Industrial soft measurement method based on quadratic inequality constraint robust random configuration network

By introducing quadratic inequality constraints and singular value decomposition in the random configuration network, the problems of instability in output weight calculation and the impact of sample outliers are solved, and robustness and soft measurement accuracy are improved, and better industrial soft measurement effects are achieved. .

CN119069019APending Publication Date: 2024-12-03CHINA UNIV OF MINING & TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410631725.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-05-21
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing random configuration networks have problems such as instability in output weight calculations and high impact on sample outliers in industrial soft measurements, making it difficult to achieve high robustness and accuracy.

Method used

Propose a robust random configuration network (QI-RSCN) based on quadratic inequality constraints, and reduce the exception sample by converting the output weight optimization problem into the least squares problem under quadratic inequality constraints. The influence on the training effect is introduced and weight matrix weighting is introduced in the output weight calculation, and the least squares problem is transformed into the weighted least squares problem.

Benefits of technology

The robustness and soft measurement accuracy of the random configuration network are improved, the resistance to outliers and pathological matrix effects is enhanced, and the stability and generalization performance of learning results are significantly improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
  • Figure FT_3
    Figure FT_3
Patent Text Reader

Abstract

The invention relates to an industrial soft measurement method based on a quadratic inequality constraint robust random configuration network, which converts the optimization problem of the output weight of the random configuration network into the least square problem under the quadratic inequality constraint condition, proposes the quadratic inequality constraint random configuration network, and enables the output weight to be more stable. Furthermore, a robust random configuration network of quadratic inequality constraints is provided for suppressing the influence of abnormal values in training samples on output weights, and the process is as follows: firstly, a Laplacian degree matrix of the samples is utilized to construct a weight matrix of the samples, possible abnormal values are identified, and then the abnormal values in the samples are extracted; an SCN output weight optimization problem is converted into a weighted least square problem under quadratic inequality constraint, the proposed method is applied to industrial process soft measurement, and an experimental result shows that when an abnormal value exists in a sample, a good soft measurement effect is achieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to an industrial system operation closed-loop control method, and in particular to an industrial soft measurement method based on a quadratic inequality constrained robust random configuration network. Background Art

[0002] Real-time and accurate measurement of industrial system operation indicators is a prerequisite for achieving industrial system operation optimization and closed-loop control. However, in practice, many industrial operation indicators are difficult to detect online in real time, such as sewage treatment plant discharge water quality[1], cement clinker composition[2], waste incinerator exhaust gas composition[3], grinding particle size[4], etc. For industrial operation indicators that are difficult to detect online in real time, researchers have proposed detection methods using soft measurements[5]. Soft measurement is a method that uses computer technology combined with production process knowledge (or historical data) to measure important variables that are difficult to measure directly or temporarily unmeasurable. The research on soft measurement modeling mainly includes two aspects: mechanism-based modeling methods and data-driven modeling methods[6]. [N] represents the Nth reference, the same below.

[0003] The mechanism modeling method requires an in-depth understanding of the internal mechanism of the modeling object, accurate detection of the parameters of the modeling object, and certain modeling assumptions. In practice, the working mechanism of complex industrial systems is difficult to accurately describe, the internal parameters are difficult to accurately obtain, and there is a certain deviation between the actual working conditions and the theoretical assumptions, which leads to a certain deviation in the soft measurement based on the mechanism modeling method, and sometimes it is difficult to implement the literature [7]. The data-driven soft measurement modeling method does not need to understand the internal mechanism of the modeling object, but only requires input and output data. Therefore, this method has obvious advantages for the soft measurement problem of complex industrial operation indicators [8]. With the development of neural networks, data-driven soft measurement modeling methods using neural networks are becoming more and more popular [9].

[0004] Random weight neural network (RWNN) has fast learning speed and strong learning ability, and is a powerful soft sensor modeling tool [10,11,12]. The basic idea of ​​RWNN is: first determine the number of hidden nodes in the network, then randomly generate input weights and biases within a certain interval, and then use analytical methods to determine the output weights. A typical representative of the RWNN model is the Random Vector Function-link Network (RVFLN)

[13] . At present, there are many derivative models for RVFLN, such as variational mode decomposition RVFLN

[14] , robust multi-core RVFLN

[15] , incremental RVFLN

[16] , sparse pre-training RVFLN

[17] , etc. Since the random parameter interval of RVFLN is fixed and the network structure is fixed, it is necessary to adjust the parameters and network structure multiple times to obtain the optimal learning effect. To address this problem, the literature

[18] proposed a stochastic configuration network (SCN), which realizes the adaptive selection of network parameters and incremental learning of network structure under constraints.

[0005] Researchers have conducted in-depth research on the SCN model and proposed multiple improvement measures. Reference

[19] proposed a block incremental random configuration network (BSCN), which adds a set of hidden nodes to the network each time. BSCN speeds up the learning speed of SCN

[20] . Reference

[21] proposed a hybrid learning SCN, which uses a data and structure parallel learning method to effectively deal with the learning problem of big data. In response to the problem of ill-conditioned matrices that may be encountered when calculating the output weights of SCN, researchers have proposed a variety of solutions, such as Truncated singular-value-decomposition, TSVD-SCN

[22] , L 2 Regularization method

[23] , Baysian method, etc.

[24] . The above methods solve the problem of stable calculation of output weights, but do not consider the impact of sample outliers on the model. In response to the problem of outlier interference in learning samples, Wang and Li proposed the maximum related entropy robust random configuration network (Robust SCN, RSCN) and kernel density robust random configuration network (KDEN-SCN) in references

[25] and

[26] respectively. In addition, reference

[27] proposed a weighted RSCN and applied the proposed model to the soft measurement of operating indicators of hematite grinding process. Reference

[28] proposed a robust SCN based on M-Estimator. The above methods can suppress the impact of abnormal samples on learning effects, but the above robust models do not consider the stability of output weight calculation.

