Submerged arc furnace electrode insertion depth soft measurement method based on radial basis function neural network

Through the soft measurement method based on the radial basis function neural network, the problem of difficult to accurately measure the electrode insertion depth of the mineral furnace is solved, and the precise measurement and stability adjustment of the electrode insertion depth are realized, which improves the stability and data of the smelting process reliability.

CN119066320APending Publication Date: 2024-12-03DALIAN HEAVY MECHANICAL & ELECTRICAL EQUIP ENG CO +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411137248.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-19
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The electrode insertion depth of the ore furnace is difficult to accurately measure, resulting in three phases imbalance and difficulty in adjusting, affecting the stability of the smelting process.

Method used

Using a soft measurement method based on radial basis function neural network, the electrical and working conditions parameters are reduced in dimensionality through the PLS unit, and the soft measurement calculation of electrode insertion depth is used by the RBFN unit. Combined with the online correction mechanism, the model is adjusted in real time to improve the Measurement accuracy.

Benefits of technology

It realizes accurate measurement of the electrode insertion depth of the ore furnace, improves the stability of the production process and data reliability, can effectively track changes in production conditions, and reduces errors and adjustment difficulties.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066320A_ABST
    Figure CN119066320A_ABST
Patent Text Reader

Abstract

The invention provides a submerged arc furnace electrode insertion depth soft measurement method based on a radial basis function neural network. The method comprises the following steps that S1, electrical parameters and working condition parameters in the production and operation process of the submerged arc furnace are collected, and the electrical parameters and the working condition parameters are sent to a PLS unit; s2, the PLS unit performs dimension reduction processing on the electrical parameters and the working condition parameters, and sends low-dimension data after dimension reduction to an RBFN unit; and S3, the RBFN unit receives the low-dimensional data, performs soft measurement calculation on the electrode insertion depth based on the low-dimensional data, and outputs a real-time electrode insertion depth estimation value. According to the invention, soft measurement of the insertion depth of the electrode can be realized, technical parameter guidance is provided for technological operators or data obtained by soft measurement is used for other automatic electrode control systems, production is stabilized, economic benefits are improved, and the overall intelligent level of the submerged arc furnace is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of ore-bearing furnace smelting, and in particular to a soft measurement method for electrode insertion depth of an ore-bearing furnace based on a radial basis function neural network. Background Art

[0002] At present, the design and production capacity of smelting submerged arc furnaces is huge, and some have a capacity close to 100MVA or more. Generally, three AC single-phase transformers are used for power supply. The primary high-voltage side is connected to the high-voltage power supply system by star connection or delta connection, and the secondary low-voltage side is connected to three electrodes by delta connection. The production furnace condition is closely related to the electrode insertion depth. Controlling the electrode insertion depth is the main means to ensure the stability of the furnace condition. The electrode is inserted into the charge for submerged arc smelting, and the electrode tip is constantly consumed. The actual insertion depth is difficult to measure. The process operator needs to roughly estimate the electrode insertion depth based on many parameters such as electrode current, voltage, active power, furnace body temperature, flue gas temperature, etc. The empirical method is used for operation, which often causes three-phase imbalance and is difficult to adjust. Therefore, in order to obtain the actual insertion depth, the production personnel need to shut down the power every 1 to 2 days to measure the actual insertion depth. Some of them are calculated by inserting a steel chisel according to the angle, while others are directly pulling out all the electrodes and reinserting them into the furnace after the measurement is completed. It is still difficult to accurately measure during the production process. Therefore, it is necessary to invent a method for accurately measuring the electrode insertion depth value of the submerged arc furnace. Summary of the invention

[0003] According to the technical problem that the electrode insertion depth of an AC submerged arc furnace cannot be accurately measured, a submerged arc furnace electrode insertion depth soft measurement method based on a radial basis function neural network is provided. The present invention mainly utilizes an intelligent algorithm for the electrode insertion depth of a submerged arc furnace based on soft measurement technology to realize soft measurement estimation of the electrode insertion depth of a submerged arc furnace, and provides reliable and effective data support for submerged arc furnace production.

