and performance of materials at different scales.

[0003] However, it is only applicable to specific systems. When faced with complex systems, there is often an unbearable amount of calculation. The development of some theoretical methods cannot meet the requirements of quantitative description of material properties. These have restricted the pace of new material research and development.

[0004] In recent years, with the development of computer technology and network technology, neural networks have gradually been applied to the technical field of glass development. For example, the patent document with Chinese patent publication number CN105095273A discloses a glass tempering process parameter setting method based on a fuzzy BP neural network, and specifically discloses the following technical content: In the process parameter setting method, the BP neural network is trained according to the initial sample, and the existing tempered glass is divided into several categories. The trained BP neural network is used to obtain the optimal process parameters of each category, and a process database is constructed according to the process parameters of all categories. The category of the glass to be tempered is directly determined, and the process parameters corresponding to the category are directly selected from the process database. The selected process parameters are used to set the process parameters of the tempered glass to be tempered.

[0005] However, the above-mentioned patent document No. CN105095273A is only an optimization of existing process parameters, rather than a design of new glass materials. How to use neural networks and improvements to neural networks to predict the hardness performance of glass during glass research and development and shorten the process and cycle of glass research and development has become a technical problem that technicians in this field need to solve. Summary of the invention

[0006] The purpose of this application is to provide a glass hardness prediction method based on a hierarchical prediction neural network, which can predict the hardness of glass through existing data before manufacturing the glass, thereby accelerating glass design and shortening the glass research and development cycle.

[0007] To achieve the above objectives, this application is implemented through the following technical solutions:

[0008] A method for predicting glass hardness based on a hierarchical prediction neural network, the method comprising the following steps:

[0009] S1. Data acquisition, including chemical composition, melting temperature, clarification temperature, forming temperature, heat treatment process, cold processing process of glass, and hardness of glass obtained through measuring equipment;

[0010] S2, data preprocessing, including normalizing the chemical composition, melting temperature, clarification temperature, and molding temperature obtained in step S1, and one-hot encoding the heat treatment process and cold processing process data obtained in step S1 to construct data that meets the model input type of step S4;

[0011] S3, constructing a data set, including dividing the data formed in step S2 into a training set and a validation set, with the ratio of the data volume of the training set to that of the validation set being 4:1;

[0012] S4. Building a neural network model, including building a hierarchical prediction neural network, which includes two neural networks, namely, the Kalmogorov-Arnold network and the feedforward neural network;

[0013] S5, using the training set in step S3 to train the neural network model, using the mean absolute error to evaluate the prediction error, and training the neural network model by gradient descent and back propagation;

[0014] S6. Use the verification set in step S3 to perform performance testing and verification on the neural network model trained in step S5. When the prediction error of the neural network model meets the accuracy requirement, stop training the neural network model and save the corresponding network structure and parameters. Perform data preprocessing on the chemical composition, melting temperature, clarification temperature and forming temperature of the glass to be designed, heat treatment process and cold processing process, and input them into the trained neural network model. The output value obtained is the average hardness of the glass predicted by the model.

[0015] As one of the preferred technical solutions, in the present application, obtaining the hardness of the glass by measuring the device in step S1 includes taking 5 measuring points on average in each glass and measuring the hardness of the glass at the 5 measuring points respectively. , , The hardness of the glass at 5 points is an integer, and the average hardness of the glass is obtained. , and through formula 1:

[0016] (1), calculate the glass hardness at each point Average value of glass hardness The relative error , is an integer, and .

[0017] As one of the preferred technical solutions, in this application, in step S1, if any relative error calculated is If the relative error is greater than 15%, it can be considered that a defective point was selected when setting the sampling point. To ensure the training effect of the subsequent step S5, the sampling points of the glass should be redesigned until the relative error corresponding to all sampling points is Are less than or equal to 15%.

