Powder particle size PDF shape modeling method based on iterative learning and subspace identification

Through iterative learning, selecting Gaussky functions and subspace identification to establish a linear dynamic model, the problem of difficulty in selecting parameters and complex models in existing powder particle size PDF shape modeling is solved, and more efficient powder particle size PDF shape modeling and tracking is achieved. control.

CN115356931BActive Publication Date: 2024-11-19NORTHEASTERN UNIV CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202211016390.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-24
Publication Date
2024-11-19
Estimated Expiration
2042-08-24

AI Technical Summary

Technical Problem

The existing powder particle size PDF shape modeling method relies on manual experience to select Gaussky functions, resulting in difficulty in selecting parameters, poor modeling effect, and complex neural network model structure, making it difficult to apply in practice.

Method used

The Gaussian radial basis function neural network based on iterative learning is used to select the optimal basis function, and combine subspace identification to establish a linear dynamic model with weights, replacing the nonlinear dynamic model, simplifying the model structure and improving modeling accuracy.

Benefits of technology

Through iterative learning, selecting the optimal Gaussky function can significantly improve the modeling effect of powder particle size PDF shape, reduce model complexity, and improve modeling accuracy, which is suitable for practical application needs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115356931B_ABST
    Figure CN115356931B_ABST
Patent Text Reader

Abstract

The present invention provides a powder particle size PDF shape modeling method based on iterative learning and subspace identification, and relates to the field of powder particle size control technology. The steps are as follows: 1) collecting input and output data of a disc grinding system, including disc gap, feed amount and powder particle size PDF shape data; 2) using RBF-NN to perform decoupling calculation on the weights corresponding to the powder particle size PDF shape; 3) using iterative learning to select a set of optimal Gaussian basis functions; 4) using subspace identification to establish a linear dynamic model of weights; 5) combining the weight linear dynamic model and the RBF-NN approximation of the powder particle size PDF shape, a dynamic model of the powder particle size PDF shape can be obtained. The present invention uses iterative learning to select the Gaussian basis function, so that the approximation error of the powder particle size PDF shape is minimized, which greatly improves the modeling effect; using subspace identification to establish a weight linear dynamic model, while ensuring good modeling accuracy, the model complexity is reduced, which is particularly in line with the requirements for the model to be simple and easy to use in practical applications.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of powder particle size control, and in particular to a powder particle size PDF shape modeling method based on iterative learning and subspace identification. Background Art

[0002] In actual industrial processes, the PDF shape control problem of variables has always been extremely challenging. For example, the shape control of the fiber length distribution of paper in the papermaking process and the shape control of the flame temperature field distribution in the combustion process can be regarded as typical random system distribution control problems. Specifically, taking the grain production process in the food processing industry as an example, the particle size of the processed grain powder needs to meet certain requirements, otherwise it will affect the subsequent processes; therefore, the precise control of the powder particle size can not only provide higher quality raw materials for subsequent food processing processes, but also reduce the energy consumption of the production process. However, in the actual production process, the conventional practice is to use the mean and variance of the powder particle size as the basis for judging the quality of the powder. However, for an actual random system, the mean and variance cannot accurately reflect the full random distribution characteristics of the output variable. Therefore, it is very important to use the powder particle size probability density function (PDF) shape instead of the traditional mean and variance as an indicator of powder quality, which is bound to be more widely used in the actual production process.

[0003] In the process of using the disc grinding system to produce powder, multiple variables such as feed amount, grinding disc pressure, grinding disc speed, grinding disc gap, etc. will affect the powder particle size PDF shape. Considering the measurability of the variables and the degree of influence on the actual powder particle size PDF, the grinding disc gap and feed amount can be determined as the control input of the system to change the action of the disc grinding system. Therefore, establishing a random distribution model of the disc grinding system with grinding disc gap and feed amount as control input and powder particle size PDF shape as output is of great significance to improving powder quality and reducing production energy consumption.

[0004] Patent publication number CN109695174A discloses "PDF shape prediction method and system for fiber length distribution in pulping process". This method focuses on the prediction of PDF shape of fiber length distribution in papermaking pulping process, uses RBF neural network to approximate PDF data of fiber length distribution, and then uses random weight neural network to establish a nonlinear dynamic model of weights, thereby obtaining the PDF shape prediction model of fiber length distribution.