[0004] The technical means adopted by the present invention are as follows:

[0005] A soft measurement method for electrode insertion depth of an electric arc furnace based on a radial basis function neural network comprises the following steps:

[0006] S1. Intercept a section of data before power outage for each measurement of the actual electrode insertion depth as historical data and online correction data, collect electrical parameters and operating parameters during the production and operation of the submerged arc furnace, and send the electrical parameters and operating parameters to the PLS unit;

[0007] S2, the PLS unit performs dimensionality reduction processing on the electrical parameters and the operating parameters, and sends the low-dimensional data after dimensionality reduction to the RBFN unit;

[0008] S3, the RBFN unit receives low-dimensional data and performs soft measurement calculation of electrode insertion depth based on the low-dimensional data and outputs it, obtains the RBF network structure and parameter model through calculation and stores it in the RBFN unit, the RBF network structure and parameter model include the number of hidden layer centers, center positions, and weights of the model;

[0009] S4. Obtain a real-time electrode insertion depth estimation value based on the connection weight vector W.

[0010] Furthermore, it also includes:

[0011] S5. The sampling unit samples the actual value of the electrode insertion depth of the electric arc furnace at different intervals. The online correction unit compares the actual value of the electrode insertion depth with the estimated value of the electrode insertion depth to obtain a comparison difference. When the comparison difference exceeds the specified difference range, the comparison difference and the production conditions in the corresponding section are recorded. When the generation of the comparison difference exceeds the rated number of times and the corresponding production conditions are always in a stable operating state, the adjustment amount is obtained according to the comparison difference, and the adjustment amount is superimposed on the output layer node of the soft measurement instrument. At the same time, the samples when the deviation occurs are recorded so that new samples can be added for retraining of the network after accumulating a certain number.

[0012] Furthermore, S2 specifically includes the following steps:

[0013] S200, initialization program, selecting electrode voltage, electrode current, electrode power, electrode bearing chamber pressure of electrode lifting hydraulic cylinder, furnace bottom temperature, flue gas temperature, oxygen content in flue gas, hydrogen in flue gas as electrical parameters and working condition parameters as input variable X, selecting the actually measured three-phase electrode insertion depth as network output variable Y, and performing column-by-column standardization / normalization processing on X and Y;

[0014] S201, let E = X, F = Y, u equal to a column of Y, or take the column vector with the largest variance;

[0015] S202, calculate input weight vector w=E T u / (u T u);

[0016] S203, normalizing the input weight vector w=w / ||w||;

[0017] S204, calculating the characteristic vector t=Ew;

[0018] S205, calculate output load vector q=F T t / (t T t);

[0019] S206, normalizing the output load vector q=q / ||q||;

[0020] S207, calculate output feature vector u=Fq T / (q T q);

[0021] S208, repeat steps S202 to S207 until convergence;

[0022] S209, calculate the internal model regression coefficient b=u T t / (t T t);

[0023] S210, calculate X load vector p=bX T t / (bt T bt);

[0024] S211, calculate the residual matrix E = E-btp T and F = F-btq T ;

[0025] S212, return to S202 and execute in a loop until all eigenvectors are calculated;

[0026] S213. Sort the eigenvectors by eigenvalue from large to small, select the eigenvectors corresponding to the first k eigenvalues ​​with the largest contribution to the overall feature, use these k eigenvectors as row vectors to form a dimensionality reduction matrix, multiply the dimensionality reduction matrix by the original matrix on the left to complete the dimensionality reduction, and the obtained data is the data after dimensionality reduction in step S2.

[0027] Further, in S208:

[0028] When the difference between the eigenvector and the previous one within 10 times is within the allowed range, it is judged to be converged.

[0029] Furthermore, in S3, the RBFN unit adopts the RBF network, which is a two-layer forward network. The number of inputs is equal to the number of independent variables under study. The middle layer selects the basis function as the transfer function. The transformation from the input layer to the hidden layer space is nonlinear, and the transformation from the hidden layer to the output layer is linear. The transformation function of the hidden layer unit is a locally distributed non-negative nonlinear function with radial symmetric decay about the center, that is, the radial basis function, and the output layer is a linear combiner.