[0006] At present, SCN has a good effect in soft measurement. Reference

[29] introduced the L1 regularization term into SCN, and then used the ADMM algorithm to solve the output weight, and applied the improved model to the soft measurement of underground supply air volume. Reference

[30] proposed a soft measurement method for total phosphorus concentration in sewage treatment effluent based on fuzzy random configuration network. Reference

[31] proposed a soft measurement method based on improved manifold regularized random configuration network. Reference

[32] used SCN to perform soft measurement of NOx in solid waste incineration emission gas and achieved good results. Recently, reference

[33] reviewed the development history, advantages and disadvantages, future research directions, and application prospects of SCN. Summary of the invention

[0007] The purpose of the present invention is to propose an industrial soft measurement method based on quadratic inequality constraints SCN with robustness, QI-RSCN, which reduces the influence of sample outliers and ill-conditioned matrices on learning results, enhances the robustness of SCN, and improves the soft measurement accuracy of SCN.

[0008] The present invention adopts the following technical solutions:

[0009] An industrial soft sensing method based on a quadratic inequality constraints SCN with robustness (QI-RSCN) comprises the following steps:

[0010] (1) The optimization problem of SCN output weights is transformed into a least squares problem under quadratic inequality constraints (LSQI). On this basis, a quadratic inequality constrained random configuration network (Quadratic Inequality Constraints SCN, QI-SCN) is proposed, which makes the calculation of SCN output weights more stable and improves the robustness of SCN. At the same time, singular value decomposition is used to simplify the LSQI solution process, making it easy to implement.

[0011] (2) In order to reduce the impact of abnormal samples in training samples on the training effect of SCN, a quadratic inequality constrained robust random configuration network (QI-RSCN) is proposed:

[0012] First, the Laplace matrix of the sample is constructed, and then the weight matrix of the sample is constructed using the degree matrix. In the process of optimizing the output weight, the weight matrix gives smaller weights to abnormal samples. Then, the SCN output weight optimization problem is transformed into a weighted least squares problem (Weighted LSQI, WLSQI) under quadratic inequality constraints.

[0013] Function approximation simulation experiments show that QI-SCN and QI-RSCN have good robustness when the training samples contain different proportions of outliers. The proposed method is applied to the soft measurement of particle size in the hematite grinding process, and the experimental results show that the proposed method has good soft measurement effect.

[0014] On the basis of the above technical solutions, there are the following further technical solutions:

[0015] 1. Randomly configure the network

[0016] Use the function approximation process to briefly describe the learning process of SCN. Given the objective function f: d →m corresponds to a set of training data sets {X,Y}, where X = {x 1 ,x 2 ,L,x N} represents the N input data of f, Y={y 1 ,y 2 ,L,y N} represents the corresponding output data of f, y i =[y i,1 ,y i,2 ,L,y i,m ] T ∈? m (i=1,2,L,N), assuming that the SCN with L-1 hidden nodes is as shown in formula (1):

[0017]

[0018] in represents the activation function g of the jth hidden node j The output of f L-1 (X) represents the output of the SCN with L-1 hidden nodes; parameter (j=1, 2, L, L-1) are the input weight, bias and output weight of the jth hidden node respectively. The learning process of approximating the function f with SCN is as follows:

[0019] Step 1: Calculate f L-1 The output residual e of (X) L-1 , since the true function f is unknown, the sample set Y is used to replace the actual function value, f L-1The residual is shown in formula (2):

[0020] e L-1 =f L-1 (X)-Y=[e L-1,1 ,e L-1,2 Le L-1,m ]∈? N×m (2)

[0021] in, represents the i-th dimension output residual of SCN (i=1,2,L,m);

[0022] Step 2: Towards f L-1 (X) adds the Lth hidden node g L , its input weight and bias should satisfy the constraints of formula (3):

[0023]

[0024] where r∈(0,1) is a small real number, {μ L} is a non-negative real number sequence that satisfies μ L <1-r, h L Represents the hidden node g L The output is shown in formula (4):

[0025]

[0026] According to formula (3), T max candidate nodes form a candidate pool, and then select The largest g L As the Lth hidden node;

[0027] Step 3: Calculate the output weight of SCN according to formula (5).

[0028]

[0029] Among them, H L =[h 1 ,h 2 ,L,h L ]∈? N×L , the result of the above calculation is formula (6):

[0030]

[0031] Indicates H L The Moore-Penrose generalized inverse matrix of .

[0032] Step 4: Calculate f according to formula (7)L (X) output.

[0033]

[0034] Repeat steps 1 to 4 above. If the residual || e L || is smaller than the preset value ε of the residual, or the number of hidden nodes in SCN is greater than the maximum value L of hidden nodes max Then stop learning.

[0035] 2. Randomly Assigned Networks with Quadratic Inequality Constraints

[0036] From formula (6), we can see that when calculating the output weight β, if H L If it is an ill-conditioned matrix, the output weight β is very unstable, which makes the generalization performance of SCN worse. The present invention directly constrains the output weight β and proposes a quadratic inequality constrained random configuration network (QI-SCN). The selection of random parameters in QI-SCN is the same as step 2 in the previous section, and certain constraints are imposed on its norm when solving the output weight, converting the unconstrained least squares problem into the least squares problem LSQI problem. After adding the Lth hidden node in QI-SCN, the calculation of the output weight is as shown in formula (8):

[0037]

[0038] Where α>0, the constraint ||β|| 2 ≤α ensures that the norm of the output weight β does not exceed the hypersphere with a radius of α, avoiding the instability of the calculation results caused by the ill-conditioned matrix, thereby improving the robustness of SCN.