[0005] Patent publication number CN108846178A discloses "a method and system for estimating the powder particle size distribution shape of a disc grinding system". This method uses an RBF neural network to decouple the weights, and then uses a BP neural network to construct a nonlinear dynamic model of the weights, thereby obtaining an estimated value of the weight vector, and on this basis, estimates the probability density function of the powder particle size distribution shape.

[0006] However, in the PDF shape modeling method proposed in the above patent, Gaussian radial basis function neural network (RBF-NN) is generally used to approximate the PDF shape of the output random variable. Therefore, the effect of the approximation strongly depends on the center and width of the selected Gaussian basis function. If the basis function is not selected accurately, it will greatly affect the subsequent modeling accuracy. The existing methods often rely on manual experience for the selection of basis functions, and have the disadvantages of difficulty in parameter selection and poor modeling effect. On the other hand, most of the existing methods choose to use neural networks of various structures to establish nonlinear dynamic models of weights, which have problems such as complex model structure, high modeling data requirements, and difficulty in parameter selection. Considering the actual application effect of the model, a simpler and more efficient linear dynamic model of weights should be used instead. Summary of the invention

[0007] In view of the deficiencies in the prior art, the present invention provides a powder particle size PDF shape modeling method based on iterative learning and subspace identification.

[0008] A powder particle size PDF shape modeling method based on iterative learning and subspace identification specifically includes the following steps:

[0009] Step 1, collecting the control data of the grinding disc gap and the feeding amount during the operation of the grinding system, as well as the powder particle size PDF shape data;

[0010] Step 2: Using Gaussian radial basis function neural network (RBF-NN), the weight vector corresponding to the powder particle size PDF shape is decoupled and calculated, so that the powder particle size PDF is expressed as a product of a set of basis functions and corresponding weights;

[0011] Step 2.1, the grinding disc gap and feed amount are taken as the input variables of the system, and the powder particle size PDF shape is taken as the output variable of the system.

[0012] Step 2.2: Use a Gaussian radial basis function neural network (RBF-NN) to approximate the powder particle size PDF shape.

[0013] The approximation method is specifically:

[0014] For a dynamic system with non-Gaussian random distribution, the powder particle size is represented as a uniformly bounded random variable y∈[a,ζ], where a and ξ represent the value range of y;

[0015] Let u(k)∈R m R is the control variable input for controlling the powder particle size PDF shape. m represents an m-dimensional column vector;

[0016] The shape of the powder particle size PDF in the disc grinding process is represented by its probability density function γ(y,u(k)) at each sampling time k:

[0017]

[0018] Where P(a≤y<ξ,u(k)) represents the probability that the output powder particle size falls within the interval [a,ξ] when the control input u(k) is applied to the disc grinding system;

[0019] Assuming that the interval [a, b] is given in advance, where a and b represent the left and right endpoints of the random variable y, and the PDF is continuous and bounded, the powder particle size PDF shape is approximated by the following Gaussian radial basis function neural network RBF-NN:

[0020]

[0021] In the formula, w l (u(k)) is the basis function R controlled by the control input u(k) l (y), l represents the lth network node of the Gaussian radial basis function neural network RBF-NN, and n is the number of network nodes;

[0022] The basis function R defined in the interval [a,b] l (y) is as follows:

[0023]

[0024] In the formula, μ l and σ l Respectively represent the center and width of the lth basis function;

[0025] At this time, the powder particle size PDF shape is expressed as the product of the basis function and the corresponding weight

[0026]

[0027] In the formula, C(y) is the row vector composed of n basis functions, V(k) is the corresponding weight vector, and R n (y) and w n (u(k)) represents the nth basis function and the corresponding weight respectively.

[0028] Step 2.3, decoupling calculation of the weight vector corresponding to the PDF shape of the powder particle size;

[0029] The decoupling calculation is to solve the weight vector using matrix inverse operation, as follows:

[0030] V(k)=(C T (y)C(y)) -1 C T (y)γ(y,u(k))

[0031] Step 3: Using an iterative learning method to select a set of optimal Gaussian basis functions, so that the approximation error of the powder particle size PDF shape is minimized;

[0032] The iterative learning is to continuously adjust the center and width values ​​of the RBF-NN basis function in each iteration to minimize the error between the actual measured PDF of the powder particle size and the approximate PDF of the powder particle size, thereby selecting a set of optimal Gaussian basis functions.