[0018] As one of the preferred technical solutions, in the present application, the step S2 includes the following steps: Step S21, normalizing the chemical composition of the glass, that is, scaling the data to [0, 1], and calculating using Formula 2: (2), where is the amount of substance in a chemical composition, is the maximum amount of the chemical component substance in the data set, It is the minimum value of the amount of the chemical component substance in the data set. The unit of the chemical component substance is mole.

[0019] Step S22: normalize the melting temperature, clarification temperature and forming temperature of the glass respectively, that is, scale the data to [0, 1], and calculate using Formula 3: (3), where It is the melting temperature, clarification temperature and forming temperature of glass. is the maximum value corresponding to the melting temperature, clarification temperature, and molding temperature in the data set. It is the minimum value of melting temperature, clarification temperature and molding temperature in the data set. The units of melting temperature, clarification temperature and molding temperature are all degrees Celsius.

[0020] Step S23, one-hot encoding is performed on the categorical data of the heat treatment process and the cold treatment process of the glass, wherein the one-hot encoding of the heat treatment process is performed in the following manner: Block glass sample creates a vector If the glass sample is annealed, And order is equal to the annealing temperature, otherwise , are all 0; if the sample adopts the quenching process, then And order is equal to the quenching temperature, otherwise , are all 0; if the sample adopts the tempering process, then And order Equal to the tempering temperature, otherwise , are all 0; then the temperature data corresponding to annealing, quenching and tempering in the heat treatment process are normalized by analogy with formula 3, and the normalized vector is recorded as , ,in, After normalization .

[0021] The unique hot encoding of the cold treatment process is as follows: Block glass sample creates a vector If the glass sample is ground, , otherwise let If the glass sample is polished, , otherwise let ; and so on, until all the cold treatment processes used are coded; that is, each additional cold treatment process not listed will be coded. Add a dimension.

[0022] As one of the preferred technical solutions, in the present application, the step S4 includes the following contents: Step S41: Building a hierarchical prediction neural network model, the hierarchical prediction neural network model is composed of two neural networks, the two neural networks are a Kalmogorov-Arnold network and a feedforward neural network, and the two neural networks respectively include 1 input layer, 1 hidden layer and 1 output layer, wherein the output layer of the Kalmogorov-Arnold network is also part of the input layer of the feedforward neural network.

[0023] The number of neurons in the input layer of the Kalmogorov-Arnold network Formula 4: (4), where: is the number of glass chemical components in the data set; It is the sum of the number of heat treatment processes and heat treatment temperatures in the data set; is the number of glass melting temperatures and clarification temperatures in the data set, which is always 1; therefore, Formula 4 can be simplified to Formula 5: (5).

[0024] Number of neurons in the hidden layer of the Kalmogorov-Arnold network , according to the Kalmogorov-Arnold representation theorem, should be set to Formula 6: (6).

[0025] Number of neurons in the output layer of the Kalmogorov-Arnold network As shown in formula 7: (7).

[0026] Functional Expression of Kalmogorov-Arnold Network See Formula 8 for the definition: (8), where: For input data into the Kalmogorov-Arnold network; In represents the mapping from layer 0 to layer 1, In , represents the mapping from layer 1 to layer 2, both of which are activation functions of the Kalmogorov-Arnold network , that is, SigmoidLinear Unit function and spline functions The linear combination of , see Formula 9 for the specific formula: (9), where: and is a trainable parameter.

[0027] Spline function It is a linear combination of B-spline functions, see formula 10: (10), where: is a trainable parameter, is the B-spline function.

[0028] Step 42: The network structure of the feedforward neural network includes: the number of neurons in the input layer of the feedforward neural network See formula 11: (11), where: is the number of cold treatment processes in the data set, is the number of neurons in the output layer of the Kalmogorov-Arnold network, which is always 1, so Formula 11 can be simplified to (12).

[0029] Number of neurons in the hidden layer of a feedforward neural network See formula 13: (13).