[0039] Solving the quadratic inequality constraints

[0040] The present invention uses the singular value decomposition method to calculate the LSQI solution of formula (8). In order to simplify the subsequent analysis, it is assumed that is a vector, the hidden layer matrix rank(H L )=r,H L The singular value decomposition of is formula (9)

[0041]

[0042] where U=[u 1 ,u 2 L N ]∈? N×N ,V=[v 1 ,v 2 LV N ]∈? m×m , σi >0(i=1,2,L,r) is the singular value of the matrix, then min||H L The solution of β-Y|| is shown in formula (10):

[0043]

[0044] like Then formula (10) is the solution of formula (8), if Then the solution of equation (8) is in ||β|| 2 =α, the inequality constraint of equation (8) is transformed into an equality constraint, as shown in equation (11):

[0045] min||H L β-Y|| 2

[0046] st||β|| 2 =α 2 (11)

[0047] To solve equation (11), first construct the Lagrangian function as equation (12)

[0048]

[0049] The partial derivative of L(β,λ) with respect to β,λ is equal to 0, and we get formula (13):

[0050]

[0051] Where β(λ) is a function of λ, using H L The singular value decomposition of The solution is formula (14)

[0052]

[0053] Since β(λ) satisfies the condition in formula (13) Therefore, the following equation (15) for the parameter λ holds,

[0054]

[0055] For a given α>0, solving β(λ) in equation (13) only requires finding λ that satisfies equation (15), and then substituting λ into equation (14). To solve λ, the following function is introduced, as shown in equation (16):

[0056]

[0057] Solving equation (15) is equivalent to finding the solution of the equation φ(λ) = 0. Since the least squares solution is Does not satisfy the constraints of formula (8), that is Therefore, when λ=0, φ(0)>0; when λ=∝, from formula (16), we can know that φ(∝)=-α 2 <0. Since φ(λ) is a monotonically decreasing function, there is a unique solution for φ(λ)=0 on the interval (0,∝). Let the solution of φ(λ)=0 be λ + , then the LSQI solution of equation (11) is equation (17)

[0058]

[0059] Least squares problem LSQI solution method:

[0060] H L The singular value decomposition is Among them, σ i H L The singular values ​​of u i surface

[0061] Show The i-th column, v i express The i-th column of H L ),

[0062]

[0063] Then the LSQI solution of equation (8) is

[0064] Else

[0065] Finding Lambda + >0, so that the formula φ(λ + )=0, then the LSQI solution of equation (8) is

[0066] End

[0067] 2.1 Quadratic Inequality Constraint Error Analysis

[0068] Let ρ LSQI (λ)=||H L β-Y|| 2 The square of the error norm of expression (8), β LS represents the least squares solution of equation (8) under unconstrained conditions. When , as shown in formula (18).

[0069]

[0070] At this time, the error of equation (8) is equal to the least squares error ρ LS .

[0071] When , the solution of equation (8) is shown in equation (17), and the square of the output error norm is shown in equation (19):

[0072]

[0073] The above equation ρ LSQI (λ + ) is shown in Appendix 1. The above formula shows that the error ρ of the least squares solution under the quadratic inequality constraint is LSQI Greater than or equal to the error ρ of the least squares solution under unconstrained conditions LS When λ + =0, ρ LSQI =ρ LS (At this time, the constraint condition in equation (8) does not work); when λ + >0, ρ LSQI > LS (At this time, the solution of equation (8) falls on the boundary of the constraint condition).

[0074] 2.2 Robustness Analysis of Quadratic Inequality Constraints

[0075] Under the constraint condition, the constraint condition of formula (8) makes the output weight not exceed the hypersphere The boundary of , avoiding the appearance of extremely large or extremely small output weights, and realizing the "risk sharing" of output weights, so the above constraints increase the output weight β(λ + ) calculates stability, assuming that the matrix H L The interference is δ H ,

[0076] If H L is a non-ill-conditioned matrix, we get The error when the system is disturbed is:

[0077]

[0078]

[0079] If H L is an ill-conditioned matrix, when λ + satisfy When , the disturbance error is:

[0080]

[0081] From formula (19), we can know Therefore, formula (21) can be written as:

[0082]

[0083] When 0<σr <<1(such as σ r =10 -10 ), due to λ + The existence of κ(λ + ) is greater than the first term of κ(0), that is, Choosing the right lambda + ,but That is, κ(λ + ) is much smaller than the second term of κ(0). Therefore, H L When it is an ill-conditioned matrix, LSQI can improve the robust performance of the system.

[0084] 3. Robust Random Configuration Networks with Quadratic Inequality Constraints

[0085] The quadratic inequality constraint random configuration network QI-SCN improves the robustness of the SCN model, but it is difficult to suppress the influence of outliers in training samples on the training results. In order to reduce the influence of outliers in training samples on the training results and further enhance the robustness of QI-SCN, the present invention proposes a quadratic inequality constraint SCN with robustness (QI-RSCN). The basic idea is to first identify possible noise samples, assign a smaller weight to the sample, and then calculate the output weight under the inequality constraint. The above process transforms the outlier suppression problem into a weighted least squares problem WLSQI, and then the quadratic inequality constraint robust random configuration network QI-RSCN can be expressed as follows:

[0086]

[0087] The diagonal matrix P = diag(p 1 ,p 2 Lp N ) is the weight matrix, p i represents the weight of the i-th sample. Reference

[22] proves that min||P(H L β-Y)|| 2 The SCN under the framework has universal approximation capability. At this time, the constraint condition of SCN (3) becomes (25),

[0088]

[0089] in make The output weight β can be calculated according to the LSQI solution method in Section 3.1.