[0033] First, let the vector form of the actual measurement PDF of M powder particle sizes be as follows:

[0034] g(y)=[g1(y),g2(y),...,g M (y)]

[0035] Where M represents the total number of sampling points, g M (y) represents the actual measurement PDF at the Mth sampling point;

[0036] Assume that the powder particle size approximation PDF obtained after the i-th iteration is:

[0037] γ i (y) = [γ i,1 (y),γ i,2 (y),...,γ i,M (y)]

[0038] In the formula, γ i,M (y) represents the approximate PDF at the Mth sampling point after the i-th iteration;

[0039] In order to evaluate the effect of powder particle size PDF shape approximation, the following performance index function is defined:

[0040]

[0041] In the formula, J i,m It represents the approximation error between the approximate PDF of the powder particle size at the mth sampling point after the ith iteration and the actual PDF.

[0042] Written in vector form:

[0043] E i =[J i,1 ,J i,2 ,...,J i,M ] T

[0044] In the formula, J i,M represents the approximation error between the approximate PDF of the Mth sampling point after the i-th iteration and the actual PDF, E i represents the error vector composed of the approximation errors of all sampling points after the i-th iteration, and E i All components are non-negative.

[0045] An algorithm for updating the center and width values ​​based on an iterative learning rule is used to adjust the basis function parameters:

[0046]

[0047] In the formula, μ l,i and σ l,i Respectively represent the center and width of the lth basis function after the i-th iteration, and the learning parameter α u and β σ They are defined as follows:

[0048]

[0049] In the formula, λ i and λ i ′ is the weight coefficient, ζ u and σ is the learning rate.

[0050] Step 4: Use subspace identification to establish a linear dynamic model of weights.

[0051] The method for constructing the linear dynamic model of the weights is:

[0052] Ignoring process and measurement noise, the linear dynamic model of the weights is described by the state space model of the following deterministic system:

[0053]

[0054] In the formula, u k ∈R m 、v k ∈R l and x k ∈R n Respectively represent the input vector, output vector and state vector at the kth sampling moment, R m , R l and R nThey represent m-dimensional, l-dimensional and n-dimensional column vectors respectively, and A, B, C, D are system matrices of corresponding dimensions.

[0055] Step 4.1, determining the input and output variables of subspace modeling; wherein the input variables are the control data of the grinding disc gap and the feeding amount, and the output data are the weight vector data corresponding to the PDF shape of the powder particle size;

[0056] Step 4.2, construct the Hankel data matrix:

[0057]

[0058]

[0059]

[0060]

[0061] Wherein, the Hankel matrix contains i blocks of rows and j blocks of columns, that is, each element in the above matrix is ​​a column vector;

[0062] Among them, u k ∈R m represents the input vector at the kth sampling moment, v k ∈R l Represents the output vector at the kth sampling moment, and the value range of k is [0,2i+j-2].

[0063] Step 4.3: Calculate the oblique projection O using the oblique projection formula i and O i-1 :

[0064]

[0065]

[0066] Where W p =[U p V p ] T , W p + =[U p + V p + ] T , the matrix U p + , U f - 、V p + and V f - is defined as follows:

[0067]

[0068]

[0069]

[0070]

[0071] Wherein, the superscripts + and - respectively indicate that the number of row blocks of the corresponding matrix increases by one row and decreases by one row.

[0072] Step 4.4, oblique projection O i Perform singular value decomposition to obtain the generalized energy matrix Γ i and the estimated value of the state vector sequence At the same time, determine the order n of the system:

[0073]

[0074] Where U1, U2, S1, S2, V1, V2 are all matrices O i The decomposed submatrix.

[0075] Step 4.5: Oblique projection O i-1 Perform singular value decomposition and calculate Γ i-1 and the estimated value of the state vector sequence at the next moment

[0076]

[0077] Where U a1 , U a2 , S a1 , S a2 、V a1 、V a2 Both are matrices O i-1 The sub-matrix obtained by decomposition; Γ i-1 Represents Γ i Remove the submatrix obtained by removing the last l rows.