[0030] Furthermore, S3 specifically includes the following steps:

[0031] S300, initialization program;

[0032] S301, select a hidden layer unit number M;

[0033] S302, select a group of RBF center vectors;

[0034] S303, according to the central vector, using the data output by the dimension reduction in step S2 as the input sample vector, calculate the regression matrix P;

[0035] S304, orthogonalizing each column of the regression matrix;

[0036] S305, calculate separately and ε i ;

[0037] S306, orthogonally decompose P to obtain an upper triangular matrix A, and solve the connection weight vector W by the triangular equation g=AW, where g=[g 1 , g 2 , ..., g M ] T ;

[0038] S307, check whether it is satisfied Wherein 0<ρ<1 is the selected tolerance, if it is satisfied, the calculation is stopped, otherwise jump to S302 to continue execution.

[0039] The present invention also provides a storage medium, which includes a stored program, wherein when the program is run, any of the above-mentioned soft measurement methods for electrode insertion depth of an electric arc furnace based on a radial basis function neural network is executed.

[0040] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes any one of the above-mentioned soft measurement methods for electrode insertion depth of an electric arc furnace based on a radial basis function neural network through the computer program.

[0041] Compared with the prior art, the present invention has the following advantages:

[0042] The present invention adopts the RBF (Radial Basis Function) radial basis function neural network in ANN, so that the input vector can be directly (i.e. not connected by weight) mapped to the hidden space. When the center point of RBF is determined, this mapping relationship is also determined; the mapping from the hidden layer space to the output space is linear, that is, the output of the network is the linear weighted sum of the outputs of the hidden units, and the weight here is the network adjustable parameter; in general, the mapping from input to output of the network is nonlinear (relatively slow learning speed), while the network output is linear for the adjustable parameters (faster learning speed). RBF network has the properties of global approximation and the best approximation performance at the same time, and is very suitable for real-time identification and control of the system. The model training speed of RBF neural network is relatively fast, and online model learning can be realized, that is, a batch of data under new working conditions is collected online regularly, and the model is retrained. Then, the model's hidden layer center number, center position, weight and other parameters are completely regenerated according to the latest data. Each training removes the oldest data, absorbs the latest data, and iterates repeatedly. The soft measurement model can be corrected in time through learning. This method enables the model to closely track the production conditions and can effectively solve the problem of production condition drift.

[0043] The present invention uses the partial least squares algorithm (Partial Least Square, PLS) to simultaneously perform orthogonal decomposition on the input / output data sets, so that PLS can obtain as much information as possible from fewer load vectors. It is a single-step decomposition and regression method. Compared with the principal component analysis (Principal Component Analysis, PCA) algorithm that only performs orthogonal decomposition on the input variables, PLS emphasizes the predictive effect of the input on the output, removes noise that is not helpful for regression, makes the model contain the least number of variables, has better robustness and prediction stability, and can further improve the accuracy of system modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 The figure is a flow chart of the method of the present invention.

[0046] Figure 2 This is a step diagram of the PLS learning algorithm of the present invention.

[0047] Figure 3 This is the RBF network diagram of the present invention.

[0048] Figure 4 This is a step diagram of the RBF-OLS learning algorithm of the present invention. DETAILED DESCRIPTION

[0049] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0050] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0052] Unless otherwise specifically stated, the relative arrangement of the parts and steps described in these embodiments, the numerical expressions and numerical values ​​do not limit the scope of the present invention. At the same time, it should be clear that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The technology, methods and equipment known to ordinary technicians in the relevant field may not be discussed in detail, but in appropriate cases, the technology, methods and equipment should be regarded as part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar numbers and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further discussed in subsequent drawings.

[0053] In addition, it should be noted that the use of terms such as "first" and "second" to limit components is only for the convenience of distinguishing the corresponding components. If not otherwise stated, the above terms have no special meaning and therefore cannot be understood as limiting the scope of protection of the present invention.