[0090] 3.1 Calculation method of weight matrix

[0091] For a Lipschitz function, similar inputs will have similar outputs. When the output is disturbed, the derivative at the interference point will change suddenly. Therefore, the above idea can be used to identify the interference sample. For a set of discrete samples, if the sample (x j ,y j ) is disturbed, then the sample {x j ,y j The “rate of change” (i.e., difference value) between {x} and its surrounding samples is large, so the “rate of change” can be used to j ,y j The basic principle is as follows: Assume that the sample {x 1 ,y 1}, {x 2 ,y 2}, {x 3 ,y 3} are samples that are close and different from each other. The degree of difference between the three samples can be expressed as If a 12 a 13 a 23 If a is small, the similarity of the three samples is large, and the possibility of outliers in the samples is small; if a 12 a 13 Larger, and a 23 is smaller, then {x 1 ,y 1} may be an outlier, in which case the sample {x 1 ,y 1} Smaller weights, assuming the training set is Using the above method, we can get the neighbor matrix A of the sample set as shown in formula (26):

[0092]

[0093] in

[0094]

[0095] In the above formula, a ij (i≠j) indicates the similarity between samples, and the element a on the diagonal ii =0(i=1,2,L,N),θ 1 is a small constant. A is a Laplace matrix, and the degree matrix D of A is calculated as D = diag(d 1 ,d 2 L N ), where the elements in the matrix D If sample i is an outlier, then d i Larger. Set θ 2 >0, the following matrix is ​​calculated from matrix D as shown in formula (28)

[0096]

[0097] in When the matrix Elements on the diagonal It can be considered that sample i is not an outlier; when Sample i may be an outlier. The larger i is, the more likely it is an outlier. The weight matrix P of the sample is calculated as shown in formula (29):

[0098] P = diag (p 1 ,p 2 Lp N ) (29)

[0099] in θ 3 >0 is the set parameter. When sample i is a normal sample value, its weight p i =1; when When , sample i may be an abnormal sample, and the weight of sample i is p i A smaller number.

[0100] 3.2 Algorithm of the Quadratic Inequality Constrained Robust Random Configuration Network (QI-RSCN)

[0101] Given a training set {X,Y}, where X = {x 1 ,x 2 L,x N}, x i =[x i,1 ,x i,2 L i,d ] T ∈? d ; Y = {y 1 ,y 2 Ly N},y i =[y i,1 ,y i,2 y i,m ] T ∈? m ; Given the maximum number of hidden nodes L max , expected accuracy ε, maximum number of random configurations of random parameters T max , quadratic inequality constraint α. Given a set of positive definite scalars Υ=[λ min :Δλ:λ max ], empty set W,Ω. The QI-RSCN algorithm proposed in the present invention is described as follows:

[0102] QI-RSCN Algorithm

[0103] Initialization phase: Let e 0 =[y 1 ,y 2 y N ] T , W,Ω is an empty set, 0<r<1, set parameter θ 1 ,θ 2 ,θ 3

[0104]

[0105] 4. The present invention also proposes a soft-sensing experimental method for industrial operation indicators

[0106] The present invention firstly performs numerical simulation on the proposed method through numerical simulation experiments. Then the method in this paper is applied to the soft measurement of particle size in the hematite grinding process to verify the soft measurement effect of the method proposed in the present invention in actual industry. The configuration of the experimental platform is: Core(TM)i3-10100 CPU@3.60GHz, operating system window10, and Python is used for programming test. The root mean square error (RMSE) is used as the evaluation index, and its calculation method is as shown in formula (30)

[0107]

[0108] where y Test,i is the test output of the model, y i is the actual value, and N is the number of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0109] Figure 1 is the training set and experimental results of function approximation;

[0110] Figure 2 It is the process flow chart of hematite grinding process;

[0111] Figure 3 These are the soft measurement results of different methods (when 10% of the samples are added with noise).

[0112] 4.1 Numerical simulation experiment

[0113] The method of the present invention is used to approximate the following function (x∈[0,1]) as shown in formula (31):

[0114]

[0115] Use rand(0,1) to generate 1000 random numbers uniformly distributed in the interval [0,1] as training samples, and then use rand(0,1) to generate 600 random numbers uniformly distributed in the interval [0,1] as test samples. Add noise to 10% of the training sample output data to form a training sample set containing outliers. The noise generation is shown in formula (32):

[0116]

[0117] where y i is the original sample, is the interfered sample, k(rand(0,1)-0.5) is the random noise, k=4 is the coefficient for controlling the noise size, and experiments are carried out using IRVFLN

[16] , SCN

[18] , RSCN

[23] , QI-SCN, QI-RSCN and other methods. max =100,T max =50,ε=0.001,1-r=[10 -1 ,10 -2 ,10 -3 ,10 -4 ,10 -5 ,10 -6 ,10 -7 ,10 -8 ],Υ=[λ min :Δλ:λ max ]=[0.5,1,5,10,30,50,100,150,200,250], and the other parameters are shown in Table 1.

[0118] Table 1 Parameter setting values

[0119]

[0120] When 10% of the samples in the training set contain noise, the training set Figure 1 As shown in (a), the experimental results of SCN and QI-RSCN are as follows Figure 1 (b) is shown. Figure 1 As shown in (b), the learning effect of QI-RSCN is significantly better than that of SCN.

[0121] Table 2 Comparison of test results under different noise interference (RMES)

[0122]

[0123] Using the parameters in Table 1, 50 independent experiments were conducted, and the average RMSE of the 50 results is shown in Table 2. It can be seen from Table 1 that the QI-SCN and QI-RSCN methods are better than SCN, RSCN, IRVFLN and other methods under different proportions of outliers. Since QI-RSCN suppresses outliers in input samples and constrains the output weights, compared with other methods, QI-RSCN can achieve the best effect under various proportions of outliers (as shown in bold in the table). The experimental results show that when there is no noise interference, there is no need to identify the noise. At this time, the performance of QI-SCN and QI-RSCN is basically the same, and the experimental results in this experiment are better than the SCN method.

