Film parameter prediction method based on random configuration network and subtraction averaging optimizer
By combining the method of randomly configuring network and subtraction average optimizer, the problem of low prediction accuracy in traditional algorithms in tire production environments is solved, and more accurate film parameter prediction and stronger quality control capabilities are achieved.
Patent Information
- Application Number
- CN202510025473.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-08
- Publication Date
- 2025-05-16
AI Technical Summary
Traditional neural networks and machine learning algorithms cannot effectively capture the complex and non-local relationships between features in the process of complex tire production environments, resulting in low prediction accuracy and easy to fall into local optimality.
The film parameter prediction method based on the random configuration network (SCN) and subtraction average optimizer (SABO) is used to generate more accurate film parameter prediction values through data preprocessing, correlation analysis, SCN model initialization and SABO optimization process parameters.
It improves the accuracy and practical significance of film parameter prediction, avoids local optimal problems, enhances the ability to control film quality, and improves market competitiveness.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of calender film parameter prediction, and in particular relates to a calender film parameter prediction method based on a random configuration network and a subtraction average optimizer. Background Art
[0002] The calender is a precision general-purpose machine used in tire production and rubber industry. It is mainly used for calendering rubber sheets, rubber sheets, and gluing and rubbing of textiles and steel cord fabrics. According to the number of rollers, the calender can be divided into three-roller and four-roller types. The three-roller calender is mainly used for sheeting, single-sided gluing or gluing; the four-roller calender is widely used for double-sided gluing of cord fabrics in tire manufacturing. In the calendering production process, the production quality of the film is crucial. Slight changes in width and thickness will lead to increased waste and affect the quality of lamination and vulcanization.
[0003] Since the 1990s, artificial intelligence technology has gradually matured, and neural networks, machine learning and other methods have been used for parameter prediction and process optimization in the rubber industry. However, in the complex tire production environment, traditional neural networks and machine learning algorithms cannot effectively capture the complex and non-local relationships between features when processing structured data. Random Configuration Network (SCN) is a neural network that has developed rapidly in recent years. It determines the hidden layer parameters by random methods, thereby simplifying the network training process. The main limitation of SCN as a prediction model is that its weights and biases are randomly generated by the algorithm process parameters, which may lead to the instability of the hidden layer parameters and affect the prediction accuracy and generalization ability.
[0004] In terms of optimization algorithms, many classic methods such as particle swarm optimization algorithm and genetic algorithm have shown certain advantages in optimizing prediction models, but they are prone to fall into local optimality when facing high-dimensional and multi-peak problems, and the convergence speed is difficult to reach a satisfactory solution. Summary of the invention
[0005] The present invention aims at the limited prediction accuracy of the existing methods, and provides a film parameter prediction method based on the combination of random configuration network and subtraction mean optimizer. First, the data is pre-processed, and then the processed data is predicted by SCN-SABO to obtain the predicted value of the film parameter. The prediction result is more accurate and has practical significance.
[0006] To achieve the above object, the present invention adopts the following technical solution, including the following steps:
[0007] Step 1: Load film parameters, including thickness and width, from the data set and clean the data; Step 2: Perform correlation analysis on the input and output data and remove strong and weak correlation features; Step 3: Establish the SABO-SCN model, initialize the unoptimized SCN model, generate the initial population, select the process parameters and shrinkage factors at the corresponding positions, and use SCN to predict the film parameters; Step 4: Obtain the optimal process parameters and shrinkage factors for the current iteration through the subtraction average optimizer; Step 5: Update the optimized parameter vector, generate weights and biases, update the error value, send the test set to the model for prediction, and compare all models to determine the optimal parameters.
[0008] In step 2, analyzing the feature correlation includes the following steps:
[0009] Step 2-1: Calculate the correlation coefficient matrix between characteristic variables and calendered film parameters The correlation measurement method used is the Pearson correlation coefficient, which is expressed as:
[0010]
[0011] In the formula, n is the number of samples, x is i is the characteristic value of the input data set, including the upper and lower chamber pressures of the calender, the glue output speed, temperature, pressure, current and other physical quantities; y i The output film parameter characteristics include the drive side thickness and the take-off width. and Represent the average values of input feature samples and output feature samples respectively.
[0012] In step 3, initializing the unoptimized SCN model includes the following steps:
[0013] Step 3-1: The data set obtained in step 2-1 after removing the strong and weak correlation features is used as input through the random configuration network, and the neuron input weights and biases are determined by the supervision mechanism;
[0014] Step 3-2, calculate the output weight by minimizing the global residual;
[0015] Step 3-3: Select the process parameters and shrinkage factors of the corresponding positions.