[0054] The intelligent algorithm for the electrode insertion depth of a DC ore-fired furnace based on soft measurement technology provided by the present invention adopts PLS-RBFN to realize soft measurement of the three-phase electrode insertion depth and can realize online correction, thereby providing reliable and effective data support for the production of DC ore-fired furnaces.

[0055] Artificial Neural Network (ANN) has large-scale parallelism, redundancy, fault tolerance, inherent nonlinearity, self-organization, and self-learning adaptability. Using artificial neural networks for information processing can reduce data analysis and modeling work, and complete complex nonlinear mapping in a non-programmed, adaptive processing method, making it have good application prospects in the field of automatic control. The application of ANN does not require prior knowledge of the object, but directly models the object based on its input and output data; its unique non-traditional expression method and inherent learning ability give it great potential in solving highly nonlinear and severely uncertain system control. The basic idea of ​​the RBF (Radial Basis Function) neural network in ANN is to use RBF as the "basis" of the hidden unit to form the hidden layer space, so that the input vector can be directly (that is, not connected through weights) mapped to the hidden space. When the center point of the RBF is determined, this mapping relationship is also determined; and the mapping from the hidden layer space to the output space is linear, that is, the output of the network is the linear weighted sum of the output of the hidden unit, and the weight here is the network adjustable parameter; in general, the mapping from input to output of the network is nonlinear (relatively slow learning speed), while the network output is linear to the adjustable parameters (fast learning speed). RBF network has both the properties of global approximation and the best approximation performance, which is very suitable for real-time identification and control of the system.

[0056] The model training speed of RBF neural network is relatively fast, and online model learning can be realized, that is, a batch of data under new working conditions is collected online regularly, and the model is retrained. Then, the model's hidden layer center number, center position, weight and other parameters are completely regenerated according to the latest data. Each training removes the oldest data, absorbs the latest data, and iterates repeatedly. The soft measurement model can be corrected in time through learning. This method enables the model to closely track the production conditions and can effectively solve the problem of production condition drift.

[0057] The Partial Least Square (PLS) algorithm performs orthogonal decomposition on the input / output data sets simultaneously, so that PLS can obtain as much information as possible from fewer load vectors. It is a single-step decomposition and regression method. Compared with the Principal Component Analysis (PCA) algorithm that only performs orthogonal decomposition on the input variables, PLS emphasizes the predictive effect of input on output, removes noise that is not helpful for regression, and makes the model contain the least number of variables, with better robustness and predictive stability, and can improve the accuracy of system modeling.

[0058] The soft measurement model is different from the general mathematical model. It emphasizes obtaining the best estimate of the target dominant variable through auxiliary variables. The data of each measurement of the actual three-phase electrode insertion depth before the power outage is intercepted as the historical data and online correction data. The electrical parameters and working condition parameters such as electrode voltage, electrode current, electrode power, electrode bearing chamber pressure of electrode lifting hydraulic cylinder, furnace bottom temperature, flue gas temperature, oxygen content in flue gas, and hydrogen in flue gas are selected as input variables. The actual measured three-phase electrode insertion depth is selected as the network output variable. The PLS-RBFN network model method is used for soft measurement of the three-phase electrode insertion depth. The partial least squares algorithm PLS is used to preprocess the high-dimensional input variables, and the low-dimensional principal component variables reflecting the process information are constructed. Then, the RBF radial basis function neural network is used to establish a self-correction model for the principal component variables. This method not only simplifies the structure of the neural network model, but also can detect the data error caused by instrument and process failures with the help of principal component analysis method to avoid the wrong output of the model, which is of great significance to the model-based optimization control.