[0124] 4.2 Soft measurement of particle size in hematite grinding process

[0125] Hematite grinding system Figure 2 As shown in the figure, the system mainly consists of raw ore bin, conveyor belt, ball mill and spiral classifier (at the same time, it cooperates with relevant instruments and actuators to complete detection and control). First, the ore in the raw ore bin is transported to the ball mill for grinding by the conveyor belt. In order to maintain the best working condition of the ball mill, an appropriate amount of water will be added according to the feed ratio during the grinding process. Then, the ground slurry is discharged from the outlet of the ball mill and sent to the spiral classifier. In this process, in order to adjust the concentration of the classifier slurry, a certain amount of water will be added so that large particles of ore can be precipitated. The blades of the spiral classifier will stir the slurry, so that the large particles of ore return to the ball mill for re-grinding. The fine particles of ore will flow out from the lower end of the classifier with the overflow slurry to form the final grinding product for subsequent use.

[0126] In the actual grinding process of hematite, we use the particle size (PS) as a standard to measure the quality of the grinding product. The grinding particle size refers to the proportion of particles with a diameter less than 0.074mm in the mineral after grinding in the ball mill. If the PS is too small, it may cause excessive crushing of useful minerals, muddling of ore pulp, difficulty in recovery, and increase the energy consumption of grinding. On the contrary, if the ore PS is too large, it is difficult to achieve the monomer dissociation of the mineral, making subsequent sorting difficult. The ideal sorting operation requires fewer coarse and fine mineral particles, and as many easy-to-select particles in the middle as possible.

[0127] According to the characteristics of the grinding process, the main factors affecting PS include the characteristics of the new ore and the operating parameters of the grinding equipment. The ore properties include the particle size of the new ore, B 1 Grindability B 2 The operating parameters include: ball mill feed rate a 1 , ball mill inlet water flow rate a 2, Spiral classifier overflow concentration a 3 Among them, the ore property B 1 With B 2 It is difficult to measure, but the ball mill working current c 1 The working current c of the spiral classifier 2 With ore properties B 1 With B 2 There is a certain relationship, which can reflect the properties of the ore to a certain extent. Therefore, PS can be expressed as the following function

[0128]

[0129] in Represents the functional relationship between PS and various physical quantities. The vector [x 1 ,x 2 ,x 3 ,x 4 ,x 5 ] T =[a 1 ,a 2 ,a 3 ,c 1 ,c 2 ] T express The input variables.

[0130] In this section, we use SCN

[16] , IRVFLN

[18] , RSCN

[23] , QI-SCN, QI-RSCN and other methods to perform soft measurement of mineral particle size in the hematite grinding process. In the experiment, the hematite grinding data comes from the semi-physical simulation platform

[31] . From the 1400 samples collected, 800 are selected as training samples and 600 as test samples. Normalized data processing is used to make the data input samples and output samples distributed between [0,1]. In the experiment, the maximum number of hidden nodes of IRVFLN is 50, and the L of the other models is max =25, T max =20, parameter ε,1-r,Υ=[λ min :Δλ:λ max The setting of ] is the same as that of Experiment 1, and the settings of other parameters are shown in Table 3. In the model learning stage, noise is added to 10% of the samples according to formula (32), while no noise is added to the test set. The estimation results of grinding particle size using different methods are shown in Figure 3 As shown. Figure 3 It can be seen that the estimation result of the QI-RSCN proposed in the present invention is most consistent with the actual value.

[0131] Table 3 Parameter settings of each model:

[0132]

[0133] 0%, 5%, 10%, 20%, and 30% of the training samples are selected to add noise (k=2) as shown in formula (32), while there is no noise on the test set. 50 independent experiments are carried out, and the experimental results are shown in Table 4. As can be seen from Table 4, when there is no noise in the training samples, SCN, RSCN, QI-SCN, and QI-RSCN all have good soft measurement effects. When the samples contain different proportions of outliers, the soft measurement effects of QI-SCN and QI-RSCN methods are better than other methods. Compared with QI-SCN, the QI-RSCN method suppresses outliers and constrains the output weights. Therefore, when there are outliers in the samples, QI-RSCN has better soft measurement effects than QI-SCN.

[0134] Table 4 Soft sensing results (RMSE) of different methods in hematite grinding process

[0135]

[0136]

[0137] 5 Conclusion

[0138] Aiming at the problems that SCN may encounter in industrial process soft measurement, such as unstable output weight calculation and the great influence of outliers in training samples on training results, this paper proposes a quadratic inequality robust random configuration network (QI-RSCN). First, the output weight of SCN is constrained by quadratic inequality to avoid the instability caused by ill-conditioned matrix. Secondly, the Laplace matrix is ​​constructed by differential method, and then the weights of different training samples are determined to reduce the influence of sample outliers on training results. Numerical simulation experiments and hematite grinding soft measurement experiments show that QI-RSCN has better soft measurement effect. The method proposed in this paper will be further studied from the following aspects: 1) The method in this paper involves many hyperparameters, and how to select appropriate hyperparameters needs further study; 2) When there are few samples, how to determine the outliers in the sample also needs further study. 3) The LSQI method is combined with the existing weighted SCN method [26,27,28], or combined with the outlier detection method to further study the robustness of SCN.

[0139] References:

[0140] [1]Wang G, Jia QS, Zhou MC, et al. Artificial neural networks for waterquality soft-sensing in wastewater treatment: a review [J]. Artificial Intelligence Review, 2021: 1-23. DOI: 10.1007 / s 10462-021-10038-8.

[0141] [2]Qiao Jinghui, Chai Tianyou. Data and model driven soft sensing model for cement raw material decomposition rate[J]. Acta Automatica Sinica, 2019, 45(08): 1564-1578. DOI: 10.16383 / j.aas.c180734

[0142] [3] Qiao Junfei, Guo Zihao, Tang Jian. A review of dioxin emission concentration detection methods for municipal solid waste incineration process[J]. Acta Automatica Sinica, 2020, 46(06): 1063-1089. DOI: 10.16383 / j.aas.c190005

[0143] [4]Dai W,Liu Q,Chai T.Particle size estimate of grinding processes using random vector functional link networks with improved robustness[J].Neurocomputing,2015,169 361-372.