[0016] In step 4, optimizing process parameters includes the following steps:
[0017] Step 4-1, initialize the maximum number of iterations, population size, process parameters and upper and lower bounds of the shrinkage factor, initialize the search for the main position and update the position, determine whether the new position of the current iteration replaces the old position prediction, the mathematical expression is as follows:
[0018] x i,d=lb d +r i,d ·(ub d -lb d );
[0019] i=1,…,N,d=1,…,m,
[0020]
[0021] Among them, x i,d is the dth dimension of the search space for the i-th search agent, r i,d is a random number, lb d andub d are the lower and upper bounds of the d-th decision variable, respectively; N is the total number of search individuals; -v is an m-dimensional vector whose components are random numbers selected from {1,2}, and F(A) and F(B) are the objective function values of search agents A and B, respectively.
[0022] In step 5, determining the final model includes the following steps:
[0023] Step 5-1, update the process parameter vector, and update the error value by the random configuration network, send the test set to the corresponding model for prediction to obtain the prediction result, compare the fitness values of all models, determine the optimal parameters and the final SCN model and save them to obtain the prediction model.
[0024] Compared with the prior art, the present invention has the following beneficial effects:
[0025] The method proposed in the present invention improves the problems of low prediction accuracy and easy falling into local optimality of traditional prediction models when facing problems such as large data fluctuations and complex data relationships in actual industrial production environments. Compared with traditional neural network models, the random configuration network dynamically adjusts the configuration of the network according to the characteristics of the input data. It gradually increases the number of hidden layer nodes through a supervision mechanism to better adapt to different data distributions and perform parameter prediction while avoiding global optimization problems. The calender film parameter prediction method based on the random configuration network and the subtraction average optimizer has more accurate prediction capabilities, can better realize the early prediction of film quality, further deepen the company's control over tire film quality, and improve market competitiveness. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] The present invention is further described below in conjunction with the accompanying drawings and specific implementation methods. The protection scope of the present invention is not limited to the following descriptions.
[0027] Figure 1 It is a basic model structure diagram of the present invention.
[0028] Figure 2 It is a flow chart of the prediction model proposed by the present invention.
[0029] Figure 3 It is a characteristic correlation coefficient diagram of thickness data of the present invention.
[0030] Figure 4 It is a characteristic correlation coefficient diagram of width data of the present invention.
[0031] Figure 5 It is a graph showing the thickness data prediction results of the calender 1 of the present invention.
[0032] Figure 6 It is a graph showing the thickness data prediction results of the calender 2 of the present invention.
[0033] Figure 7 It is a graph showing the prediction results of the width data of the calender 1 of the present invention.
[0034] Figure 8 It is a graph showing the prediction results of the width data of the calender 2 according to the present invention.
[0035] Fig. 9 It is the average RMSE curve of the present invention when the number of hidden layer nodes is different.
[0036] Fig.10 It is the average MAE curve of the present invention when the number of hidden layer nodes is different. DETAILED DESCRIPTION
[0037] The present invention combines the basic prediction model and the optimization algorithm to propose a design method:
[0038] Firstly, the film parameters are obtained, and feature selection is performed through correlation analysis to screen the key data in the film extrusion calendering process, and these feature data are input into the input layer of the random configuration network. The SCN model process parameters are searched by the subtraction average optimizer (SABO) to obtain the optimal solution and generate output weights to obtain the final prediction result. The present invention is described in detail below with reference to the accompanying drawings.
[0039] Specifically, the random configuration network structure diagram is as follows Figure 1 As shown in Figure 1, it contains input layer, hidden layer and output layer, which reflects its incremental configuration method and the operation mode of dynamically adding hidden nodes. Figure 2 As shown, the method includes:
[0040] S1, obtain the film production process data of the calender and extruder, check whether there are extreme missing values or abnormal values in the data, if so, use the average value of the five rows of data above and below the location to replace it, and get the initial data sequence.
[0041] Taking the 2023 tire calender and extruder parameter data set of a domestic company as an example, the platform samples the width data and thickness data of the films of the two calenders, including internal working current, temperature, pressure value and other characteristics. The feature dimensions of the thickness and width data sets are 24 and 21 dimensions respectively. The collected data are screened for abnormal data, and abnormal values with a difference of more than 3 standard deviations from the feature average are eliminated.