[0059] Based on the above, if Figure 1 As shown, the present invention provides a soft measurement method for electrode insertion depth of an electric arc furnace based on a radial basis function neural network, comprising the following steps:

[0060] S1. Collect electrical parameters and operating parameters during the production and operation of the submerged arc furnace, and send the electrical parameters and operating parameters to the PLS unit;

[0061] S2, the PLS unit performs dimensionality reduction processing on the electrical parameters and the operating parameters, and sends the low-dimensional data after dimensionality reduction to the RBFN unit;

[0062] S3, the RBFN unit receives low-dimensional data and performs soft measurement calculation of electrode insertion depth based on the low-dimensional data, and outputs an RBF network structure and parameter model, wherein the RBF network structure and parameter model include the number of hidden layer centers, center positions, and weights of the model;

[0063] S4. When running online, new data will be used as data input for the RBF network structure and parameter model after PLS dimension reduction, and the estimated value of electrode insertion depth can be output after calculation.

[0064] It also includes: S5, the sampling unit samples the actual value of the electrode insertion depth of the electric arc furnace at different intervals, and the online correction unit compares the actual value of the electrode insertion depth with the estimated value of the electrode insertion depth to obtain a comparison difference; when the comparison difference exceeds the specified difference range, the comparison difference and the production conditions in the corresponding section are recorded; when the generation of the comparison difference exceeds the rated number of times and the corresponding production conditions are always in a stable operating state, an adjustment amount is obtained according to the comparison difference, and the adjustment amount is superimposed on the output layer node of the soft measurement instrument, and the samples when the deviation occurs are recorded at the same time, so as to add new samples for retraining the network after accumulating a certain number.

[0065] In S2, PCA first performs principal component analysis on the input data set X(m×n) to obtain the principal component matrix T, and then makes the output data set Y( n ×p) regresses T. When extracting information, principal component analysis decomposes X independently without considering Y. PLS considers both X and Y on the basis of PCA, and performs orthogonal decomposition on X and Y at the same time, so that PLS can obtain as much information as possible from fewer load vectors.

[0066]

[0067]

[0068] u 1 =f 1 (t 1 )+r 1

[0069] Where n-dimensional vector t 1 、u 1 is the first pair of input and output feature vectors, the m-dimensional vector p 1 ,q 1 is the load vector corresponding to the child, E 1 、F 1 is the residual matrix. The goal of orthogonal decomposition is to make ||E 1 ||、||F 1 ||Minimize, function f 1 () selects the target to make || r 1 ||Minimize.

[0070] Compute the residual:

[0071]

[0072]

[0073] Then to E 1 、F 1 Perform orthogonal decomposition and repeat the above steps until the required kth eigenvector is extracted. At this time, E k 、F k It contains almost no useful information. After the above iterative calculation, we can get the external relationship between X and Y:

[0074]

[0075]

[0076] And the internal relations:

[0077] U=f(T)+R=[f 1 (t 1 )f 2 (t 2 )…f k (t k )]+[r 1 r 2 …r k ]

[0078] In summary, if Figure 2 As shown in Figure 1, the commonly used PLS algorithm is a nonlinear iterative partial least squares method. Assuming that X and Y have been standardized (normalized) by column, the specific data dimensionality reduction steps of the partial least squares algorithm (PLS) are as follows:

[0079] S200, initialization program, select the electrode voltage, electrode current, electrode power, electrode bearing chamber pressure of the electrode lifting hydraulic cylinder, furnace bottom temperature, flue gas temperature, oxygen content in flue gas, hydrogen in flue gas and other electrical parameters and working condition parameters as input variable X, select the actually measured three-phase electrode insertion depth as the network output variable Y, and perform standardization / normalization processing on X and Y by column;

[0080] S201, let E = X, F = Y, u equal to a column of Y, or take the column vector with the largest variance;

[0081] S202, calculate input weight vector w=E T u / (u T u);

[0082] S203, normalizing the input weight vector w=w / ||w||;

[0083] S204, calculate characteristic vector t=E w ;

[0084] S205, calculate output load vector q=F T t / (tT t);

[0085] S206, normalizing the output load vector q=q / ||q||;

[0086] S207, calculate output feature vector u=Fq T / (q T q);

[0087] S208, repeat steps S202 to S207 until convergence. The way to check convergence is to see whether the difference between t and the last time is within the allowable range. Usually, convergence will be achieved within no more than 10 times.