[0144] [5]Asogwa OC,Oladugba A V.Of Students Academic Performance RatesUsing Artificial Neural Networks(ANNs)[J].Science and Education PublishingCo.Ltd[2024-01-10].DOI:10.12691 / ajams-3-4-3.

[0145] [6] Chen Long, Liu Quanli, Wang Linqing, Zhao Jun, Wang Wei. A review of data-based process industry production index prediction methods[J]. Acta Automatica Sinica, 2017, 43(6): 944-954.

[0146] [7] Xie SW,

[0147] [8]Wang B, Li ZC, Dai ZW, Lawrence N, Yan

[0148] [9] Chen Long, Han Zhongyang, Zhao Jun, Wang Wei. A review of data-driven integrated energy system operation optimization methods[J]. Control and Decision, 2021, 36(2): 283-294.

[0149] (Chen Long, Han Zhong-Yang, Zhao Jun, Wang Wei. Review of research of data-driven methods on operational optimization of integrated energy systems [J]. Control and Decision, 2021, 36(2): 283-294.)

[0150]

[10] Huang G, Zhu Q, Siew C K.Extreme learning machine: theory and applications[J].Neurocomputing, 2006, 70(1):489-501.

[0151]

[11] Huang G,Chen L,Siew C K,et al.,Universal approximation usingincmental constructive feedforward networks with random hidden nodes[J].IEEETransactions on Neural Networks,2006,17(4):879-892.

[0152]

[12] Cao W,Wang X,Ming Z,et al.A review on neural networks with randomweights[J].Neurocomputing,2018,275:278-287.

[13] Pao Y H,Park G H,Sobajic DJ.Learning and generalization characteristics of the random vectorfunctional-link net[J].Neurocomputing,1994,6(2):163-180.

[0153]

[14] Bisoi R,Dash P K,Mishra S P.Modes decomposition method in fusionwith robust random vector functionallink network for crude oil priceforecasting[J].AppliedSoft Computing,2019,80:475-493.

[0154]

[15] Sahani M,Dash P K.Deep convolutional stack autoencoder of processadaptive VMD data with robust multikernel RVFLN for power quality eventsrecognition[J].IEEE Transactions on Instrumentation and Measurement,2021,70:1-12.

[0155]

[16] Li M,Wang D.Insights into randomized algorithms for neuralnetworks:Practical issues and common pitfalls[J].Information Sciences,2017,382-383:170-178.

[0156]

[17] Zhang Y,Wu J,Cai Z,et al.An unsupervised parameter learning modelfor RVFL neural network[J].Neural Networks,2019,112:85-97.

[0157]

[18] Wang D,Li M.Stochastic configuration networks:Fundamentals andalgorithms[J].IEEE Transactions on Cybernetics,2017,47(10):3466-3479.

[0158]

[19] Dai W,Li DP,Zhou P,Chai TY.Stochastic configuration networks withblock increments for data modeling in process industries.InformationSciences,2019,484:367–386.[doi:10.1016 / j.ins.2019.01.062]

[0159]

[20] Tian Q,Yuan SJ,Qu HQ.Intrusion signal classification usingstochastic configuration network with variable increments of hiddennodes.Optical Engineering,2019,58(2):026105.[doi:10.1117 / 1.OE.58.2.026105]

[0160]

[21] Dai W, Zhou XY, Li DP, Zhu S, Wang

[0161]

[22] Sheng ZY, Zeng ZQ, Qu HQ, Zhang Y. Optical fiber intrusion signal recognition method based on TSVD-SCN. Optical Fiber Technology, 2019, 48: 270–277. [doi:10.1016 / j.yofte.2019.01.023]

[0162]

[23] Pan Chengyan, Xu Jinxue, Weng Yongpeng. A method for chemical process fault identification based on manifold regularized random configuration network. Chinese Journal of Scientific Instruments, 2021, 42(5): 219–226. [doi:10.19650 / j.cnki.cjsi.J2107541]

[0163]

[24] Wu RZ, Lv BY, Dai CM, Wang WG. Bayesian stochastic configuration networks for robust data modeling. Concurrency and Computation: Practice and Experience, 2022, 34(1):e6495. [doi:10.1002 / cpe.6495]

[0164]

[25] Wang DH, Li M. Robust stochastic configuration networks with kerneldensity estimation for uncertain data regression. Information Sciences, 2017, 412–413:210–222. [doi:10.1016 / j.ins.2017.05.047]

[0165]

[26] Li M, Huang CQ, Wang DH. Robust stochastic configuration networks with maximum correntropy criterion for uncertain data regression. InformationSciences, 2019, 473:73–86. [doi:10.1016 / j.ins.2018.09.026]

[0166]

[27] Dai W, Li DP, Chen QX, Chai TY. Data driven particle size estimation of hematite grinding process using stochastic configuration network with robust technique. Journal of Central South University, 2019, 26(1): 43–62. [doi:10.1007 / s11771-019-3981-2]

[0167]

[28] Ding Shifei, Zhang Chenglong, Guo Lili, et al. Weighted deep random configuration network based on M-estimator function [J]. Journal of Computer Science, 2023, 46(11): 2476-2487.

[29] Wang Qianjin, Dai Wei, Lu Qun, Fu Xiaorong, Ma Xiaoping. A sparse learning method for random configuration network soft sensor model. Control and Decision, 2022, 37(12): 3171–3182. [doi:10.13195 / j.kzyjc.2021.1047]

[0168]

[30] Qiao Junfei, Li Kang, Li Dapeng, et al. A soft sensing method for effluent total phosphorus concentration based on fuzzy random configuration network: 202310564458[P][2024-01-10].