[0042] In order to analyze the prediction results of the model, the target variables (film thickness and width) were statistically analyzed before modeling. After removing outliers from the training set and data set obtained above and performing feature selection, the correlation coefficient matrix between the feature variables and the calendered film parameters was calculated. The correlation measurement method used in this work is the Pearson correlation coefficient, which is expressed as
[0043]
[0044] In the formula, n is the number of samples, x is i It is the characteristic value of the input data set, including the upper and lower chamber pressures of the calender, the glue output speed, temperature, pressure, current and other physical quantities. i The output film parameter characteristics include the drive side thickness and the take-off width. and Represent the average values of input feature samples and output feature samples respectively.
[0045] The correlation coefficient is shown in Figure 3-4 , the horizontal axis is the feature number, and the vertical axis is the correlation coefficient between the input feature and the output feature. Most of the features of the thickness data are negatively correlated with the output features and the correlation is within -0.25. The plasticizing temperature control temperature of feature 20 extruder is constant, and the correlation is 0. In order to improve the generalization ability of the model, the input data in the prediction model does not introduce features with correlations higher than 0.9 or lower than 0.1. Most of the input features of the width data are significantly linearly positively correlated with the prediction parameters, and the correlation is between 0.2-0.4. Weakly correlated and strongly correlated data are deleted to improve the model prediction ability and determine the final feature subset.
[0046] S2.1, send the training set to the SABO-SCN model, set the SCN initial parameters and save the results. The initial parameter settings include the initial r and λ, the maximum hidden layer node L_max, the maximum random assignment number T_max, and the error threshold tol.
[0047] The random configuration network (SCN) is an emerging incremental neural network model. SCN is adaptive in the selection of network structure and parameters. It dynamically adjusts the network configuration according to the characteristics of the input data to better adapt to different data distributions. Traditional neural network models rely on global optimization algorithms such as back propagation and gradient descent. SCN gradually increases the number of hidden layer nodes through a supervision mechanism to further optimize node parameters.
[0048] The implementation steps of randomly configuring the network are as follows:
[0049] (1) The input data is a training data set {X,Y} with N samples and p features, where X = [x1, x2, …, x N ] is the input data, where i=1,2,…,N,output data set Y=[y1,y2…,y n ], the input weight is expressed as b j and β j ,j=1,2,…L Max Represent the input bias and output weight respectively, input weight w j and deviation b j are randomly selected from [-λ,λ], λ>0. L Max Represents the maximum number of hidden neurons, which is a finite value hyperparameter. The network structure diagram is as follows Figure 1 Shown
[0050] Among them, h j =[g(w j x1+b j ),…,g(w j x N +b j )] T , g is the activation function, given an objective function f, the output when the number of hidden layer nodes is increased to L-1 is:
[0051]
[0052] (2) The input weight and bias of the Lth hidden neuron are determined by the following supervision mechanism:
[0053]
[0054] Among them, e L-1 represents f L and f L-1 The residual between , 0<r<1, and a non-negative real number sequence μ L , and μ L <1-r, limμ L= 0. The candidate node with the maximum value is used as the L-th node parameter.
[0055] (3) After adding the L-th hidden layer, the output weight β is calculated by minimizing the global residual, and finally the predicted value is calculated through the output layer. L The calculation of, and finally the predicted value is calculated through the output layer.
[0056]
[0057] Among them, is obtained by the least squares method, represents the generalized inverse of H.
[0058] S2.2. Initialize the parameters of the optimization algorithm SABO model, including initializing the maximum number of iterations, the population size Np, and the upper bound u b and the lower bound l b .
[0059] S3.1. Obtain the best process parameters r and the contraction factor λ of the current iteration through the subtraction mean optimizer. The specific steps are as follows:
[0060] (1) Define the outer loop while t < iter_num and the inner double loop for i = 1:Np for j = 1:Np.
[0061] (2) Initialize the main position of the search SCN process parameters through Equation 5
[0062] x i,d = lb d + r i,d · (ub d - lb d )
[0063] i = 1,…, N, d = 1,…, m, ⑸
[0064] Among them, x i,d is the d-th dimension (decision variable) of the i-th search agent (population member) in the search space, which can be regarded as a candidate solution to the problem. r i,d is a random number, lb d and ub d are the lower and upper bounds of the d-th decision variable respectively. N is the total number of search individuals.
[0065] S3.2. The algorithm designs a new calculation concept, "-v", called "v-subtraction" of search agent A and search agent B, and its definition is as Equation (6), and the position is updated by Equation (7).
[0066]
[0067] Where -v is an m-dimensional vector whose components are random numbers selected from {1,2}, and F(A) and F(B) are the objective function values of search agents A and B, respectively.