[0030] Number of neurons in the output layer of a feedforward neural network See formula 14: (14).

[0031] The activation function of the hidden layer of the feedforward neural network is the ReLU function with leakage, and the activation function of the output layer of the feedforward neural network is the Sigmoid function; the optimization algorithm is preferentially selected as the Adam algorithm, which can use momentum as the parameter update direction and can adaptively adjust the learning rate.

[0032] As one of the preferred technical solutions, in this application, the leaky ReLU function in step 42 is as shown in Formula 15: (15), where Take 0.01.

[0033] The Sigmoid function is shown in Formula 16: (16).

[0034] As one of the preferred technical solutions, in the present application, the step S5 includes: S51, the calculation formula of the mean absolute error is shown in Formula 17: (17), where m is the number of samples in the training set and the validation set, is the predicted glass hardness value, is the average hardness value measured; the neural network model is trained by gradient descent method and back propagation. In order to avoid the degradation of model performance caused by overfitting, the early stopping method is used to stop training in advance before overfitting.

[0035] As one of the preferred technical solutions, in the present application, the step S5 also includes step S52, constructing a prediction model for glass hardness according to a hierarchical neural network structure, and obtaining parameters of each layer of the neural network through the optimal structure of the trained neural network; the prediction model is constructed in the form of formula 18: (18), where: It is the chemical composition data of glass after normalization; These are the normalized melting temperature, clarification temperature and molding temperature data; is the heat treatment process data after normalization; is the normalized cold working process data; For the feedforward neural network The weight of the layer; The feedforward neural network is Bias of the layer; and are the activation functions of the input layer and hidden layer in the Kalmogorov-Arnold network, see Formula 9; The function is shown in Formula 15; The function is shown in Formula 16.

[0036] Compared with the prior art, the beneficial effects of this application are:

[0037] The present application can provide a method for quickly predicting glass performance during the glass development process, from data acquisition and data preprocessing to selecting a suitable hierarchical neural network, selecting the hierarchical neural network as Kolmogorov-Arnold Networks (KAN) and Feed-Forward Neural Network, and adding the mean absolute error to evaluate the prediction error during the training process, and training the neural network model through the gradient descent method mixed with back propagation to increase the prediction accuracy of the prediction model, thereby accelerating glass design and shortening the glass research and development cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flowchart of this application. DETAILED DESCRIPTION

[0039] The technical solution described in this application is further described below in conjunction with the accompanying drawings and embodiments.

[0040] Example 1: Figure 1 As shown, a glass hardness prediction method based on a hierarchical prediction neural network comprises the following steps: S1, acquisition of data, including the chemical composition of the glass, melting temperature, clarification temperature, molding temperature, heat treatment process, cold processing process, and acquisition of the hardness of the glass through measuring equipment.

[0041] S2. Data preprocessing, including normalizing the chemical composition, melting temperature, clarification temperature, and molding temperature obtained in step S1, and one-hot encoding the heat treatment process and cold processing process data obtained in step S1 to construct data that meets the model input type of step S4.

[0042] S3, constructing a data set, including dividing the data formed in step S2 into a training set and a validation set, with the ratio of the data volume of the training set to that of the validation set being 4:1.

[0043] S4. Building a neural network model, including building a hierarchical prediction neural network. The hierarchical prediction neural network includes two neural networks, namely the Kalmogorov-Arnold network and the feedforward neural network.

[0044] S5. Use the training set in step S3 to train the neural network model, use the mean absolute error to evaluate the prediction error, and train the neural network model through gradient descent and back propagation.

[0045] S6. Use the verification set in step S3 to perform performance testing and verification on the neural network model trained in step S5. When the prediction error of the neural network model meets the accuracy requirement, stop training the neural network model and save the corresponding network structure and parameters. Perform data preprocessing on the chemical composition, melting temperature, clarification temperature and forming temperature of the glass to be designed, heat treatment process and cold processing process, and input them into the trained neural network model. The output value obtained is the average hardness of the glass predicted by the model.