[0169]

[31] Du Zhiwei, Zhang Fengnan, Yang Hai. Soft sensing based on improved manifold regularized random configuration network[J]. Control Engineering, 2023, 30(5): 872-880.

[0170]

[32] Wang RR, Li FY, Yan AJ, Modular stochastic configuration network-based prediction model for NOx emissions in municipal solid waste incineration process. Engineering Applications of Artificial Intelligence[J], 2024.

[0171]

[33] Zhang Chenglong, Ding Shifei, Guo Lili, et al. Research progress on random configuration networks[J / OL]. Journal of Software, 1-21[2024-01-10]https: / / doi.org / 10.13328 / j.

[0172] cnki.jos.006804.

Claims

1. An industrial soft-sensing method based on a quadratic inequality constrained robust random configuration network, characterized by The following steps are involved: (1) The optimization problem of the output weight of the random configuration network (SCN) is transformed into a least squares problem under quadratic inequality constraints (LSQI). On this basis, a random configuration network with quadratic inequality constraints (Quadratic Inequality Constraints SCN), namely QI-SCN, is proposed. This makes the calculation of the SCN output weight more stable and improves the robustness of SCN. At the same time, singular value decomposition is used to simplify the LSQI solution process, making it easy to implement. (2) In order to reduce the impact of abnormal samples in training samples on the training effect of SCN, a quadratic inequality constrained robust random configuration network QI-RSCN is proposed: First, the Laplace matrix of the sample is constructed, and then the weight matrix of the sample is constructed using the degree matrix. In the process of optimizing the output weight, the weight matrix gives smaller weights to abnormal samples. Then, the SCN output weight optimization problem is transformed into a weighted least squares problem under quadratic inequality constraints, Weighted LSQI, WLSQI.