[0088] S209, calculate the internal model regression coefficient b=u T t / (t T t);

[0089] S210, calculate X load vector p=bX T t / (bt T bt);

[0090] S211, calculate the residual matrix E = E-btp T and F = F-btq T ;

[0091] S212, return to S202 and execute in a loop until all eigenvectors are calculated.

[0092] S213. Sort the eigenvectors by eigenvalue from large to small, select the eigenvectors corresponding to the first k eigenvalues ​​with the largest contribution to the overall feature, use these k eigenvectors as row vectors to form a dimensionality reduction matrix, multiply the dimensionality reduction matrix by the original matrix on the left to complete the dimensionality reduction, and the obtained data is the data after dimensionality reduction in step S2.

[0093] In S3, the RBF network used is as follows: Figure 3 As shown in the figure, it is a two-layer forward network. The number of inputs is equal to the number of independent variables studied. The middle layer selects the basis function as the transfer function. The transformation from the input layer to the hidden layer space is nonlinear, and the transformation from the hidden layer to the output layer is linear. The transformation function of the hidden layer unit is a locally distributed non-negative nonlinear function with radial symmetric decay about the center, that is, the radial basis function. The output layer is a linear combiner. For this RBF network, the orthogonal least square (OLS) learning algorithm is used. This method is simple and easy to use, and has a fast calculation speed. In order to simplify the calculation process, it is assumed that there is only one unit in the network output layer, and the network training pair is {X n , d (n)}(n=1,2,...,N). Where N is the number of training samples; Xn ∈R ni is the input data vector of the network; d(n)∈R 1 is the expected output response of the network. According to the linear regression model, the expected output response of the network can be expressed as

[0094]

[0095] Where M is the number of hidden layer units, M<N; p i (n) is the regression operator, which is actually the response of the hidden layer RBF under certain parameters and can be expressed as

[0096] p i (n) = G(||X n -t i ||) (n=1, 2, ..., N; i=1, 2, ..., M)

[0097] w i is the model parameter, which is actually the connection weight between the output layer and the hidden layer; e(n) is the residual.

[0098] Written in matrix form

[0099] d=PW+e

[0100] in:

[0101] d=[d(1), d(2),..., d(N)] T

[0102] W=[w 1 , w 2 , ..., w M ]

[0103] P=[P 1 , P 2 , ..., P M ]

[0104] p i =[p i (1), p i (2), ..., p i (N)]

[0105] e=[e(1),e(2),...,e(N)] T

[0106] Where P is the regression matrix. The key problem in solving the above regression equation is the regression operator vector p i The center of the RBF is generally a subset of the input sample data vector set, and each set of centers can obtain a regression matrix P corresponding to the input sample.

[0107] The task of the orthogonal least squares method (OLS) is to select the appropriate regression operator vector and its number through learning so that the network output meets the requirements of the quadratic performance index. Its basic idea is: by orthogonalizing the regression operator vector, analyzing its contribution to reducing the residual, selecting the appropriate regression operator, and determining the number of regression operators based on the performance index. Calculation process method:

[0108] Perform an orthogonal-triangular decomposition of P

[0109] P=UA

[0110] A is an M×M upper triangular matrix with all diagonal elements set to 1; U is an N×M matrix with columns u i Orthogonal, that is, U T U=H, H is the diagonal element h i The diagonal matrix

[0111]

[0112] Therefore, d=UAW+e=Ug, g=AW

[0113] The orthogonal least squares solution of the formula g = AW is

[0114] g=H -1 U T d or

[0115] In the formula For vector The weight, and The trigonometric equations should be satisfied

[0116] Here, the orthogonalization adopts the Gram-Schmidt method, which calculates one column of A each time and performs the following orthogonalization

[0117]

[0118] (1≤i≤k; k=2,...,M)

[0119] Since Ag and e are independent of each other, the energy of the output response can be expressed as

[0120]

[0121] Divide both sides of the above equation by d T d, get

[0122]

[0123] in Defined as the error compression ratio, and q = e T e / d T d is the relative quadratic error.