[0046] Embodiment 2: A method for predicting glass hardness based on a hierarchical prediction neural network, wherein the step S1 of obtaining the hardness of the glass by measuring the device includes taking 5 measuring points on average in each glass (the number of measuring points can also be appropriately increased within the glass range according to the actual area of ​​the glass), and measuring the glass hardness at the 5 measuring points respectively. , , The hardness of the glass at 5 points is an integer, and the average hardness of the glass is obtained. , and through formula 1:

[0047] (1), calculate the glass hardness at each point Average value of glass hardness The relative error , is an integer, and .

[0048] In step S1, if any relative error calculated If the relative error is greater than 15%, it can be considered that a defective point was selected when setting the sampling point. To ensure the training effect of the subsequent step S5, the sampling points of the glass should be redesigned until the relative error corresponding to all sampling points is Are less than or equal to 15%.

[0049] The step S2 comprises the following steps: Step S21, normalizing the chemical composition of the glass, that is, scaling the data to [0, 1], and calculating using Formula 2: (2), where is the amount of substance in a chemical composition, is the maximum amount of the chemical component substance in the data set, It is the minimum value of the amount of the chemical component substance in the data set. The unit of the chemical component substance is mole.

[0050] Step S22: normalize the melting temperature, clarification temperature and forming temperature of the glass respectively, that is, scale the data to [0, 1], and calculate using Formula 3: (3), where It is the melting temperature, clarification temperature and forming temperature of glass. is the maximum value corresponding to the melting temperature, clarification temperature, and molding temperature in the data set. It is the minimum value of melting temperature, clarification temperature and molding temperature in the data set. The units of melting temperature, clarification temperature and molding temperature are all degrees Celsius.

[0051] Step S23, one-hot encoding is performed on the categorical data of the heat treatment process and the cold treatment process of the glass, wherein the one-hot encoding of the heat treatment process is performed in the following manner: Block glass sample creates a vector If the glass sample is annealed, And order is equal to the annealing temperature, otherwise , are all 0; if the sample adopts the quenching process, then And order is equal to the quenching temperature, otherwise , are all 0; if the sample adopts the tempering process, then And order Equal to the tempering temperature, otherwise , are all 0; then the temperature data corresponding to annealing, quenching and tempering in the heat treatment process are normalized by analogy with formula 3, and the normalized vector is recorded as , ,in, After normalization .

[0052] The heat treatment process of glass mainly includes annealing, quenching and tempering. Annealing refers to heating the glass to a certain temperature and then slowly cooling it to eliminate the application and defects inside the glass; quenching refers to rapidly cooling the glass to change the surface state and internal structure of the glass; tempering refers to heating the glass to a certain temperature again and then slowly cooling it to eliminate the stress and defects generated during the quenching process. The cold processing technology of glass mainly includes mechanical methods such as grinding, polishing, edging, cutting, drilling, frosting, sandblasting, and engraving, which can change the appearance and surface morphology of glass and glass products.

[0053] The unique hot encoding of the cold treatment process is as follows: Block glass sample creates a vector If the glass sample is ground, , otherwise let If the glass sample is polished, , otherwise let ; and so on, until all the cold treatment processes used are coded; that is, each additional cold treatment process not listed will be coded. Add a dimension.

[0054] The technical solutions for the remaining parts are the same as those described in Example 1, and will not be described again here to avoid cumbersome writing.

[0055] Embodiment 3: A glass hardness prediction method based on a hierarchical prediction neural network, wherein step S4 comprises the following contents: step S41: constructing a hierarchical prediction neural network model, wherein the hierarchical prediction neural network model is composed of two neural networks, namely a Kalmogorov-Arnold network and a feedforward neural network, and the two neural networks respectively include 1 input layer, 1 hidden layer and 1 output layer, wherein the output layer of the Kalmogorov-Arnold network is also part of the input layer of the feedforward neural network.