2. The industrial soft-sensing method based on quadratic inequality constrained robust random configuration network according to claim 1 is characterized in that The quadratic inequality constrained robust random configuration network QI-RSCN comprises: (1) Randomly configure the network Use the function approximation process to briefly describe the learning process of SCN. Given the objective function f: d → m A set of training data sets corresponding to {X,Y}, where X = {x1,x2,L,x N } represents the N input data of f, Y={y1,y2,L,y N } represents the corresponding output data of f, y i =[y i,1 ,y i,2 ,L,y i,m ] T ∈? m (i=1,2,L,N), assuming that the SCN with L-1 hidden nodes is as shown in formula (1): in represents the activation function g of the jth hidden node j The output, f L-1 (X) represents the output of the SCN with L-1 hidden nodes; parameter w j ,b j , (j=1, 2, L, L-1) are the input weight, bias and output weight of the jth hidden node respectively. The learning process of approximating the function f with SCN is as follows: Step 1: Calculate f L-1 The output residual e of (X) L-1 , since the true function f is unknown, the sample set Y is used to replace the actual function value, f L-1 The residual is shown in formula (2): And L-1 =f L-1 (X)-Y=[and L-1,1 ,And L-1,2 The e L-1,m ]∈? N×m (2) in, represents the i-th dimension output residual of SCN (i=1,2,L,m); Step 2: Towards f L-1 (X) adds the Lth hidden node g L , its input weight and bias should satisfy the constraints of formula (3): where r∈(0,1) is a small real number, {μ L } is a non-negative real number sequence that satisfies μ L <1-r, h L Represents the hidden node g L The output is as follows According to formula (3), T max candidate nodes form a candidate pool, and then select The largest g L As the Lth hidden node; Step 3: Calculate the output weight of SCN according to formula (5). Among them, H L =[h1,h2,L,h L ]∈? N×L , the result of the above calculation is formula (6): Indicates H L The Moore-Penrose generalized inverse matrix of . Step 4: Calculate f according to formula (7) L (X) output. Repeat steps 1 to 4 above. If the residual || e L || is smaller than the preset value ε of the residual, or the number of hidden nodes in SCN is greater than the maximum value L of hidden nodes max Then stop learning; (2) Random Configuration Network with Quadratic Inequality Constraints From formula (6), we can see that when calculating the output weight β, if H L If it is an ill-conditioned matrix, the output weight β is very unstable, which makes the generalization performance of SCN worse. The present invention directly constrains the output weight β and proposes a random configuration network QI-SCN with quadratic inequality constraints. The selection of random parameters in QI-SCN is the same as the second step in the random configuration network in the previous section. When solving the output weight, a certain constraint is imposed on its norm, and the unconstrained least squares problem is converted into a least squares problem LSQI problem. After adding the Lth hidden node in QI-SCN, the calculation of the output weight is as shown in formula (8): Where α>0, the constraint condition ||β||2≤α ensures that the norm of the output weight β does not exceed the hypersphere with a radius of α, avoiding the instability of the calculation results caused by the ill-conditioned matrix, thereby improving the robustness of SCN; Solving quadratic inequality constraints The present invention uses the singular value decomposition method to calculate the LSQI solution of formula (8). In order to simplify the subsequent analysis, it is assumed that is a vector, the hidden layer matrix rank(H L )=r,H L The singular value decomposition of is formula (9) Where U=[u1,u2L u N ]∈? N×N ,V=[v1,v2L v N ]∈? m×m , σ i >0(i=1,2,L,r) is the singular value of the matrix, then min||H L The solution of β-Y|| is shown in formula (10): like Then formula (10) is the solution of formula (8), if Then the solution of equation (8) is on the boundary of ||β||2=α. At this time, the inequality constraint of equation (8) is converted into an equality constraint, as shown in equation (11): min||H L β-Y|| 2 st||β|| 2 =a 2 (11) To solve equation (11), first construct the Lagrangian function as equation (12) The partial derivative of L(β,λ) with respect to β,λ is equal to 0, and we get formula (13): Where β(λ) is a function of λ, using H L The singular value decomposition of The solution is formula (14) Since β(λ) satisfies the condition in formula (13) Therefore, the following equation (15) holds for parameter λ: For a given α>0, solving β(λ) in equation (13) only requires finding λ that satisfies equation (15), and then substituting λ into equation (14). To solve λ, the following function is introduced, as shown in equation (16): Solving equation (15) is equivalent to finding the solution of the equation φ(λ) = 0. Since the least squares solution is Does not satisfy the constraints of formula (8), that is Therefore, when λ=0, φ(0)>0; when λ=∝, from formula (16), we can know that φ(∝)=-α 2 <0. Since φ(λ) is a monotonically decreasing function, there is a unique solution for φ(λ)=0 on the interval (0,∝). Let the solution of φ(λ)=0 be λ + , then the LSQI solution of equation (11) is equation (17) Least squares problem LSQI solution method: H L The singular value decomposition is Among them, σ i H L The singular values ​​of u i surface Show The i-th column, v i express The i-th column of L ), If Then the LSQI solution of equation (8) is Else Finding Lambda + >0, so that the formula φ(λ + )=0, then the LSQI solution of equation (8) is End; (3) Robust Random Configuration Networks with Quadratic Inequality Constraints First, possible noise samples are identified and assigned a smaller weight. Then, the output weight is calculated under the inequality constraint. The above process transforms the outlier suppression problem into a weighted least squares problem WLSQI. Then, the quadratic inequality constrained robust random configuration network QI-RSCN can be expressed as follows: The diagonal matrix P = diag(p1,p2L p N ) is the weight matrix, p i represents the weight of the i-th sample. At this time, the constraint condition of SCN (3) becomes (25), in make The output weight β can be calculated according to the LSQI solution method in Section 3.1; The calculation method of the weight matrix is: For a Lipschitz function, similar inputs will have similar outputs. When the output is disturbed, the derivative at the interference point will show a sudden change. Therefore, the above idea can be used to identify the interference samples. For a set of discrete samples, if the sample (x j ,y j ) is disturbed, then the sample {x j ,y j The "rate of change" (i.e., difference value) between {x} and its surrounding samples is large, so the "rate of change" can be used to j ,y j The basic principle is as follows: Assuming that samples {x1, y1}, {x2, y2}, and {x3, y3} are close and different samples, the difference between the three samples can be expressed as If a 12 a 13 a 23 If a is small, the similarity of the three samples is large, and the possibility of outliers in the samples is small; if a 12 a 13 Larger, and a 23 If {x1, y1} is smaller, then {x1, y1} may be an outlier. In this case, a smaller weight is given to the sample {x1, y1}. Assuming that the training set is Using the above method, we can get the neighbor matrix A of the sample set as shown in formula (26): in In the above formula, a ij (i≠j) indicates the similarity between samples, and the element a on the diagonal ii =0(i=1,2,L,N), θ1 is a small constant, A is the Laplace matrix, calculate the degree matrix of A D=diag(d1,d2L d N ), where the elements in the matrix D If sample i is an outlier, then d i Larger. Set θ2>0, and calculate the following matrix from matrix D as shown in formula (28) in When the matrix Elements on the diagonal It can be considered that sample i is not an outlier; when Sample i may be an outlier. The larger i is, the more likely it is an outlier. The weight matrix P of the sample is calculated as shown in formula (29): P=diag(p1,p2L p N ) (29) in θ3>0 is the set parameter. When sample i is a normal sample value, its weight p i =1; when When , sample i may be an abnormal sample, and the weight of sample i is p i is a smaller number; The algorithm of the quadratic inequality constrained robust random configuration network QI-RSCN is: Given a training set {X,Y}, where X = {x1,x2L,x N }, x i =[x i,1 ,x i,2 L i,d ] T ∈? d ; Y={y1,y2L y N },y i =[y i,1 ,y i,2 y i,m ] T ∈? m ; Given the maximum number of hidden nodes L max , expected accuracy ε, maximum number of random configurations of random parameters T max , quadratic inequality constraint α, given a set of positive definite scalars Υ = [λ min :Δλ:λ max ], empty set W,Ω, the proposed QI-RSCN algorithm is described as follows: Initialization stage: let e0=[y1,y2L y N ] T , W,Ω is an empty set, 0<r<1, set parameters θ1,θ2,θ3 1. Calculate the weight matrix P according to formula (29) in Section 4.1 2.While L<L max and||e L || F >εDo Phase 1: Random parameter configuration 3.For λ∈Υ Do 4.For k=1,2,L,T max Do 5. From [-λ,λ] d and a random value in [-λ,λ] as the input weight w L With the offset b L ; 6. Calculation of ξ L,q With h L , set μ L =(1-r) / (L+1) 7.If min{ξ L,1 ,x L,2 ,L,ξ L,m }≥0 8. Save w L With b L to W; Save to Ω 9. Else return to step 4 10.End if 11.End for 12.If W is not empty 13. Find the largest ξ in Ω L Corresponding and Structure H L =[H L ,h L ], go to step (17); 14.Else 15. Randomly select τ∈(0,1-r), update r=r+τ, and return to step (4); 16.End if 17.End For (End step (3)) Phase 2: Calculate output weights 18. Solve β according to the LSQI calculation method 19.End While 20.Returnβ, 3. The industrial soft-sensing method based on quadratic inequality constrained robust random configuration network according to claim 1 or 2, characterized in that The industrial operation soft sensing method is: Firstly, the proposed method is numerically simulated by numerical simulation experiments to verify the soft measurement effect of the proposed method in actual industry. The configuration of the experimental platform is: Core(TM)i3-10100 CPU@3.60GHz, operating system window10, Python is used for programming test, and the root mean square error (RMSE) is used as the evaluation index. Its calculation method is as shown in formula (30): where y Test,i is the test output of the model, y i is the actual value, and N is the number of data.

Citation Information

Cited By

  • Method and system for predicting cognitive competence of old people based on voice task

    CN119314658A