[0124] In summary, if Figure 4 As shown, the RBF network adopts the orthogonal minimum multiplication learning algorithm steps:

[0125] S300. Initialization program;

[0126] S301. Select a hidden layer unit number M;

[0127] S302. Select a set of RBF center vectors;

[0128] S303. Calculate the regression matrix P based on the center vector and the input sample vector;

[0129] S304. Orthogonalize the columns of the regression matrix;

[0130] S305. Calculate separately and ε i ;

[0131] S306. Orthogonal-triangular decomposition of P is performed to obtain an upper triangular matrix A, and the connection weight vector W is solved by the triangular equation g = AW, where g = [g 1 , g 2 , ..., g M ] T ;

[0132] S307. Check whether it is satisfied Wherein 0<ρ<1 is the selected tolerance, if it is satisfied, the calculation is stopped, otherwise jump to S302 to continue execution.

[0133] The RBF network structure and parameter model are obtained through the above S3 step, including the number of hidden layer centers, center positions, weights, etc. of the model. When running online, new data will be used as the data input of this part after PLS dimension reduction. The estimated value of electrode insertion depth can be output through calculation of this model.

[0134] In S5, a large amount of field data is integrated in the advanced control and optimization system based on soft measurement. The failure of any data may lead to the degradation of the overall performance of the system or even complete failure. Since the actual measurement data will inevitably contain errors and are incomplete, it is indispensable to correct the measurement data. Due to the time-varying, nonlinear and incompleteness of the soft measurement object, the online correction of the model must be considered to adapt to the new working conditions. Therefore, we developed the online learning correction function of the soft measurement instrument to ensure its long-term stable and reliable operation. The specific approach is to independently add a correction module to the soft measurement instrument to monitor the output of the soft measurement instrument, the actual electrode insertion depth measurement value and the production operation conditions. When there is a large difference between the output (estimated value) of the soft measurement instrument and the actual electrode insertion depth measurement value, the difference and the production conditions in the corresponding section are recorded; when this difference occurs several times in a row and the corresponding production conditions are always in a stable operation state, an adjustment amount is given by calculating the deviation, which is superimposed on the output layer node of the soft measurement instrument, and the samples when the deviation occurs are recorded at the same time, so that new samples can be added for retraining the network after accumulating a certain number. Since the RBF neural network model training speed is relatively fast, online model learning can be achieved. That is, a batch of data under new working conditions is collected regularly and the model is retrained. Then the model's hidden layer center book, center position, weight and other parameters are completely generated based on the latest data. Each training removes the oldest data, absorbs the latest data, and iterates repeatedly. This method can make the model closely track the production conditions and effectively solve the problem of production condition drift.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A soft measurement method for electrode insertion depth of submerged arc furnace based on radial basis function neural network. It is characterized in that The steps include: S1. Intercept a section of data before power outage for each measurement of the actual electrode insertion depth as historical data and online correction data, collect electrical parameters and operating parameters during the production and operation of the submerged arc furnace, and send the electrical parameters and operating parameters to the PLS unit; S2, the PLS unit performs dimensionality reduction processing on the electrical parameters and the operating parameters, and sends the low-dimensional data after dimensionality reduction to the RBFN unit; S3, the RBFN unit receives low-dimensional data and performs soft measurement calculation of electrode insertion depth based on the low-dimensional data and outputs it, obtains the RBF network structure and parameter model through calculation and stores it in the RBFN unit, the RBF network structure and parameter model include the number of hidden layer centers, center positions, and weights of the model; S4. When running online, new data will be used as data input for the RBF network structure and parameter model after PLS dimension reduction, and the estimated value of electrode insertion depth can be output after calculation.

2. According to the radial basis function neural network based soft measurement method of electrode insertion depth of submerged arc furnace in claim 1, It is characterized in that Also includes: S5. The sampling unit samples the actual value of the electrode insertion depth of the ore-fired arc furnace at different intervals, and the online correction unit compares the actual value of the electrode insertion depth with the estimated value of the electrode insertion depth to obtain a comparison difference; when the comparison difference exceeds the specified difference range, the comparison difference and the production conditions in the corresponding section are recorded; When the comparison difference occurs more than the rated number of times and the corresponding production condition is always in a stable operating state, the adjustment amount is obtained according to the comparison difference, and the adjustment amount is superimposed on the output layer node of the soft measurement instrument. At the same time, the samples when the deviation occurs are recorded so that new samples can be added for retraining of the network after accumulating a certain number.