[0056] The number of neurons in the input layer of the Kalmogorov-Arnold network Formula 4: (4), where: is the number of glass chemical components in the data set; It is the sum of the number of heat treatment processes and heat treatment temperatures in the data set; is the number of glass melting temperatures and clarification temperatures in the data set, which is always 1; therefore, Formula 4 can be simplified to Formula 5: (5).

[0057] Number of neurons in the hidden layer of the Kalmogorov-Arnold network , according to the Kalmogorov-Arnold representation theorem, should be set to Formula 6: (6).

[0058] Number of neurons in the output layer of the Kalmogorov-Arnold network As shown in formula 7: (7).

[0059] Functional Expression of Kalmogorov-Arnold Network See Formula 8 for the definition: (8), where: For input data into the Kalmogorov-Arnold network; In represents the mapping from layer 0 to layer 1, In , represents the mapping from layer 1 to layer 2, both of which are activation functions of the Kalmogorov-Arnold network , that is, SigmoidLinear Unit function and spline functions The linear combination of , see Formula 9 for the specific formula: (9), where: and is a trainable parameter.

[0060] Spline function It is a linear combination of B-spline functions, see formula 10: (10), where: is a trainable parameter, is the B-spline function.

[0061] Step 42: The network structure of the feedforward neural network includes: the number of neurons in the input layer of the feedforward neural network See formula 11: (11), where: is the number of cold treatment processes in the data set, is the number of neurons in the output layer of the Kalmogorov-Arnold network, which is always 1, so Formula 11 can be simplified to (12).

[0062] Number of neurons in the hidden layer of a feedforward neural network See formula 13: (13).

[0063] Number of neurons in the output layer of a feedforward neural network See formula 14: (14).

[0064] The activation function of the hidden layer of the feedforward neural network is the ReLU function with leakage, and the activation function of the output layer of the feedforward neural network is the Sigmoid function; the optimization algorithm is preferentially selected as the Adam algorithm, which can use momentum as the parameter update direction and can adaptively adjust the learning rate.

[0065] The leaky ReLU function in step 42 is shown in Formula 15: (15), where Take 0.01.

[0066] The Sigmoid function is shown in Formula 16: (16).

[0067] The technical solutions for the remaining parts are the same as those described in the aforementioned embodiments, and will not be described in detail here to avoid cumbersome writing.

[0068] Embodiment 4: A method for predicting glass hardness based on a hierarchical prediction neural network, wherein the step S5 comprises: S51, the calculation formula of the mean absolute error is shown in Formula 17: (17), where m is the number of samples in the training set and the validation set, is the predicted glass hardness value, is the average hardness value measured.

[0069] The neural network model is trained by gradient descent and back propagation. In order to avoid the degradation of model performance caused by overfitting, the early stopping method is used to stop the training in advance before overfitting.

[0070] The step S5 also includes a step S52, constructing a prediction model for glass hardness according to the hierarchical neural network structure, obtaining parameters of each layer of the neural network through the optimal structure of the trained neural network, and terminating the training of the neural network.

[0071] The form of constructing the prediction model is formula 18: (18), where: It is the chemical composition data of glass after normalization; These are the normalized melting temperature, clarification temperature and molding temperature data; is the heat treatment process data after normalization; is the normalized cold working process data; For the feedforward neural network The weight of the layer; The feedforward neural network is Bias of the layer; and are the activation functions of the input layer and hidden layer in the Kalmogorov-Arnold network, see Formula 9; The function is shown in Formula 15; The function is shown in Formula 16.

[0072] S6. Save the optimal structure of the trained neural network and the parameters of each layer of the neural network to predict the hardness of glass and guide glass research and development.

[0073] The technical solutions for the remaining parts are the same as those for the aforementioned parts and will not be elaborated here to avoid cumbersome writing.