[0068] The particle fitness value fit_new_P1 is calculated based on the new position, and formula (8) is used to determine whether the new position in this iteration replaces the old position.
[0069]
[0070] F i new and F i They are particle X i and X i new The objective function value of .
[0071] S3.3, determine whether the current number of nodes reaches or the error is less than the set threshold.
[0072] S3.4, when L is less than the setting or the error is greater than the threshold, configure the hidden parameters, increase the number of hidden layer nodes, and calculate the weights and biases. Return to S3.2.
[0073] S3.5, when the error meets the requirement or L reaches the maximum, determine whether the population reaches Np or the number of iterations reaches the maximum.
[0074] S4, when the population size reaches the maximum or the number of iterations reaches the maximum, update the vector v i = {v iγ ,v iλ}, and according to the vector v i The weights and biases are randomly generated and selected and saved by formula (3), and the output weight update error value is calculated by formula (3). The test set is sent to the corresponding model for prediction to obtain the prediction result.
[0075] S5, compare the fitness values f of all models best , determine the optimal parameter v best = {v best,γ ,v best,λ} and the final SCN model and save it.
[0076] In order to verify the prediction performance of the method proposed in this work, and because of the similarities in the strategies for updating the solutions, this work introduces two optimization algorithms, the vector weighted average algorithm (INFO) and the particle swarm algorithm (PSO), which are used in combination with SCN and compared with SCN-SABO to further explore the potential and effect of each optimization algorithm in improving the performance of SCN.
[0077] The vector weighted average algorithm (info) is an optimization algorithm that iteratively searches for the optimal solution through update rules, vector combination and local search. The update rules are used to generate the basic vectors of candidate solutions, and different vectors are combined by weighted average. Assume that the current iteration number and solution vector position are g and l , and its update rule is
[0078] MeanRule=k·WM1 gl +(1-k)WM2 gl ⑼
[0079] WM1 gl and WM2 gl are the weighted differences of two different sets of vectors, and k is the contribution coefficient for adjusting the two. After generating a new vector, the new vector is fused with the current solution through vector combination to generate a better solution. The final local search is a detailed search in the neighborhood of the current solution, which enhances the local optimization ability of the algorithm. Combining the info algorithm with SCN to search for parameter vectors can effectively explore the solution space.
[0080] The particle swarm optimization algorithm simulates the process of bird flocks preying in nature, and finds the global optimal solution to the problem through team collaboration. The bird flock corresponds to the population size, the foraging space corresponds to the search space dimension D of the problem, and the flight speed and position represent the speed vector of the solution respectively. and the position vector Each particle i updates its speed and position according to its own historical optimal position and the global optimal position of the group and finds the global optimal solution. In SCN-PSO, each particle represents a set of potential weights and biases in the network, and the particle's position and velocity vectors correspond to the current estimated value and update direction of this set of parameters, respectively.
[0081] The root mean square error (RMSE) is selected as the model loss function, and the model evaluation function takes the mean absolute error (MAE), which is expressed as:
[0082]
[0083] Where: n is the number of samples, y i is the actual value of the i-th sample, is the predicted value of the ith sample. Figure 5-8 They are the prediction results of the film width and thickness data of calender 1 and calender 2 respectively. Table 1 shows the prediction results of the hidden layer nodes reaching L max The results of each evaluation index are displayed at the same time.
[0084] Table 1 Evaluation index table of each model
[0085]
[0086] like Figures 5 to 8 As shown in Table 1, the horizontal axis is the sample index and the vertical axis is the value, including the true value and the predicted value. The average absolute error of the thickness data of calender 1 is reduced by 44.27% and the root mean square error is reduced by 56.22% relative to the SCN model. The average absolute error of the thickness data of calender 2 is reduced by 43.1% and the root mean square error is reduced by 50.63%. The average absolute error of the width data of calender 1 is reduced by 14.88% and the root mean square error is reduced by 27.93% relative to the SCN average absolute error. The average absolute error of the width data of calender 2 is reduced by 15.19% and the root mean square error is reduced by 25.69%.