3. The soft measurement method for electrode insertion depth of a submerged arc furnace based on radial basis function neural network according to claim 1, It is characterized in that S2 specifically includes the following steps: S200, initialization program, selecting electrode voltage, electrode current, electrode power, electrode bearing chamber pressure of electrode lifting hydraulic cylinder, furnace bottom temperature, flue gas temperature, oxygen content in flue gas, hydrogen in flue gas as electrical parameters and working condition parameters as input variable X, selecting the actually measured three-phase electrode insertion depth as network output variable Y, and performing column-by-column standardization / normalization processing on X and Y; S201, let E = X, F = Y, u equal to a column of Y, or take the column vector with the largest variance; S202, calculate input weight vector w=E T u / (u T u); S203, normalizing the input weight vector w=w / ||w||; S204, calculating the characteristic vector t=Ew; S205, calculate output load vector q=F T t / (t T t); S206, normalizing the output load vector q=q / ||q||; S207, calculate output feature vector u=Fq T / (q T q); S208, repeat steps S202 to S207 until convergence; S209, calculate the internal model regression coefficient b=u T t / (t T t); S210, calculate X load vector p=bX T t / (bt T bt); S211, calculate the residual matrix E = E-btp T and F = F-btq T ; S212, return to S202 and execute in a loop until all eigenvectors are calculated; S213. Sort the eigenvectors by eigenvalue from large to small, select the eigenvectors corresponding to the first k eigenvalues ​​with the largest contribution to the overall feature, use these k eigenvectors as row vectors to form a dimensionality reduction matrix, multiply the dimensionality reduction matrix by the original matrix on the left to complete the dimensionality reduction, and the obtained data is the data after dimensionality reduction in step S2.

4. The soft measurement method for electrode insertion depth of an electric arc furnace based on radial basis function neural network according to claim 3, It is characterized in that In S208: When the difference between the eigenvector and the previous one within 10 times is within the allowed range, it is judged to be converged.

5. The soft measurement method for electrode insertion depth of a submerged arc furnace based on radial basis function neural network according to claim 1, It is characterized in that In S3, the RBFN unit uses the RBF network. The RBF network is a two-layer forward network. The number of inputs is equal to the number of independent variables studied. The middle layer selects the basis function as the transfer function. The transformation from the input layer to the hidden layer space is nonlinear, and the transformation from the hidden layer to the output layer is linear. The transformation function of the hidden layer unit is a locally distributed non-negative nonlinear function with radial symmetric decay about the center, that is, the radial basis function. The output layer is a linear combiner.

6. The soft measurement method for electrode insertion depth of a submerged arc furnace based on radial basis function neural network according to claim 5, It is characterized in that S3 specifically includes the following steps: S300, initialization program; S301, select a hidden layer unit number M; S302, select a group of RBF center vectors; S303, according to the central vector, using the data output by the dimension reduction in step S2 as the input sample vector, calculate the regression matrix P; S304, orthogonalizing each column of the regression matrix; S305, calculate separately and ε i ; S306, orthogonally decompose P to obtain an upper triangular matrix A, and solve the connection weight vector W by the triangular equation g=AW, where g=[g 1 , g 2 , ..., g M ] T ; S307, check whether it is satisfied Wherein 0<ρ<1 is the selected tolerance. If it is satisfied, the calculation is stopped; otherwise, the process jumps to S302 to continue the execution.

7. A storage medium, It is characterized in that The storage medium includes a stored program, wherein when the program is run, the soft measurement method for the electrode insertion depth of an electric arc furnace based on a radial basis function neural network as described in any one of claims 1 to 6 is executed.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, It is characterized in that The processor executes the soft measurement method for electrode insertion depth of an electric arc furnace based on a radial basis function neural network as described in any one of claims 1 to 6 through the computer program.