[0074] Finally, although this specification is described according to implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A glass hardness prediction method based on a hierarchical prediction neural network, characterized in that: The method comprises the following steps: S1. Data acquisition, including chemical composition, melting temperature, clarification temperature, forming temperature, heat treatment process, cold processing process of glass, and hardness of glass obtained through measuring equipment; S2, data preprocessing, including normalizing the chemical composition, melting temperature, clarification temperature, and molding temperature obtained in step S1, and one-hot encoding the heat treatment process and cold processing process data obtained in step S1 to construct data that meets the model input type of step S4; S3, constructing a data set, including dividing the data formed in step S2 into a training set and a validation set, with the ratio of the data volume of the training set to that of the validation set being 4:1; S4. Building a neural network model, including building a hierarchical prediction neural network, which includes two neural networks, namely, the Kalmogorov-Arnold network and the feedforward neural network; S5, using the training set in step S3 to train the neural network model, using the mean absolute error to evaluate the prediction error, and training the neural network model by gradient descent and back propagation; S6, using the validation set in step S3 to perform performance testing and validation on the neural network model trained in step S5, after the prediction error of the neural network model meets the accuracy requirement, stop training the neural network model and save the corresponding network structure and parameters, perform data preprocessing on the chemical composition, melting temperature, clarification temperature and forming temperature of the glass to be designed, heat treatment process and cold processing process, and input the data into the trained neural network model, and the output value obtained is the average hardness of the glass predicted by the model; Wherein, the step S4 includes the following contents: Step S41: constructing a hierarchical prediction neural network model, wherein the hierarchical prediction neural network model is composed of two neural networks, the two neural networks are a Kalmogorov-Arnold network and a feedforward neural network, and the two neural networks respectively include an input layer, a hidden layer and an output layer, wherein the output layer of the Kalmogorov-Arnold network is also a part of the input layer of the feedforward neural network; The number of neurons in the input layer of the Kalmogorov-Arnold network Formula (4): （4） in: is the number of glass chemical components in the data set; It is the sum of the number of heat treatment processes and heat treatment temperatures in the data set; is the number of glass melting temperatures and clarification temperatures in the data set, always 2; Therefore, formula (4) can be simplified to formula (5): I1=N x +N hot +2(5) Number of neurons in the hidden layer of the Kalmogorov-Arnold network , according to the Kalmogorov-Arnold representation theorem, should be set to formula (6): （6） Number of neurons in the output layer of the Kalmogorov-Arnold network As shown in formula (7): （7） Functional Expression of Kalmogorov-Arnold Network See formula (8) for the definition: （8） in: For input data into the Kalmogorov-Arnold network; In represents the mapping from layer 0 to layer 1, In , represents the mapping from layer 1 to layer 2, both of which are activation functions of the Kalmogorov-Arnold network , that is, the Sigmoid Linear Unit function and spline functions The linear combination of , see formula (9) for the specific formula: （9） in: and is a trainable parameter; Spline function is a linear combination of B-spline functions, see formula (10): （10） in: is a trainable parameter, is the B-spline function; Step 42: The network structure of the feedforward neural network includes: the number of neurons in the input layer of the feedforward neural network See formula (11): （11） in: is the number of cold treatment processes in the data set, is the number of neurons in the output layer of the Kalmogorov-Arnold network, which is always 1. Therefore, formula (11) can be simplified to (12) Number of neurons in the hidden layer of a feedforward neural network See formula (13): （13） Number of neurons in the output layer of a feedforward neural network See formula (14): （14） The activation function of the hidden layer of the feedforward neural network is the ReLU function with leakage, and the activation function of the output layer of the feedforward neural network is the Sigmoid function; the optimization algorithm is preferably the Adam algorithm, which can use momentum as the parameter update direction and can adaptively adjust the learning rate; The leaky ReLU function in step 42 is shown in formula (15): （15） in Take 0.01; The Sigmoid function is shown in formula (16): （16）； The step S5 comprises: S51. The calculation formula of mean absolute error is shown in formula (17): （17） Where: m is the number of samples in the training set and the validation set, is the predicted glass hardness value, is the average hardness value measured; The neural network model is trained by gradient descent and back propagation. To avoid the degradation of model performance caused by overfitting, the early stopping method is used to stop the training in advance before overfitting. The step S5 further includes a step S52, constructing a prediction model for glass hardness according to the hierarchical neural network structure, and obtaining parameters of each layer of the neural network through the optimal structure of the trained neural network; The form of constructing the prediction model is formula (18): （18） in: It is the chemical composition data of glass after normalization; These are the normalized melting temperature, clarification temperature and molding temperature data; is the heat treatment process data after normalization; is the normalized cold working process data; For the feedforward neural network The weight of the layer; The feedforward neural network is Bias of the layer; and are the activation functions of the input layer and hidden layer in the Kalmogorov-Arnold network, see formula (9); The function is shown in formula (15); The function is shown in formula (16).