[0087] Fig. 9 and Fig.10 The average RMSE and MAE of the thickness parameters predicted by the test set of calender 1 are shown, where the horizontal axis is the number of hidden layer nodes and the vertical axis is the evaluation index under the current node. The model is configured with 30 hidden nodes. Experiments show that the convergence speed of SCN-SABO is faster than that of SCN-INFO and SCN-PSO, and when the hidden layer is added to 30 nodes, the average absolute error and root mean square error of the calender parameter prediction model based on SCN-SABO have more obvious effects compared with other prediction models. Compared with the SCN-SABO algorithm, the SCN network optimized by the INFO algorithm performs better when the number of hidden layers is small, but the convergence speed decreases with the increase of the number of hidden layers. The complexity of the introduced optimization model is greater than that of SABO, and the improper selection of key parameters leads to slow convergence or falling into the local optimum; the performance of the PSO algorithm is sensitive to the initial value, and different initial populations may cause differences in the potential optimal solutions of the problem space during the search process, which affects the final performance of the model. Figure 5 The reason why SCN-SPO has high error in the optimization process.
[0088] In summary, this paper proposes a tire film parameter prediction model based on SCN and subtraction average optimizer (SABO). First, the data set is subjected to correlation analysis, dimensionality reduction, and outlier processing. The low-dimensional eigenvalues obtained are used as the input values of SCN. The SABO algorithm is used to optimize the SCN model process parameters to obtain the optimal solution of the film parameter model. By comparing different model optimization strategies, the problem that the parameter setting of SCN has a great influence on the network performance is solved and the practicality and effect of the method are verified. Therefore, this combined prediction method is suitable for industrial production and has a higher reference value.
[0089] It can be understood that the above specific description of the present invention is only used to illustrate the present invention and is not limited to the technical solutions described in the embodiments of the present invention. Those skilled in the art should understand that the present invention can still be modified or replaced by equivalents to achieve the same technical effects; as long as the use requirements are met, they are within the protection scope of the present invention.
Claims
1. Film parameter prediction method based on random configuration network and subtraction average optimizer, characterized by The following steps are involved: Step 1: Load film parameters, including thickness and width, from the data set and clean the data; Step 2: Perform correlation analysis on the input and output data and remove strong and weak correlation features; Step 3: Establish the SABO-SCN model, initialize the unoptimized SCN model, generate the initial population, select the process parameters and shrinkage factors at the corresponding positions, and use SCN to predict the film parameters; Step 4: Obtain the optimal process parameters and shrinkage factors for the current iteration through the subtraction average optimizer; Step 5: Update the optimized parameter vector, generate weights and biases, update the error value, send the test set to the model for prediction, and compare all models to determine the optimal parameters.
2. The film parameter prediction method based on random configuration network and subtraction average optimizer according to claim 1 is characterized in that: In step 2, analyzing the feature correlation includes the following steps: Step 2-1: Calculate the correlation coefficient matrix between characteristic variables and calendered film parameters The correlation measurement method used is the Pearson correlation coefficient, which is expressed as: In the formula, n is the number of samples, x is i is the characteristic value of the input data set, including the upper and lower chamber pressures of the calender, the glue output speed, temperature, pressure, current and other physical quantities; y i To produce film parameter characteristics, including drive side thickness and take-off width, and Represent the average values of input feature samples and output feature samples respectively.
3. The film parameter prediction method based on random configuration network and subtraction average optimizer according to claim 2 is characterized in that: In step 3, initializing the unoptimized SCN model includes the following steps: Step 3-1: The data set obtained in step 2-1 after removing the strong and weak correlation features is used as input through the random configuration network, and the neuron input weights and biases are determined by the supervision mechanism; Step 3-2, calculate the output weight by minimizing the global residual; Step 3-3: Select the process parameters and shrinkage factors of the corresponding positions.
4. The film parameter prediction method based on random configuration network and subtraction average optimizer according to claim 1 is characterized in that: In step 4, optimizing process parameters includes the following steps: Step 4-1, initialize the maximum number of iterations, population size, process parameters and upper and lower bounds of the shrinkage factor, initialize the search for the main position and update the position, determine whether the new position of the current iteration replaces the old position prediction, the mathematical expression is as follows: x i,d =lb d +r i,d ·(ub d -lb d ); i=1,…,N,d=1,…,m, Among them, x i,d is the dth dimension of the search space for the i-th search agent, r i,d is a random number, lb d andub d are the lower and upper bounds of the d-th decision variable, respectively; N is the total number of search individuals; -v is an m-dimensional vector whose components are random numbers selected from {1,2}, and F(A) and F(B) are the objective function values of search agents A and B, respectively.
5. The film parameter prediction method based on random configuration network and subtraction average optimizer according to claim 1, characterized in that: In step 5, determining the final model includes the following steps: Step 5-1, update the process parameter vector, and update the error value by the random configuration network, send the test set to the corresponding model for prediction to obtain the prediction result, compare the fitness values of all models, determine the optimal parameters and the final SCN model and save them to obtain the prediction model.