2. The glass hardness prediction method based on a hierarchical prediction neural network according to claim 1, characterized in that: The step S1 of obtaining the hardness of the glass by measuring the device includes taking 5 measuring points on average in each piece of glass and measuring the hardness of the glass at the 5 measuring points respectively. , 1≤i≤5, i is an integer, and the average hardness of the glass is obtained by measuring the hardness of the glass at 5 points , and through formula (1): （1） Calculate the glass hardness at each point Average value of glass hardness The relative error , i is an integer, and 1≤i≤5.

3. The glass hardness prediction method based on hierarchical prediction neural network according to claim 2, characterized in that: In step S1, if any relative error calculated If the relative error is greater than 15%, it can be considered that a defective point was selected when setting the sampling point. To ensure the training effect of the subsequent step S5, the sampling points of the glass should be redesigned until the relative error corresponding to all sampling points is Are less than or equal to 15%.

4. The glass hardness prediction method based on hierarchical prediction neural network according to claim 3, characterized in that: The step S2 comprises the following steps: Step S21, normalize the chemical composition of the glass, that is, scale the data to [0, 1], and use formula (2) to calculate: （2） in, is the amount of substance in a chemical composition, is the maximum amount of the chemical component substance in the data set, is the minimum value of the amount of the chemical component substance in the data set, and the unit of the chemical component substance is mole; Step S22: normalize the melting temperature, clarification temperature and forming temperature of the glass respectively, that is, scale the data to [0, 1], and calculate using formula (3): （3） in It is the melting temperature, clarification temperature and forming temperature of glass. is the maximum value corresponding to the melting temperature, clarification temperature, and molding temperature in the data set. is the minimum value of melting temperature, clarification temperature and molding temperature in the data set. The units of melting temperature, clarification temperature and molding temperature are all in degrees Celsius; Step S23, one-hot encoding is performed on the categorical data of the heat treatment process and the cold treatment process of the glass, wherein the one-hot encoding of the heat treatment process is performed in the following manner: a vector is created for the i-th glass sample If the glass sample is annealed, And order is equal to the annealing temperature, otherwise , are all 0; if the sample adopts the quenching process, then And order is equal to the quenching temperature, otherwise , are all 0; if the sample adopts the tempering process, then And order Equal to the tempering temperature, otherwise , are all 0; then the temperature data corresponding to annealing, quenching and tempering in the heat treatment process are normalized by analogy with formula (3), and the normalized vector is recorded as , ,in, After normalization ; The unique hot encoding of the cold treatment process is as follows: Block glass sample creates a vector If the glass sample is ground, , otherwise let If the glass sample is polished, , otherwise let ; and so on, until all the cold treatment processes used are coded; that is, each additional cold treatment process not listed will be coded. Add a dimension.

Images