[0078] Step 4.6: Extract the system matrices A, B, C, and D by solving the following least squares problem:

[0079]

[0080] Where U i =[u i u i+1 ... u i+j-1 ],V i =[v i vi+1 ... v i+j-1 ].

[0081] Step 5, using the linear dynamic model of weights established in step 4, combined with the RBF-NN approximation of the PDF shape of the powder particle size in steps 2 and 3, a random distribution model of the disc grinding system, i.e., a dynamic model of the PDF shape of the powder particle size, is obtained;

[0082] The dynamic model of the powder particle size PDF shape is:

[0083]

[0084] The beneficial effects of adopting the above technical solution are:

[0085] The present invention provides a powder particle size PDF shape modeling method based on iterative learning and subspace identification. In terms of RBF-NN basis function selection, the previous method of manually selecting parameters based on experience is abandoned, and an iterative learning algorithm is used to select a group of Gaussian basis functions with optimal center values ​​and widths, so that the approximation error of the powder particle size PDF shape is minimized, which greatly improves the modeling effect. On the other hand, in terms of dynamic modeling of weights, in view of the problems of complex model structure, difficult parameter selection, and difficulty in practical application caused by the current common use of neural networks to establish nonlinear dynamic models of weights, subspace identification is used to establish a linear dynamic model between control quantity input and weight vectors. Subspace identification directly establishes a model based on historical data without any prior knowledge. The system matrix of the model is directly extracted from the rows and columns of a specific matrix, which can ensure good modeling accuracy and reduce model complexity, and is particularly in line with the requirements for simple and easy-to-use models in practical applications. While ensuring a good powder particle size PDF shape modeling effect, a good model foundation is laid for the subsequent tracking control of the powder particle size PDF shape. The present invention is reasonably designed, easy to implement, and has good practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 It is a structural schematic diagram of the disc grinding system in a specific embodiment of the present invention;

[0087] In the figure: 1-feeding device; 2-grinding disc grinding device; 3-grinding disc speed adjustment device; 4-grinding disc gap servo control device; 5-sampling device; 6-laser particle size analyzer; 7-powder recovery device;

[0088] Figure 2 is a flow chart of a powder particle size PDF shape modeling method of a disc grinding system in a specific embodiment of the present invention;

[0089] Figure 3It is a schematic diagram of the actual measured value of the powder particle size PDF shape in the specific embodiment of the present invention;

[0090] Figure 4 is a schematic diagram of the powder particle size PDF shape modeling effect when iterative learning is not used in a specific embodiment of the present invention;

[0091] Figure 5 is a schematic diagram of the powder particle size PDF shape modeling error when iterative learning is not used in a specific embodiment of the present invention;

[0092] Figure 6 It is a schematic diagram of the powder particle size PDF shape modeling effect when iterative learning is adopted in a specific embodiment of the present invention;

[0093] Figure 7 It is a schematic diagram of the powder particle size PDF shape modeling error when iterative learning is adopted in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0094] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0095] This embodiment provides a disc grinding system, such as Figure 1 As shown, it includes a feeding device 1, a grinding disc grinding device 2, a grinding disc speed regulating device 3, a grinding disc gap servo control device 4, a sampling device 5, a laser particle size analyzer 6, and a powder recovery device 7;

[0096] The operation process of the disc grinding system is as follows: first, the material (such as wheat, rice, etc.) is uniformly fed into the disc grinding device 2 through the feeding device 1; then, the disc grinding device 2 grinds the material into powder at a certain speed and gap under the control of the disc speed regulating device 3 and the disc gap servo control device 4; the obtained powder is intermittently sampled by the sampling device 5 and then sent to the laser particle size analyzer 6 for detection to obtain the powder particle size distribution PDF shape data; the excess powder is recovered by the powder recovery device 7;

[0097] This embodiment establishes a random distribution model of the disc grinding system through the on-site input and output data collected during the actual operation of the disc grinding system. The input data is the control data of the grinding disc gap and the feeding amount, and the output data is the powder particle size PDF shape data.

[0098] A powder particle size PDF shape modeling method based on iterative learning and subspace identification, such as Figure 2 As shown, the specific steps include:

[0099] Step 1, collecting the control data of the grinding disc gap and the feeding amount during the operation of the grinding system, as well as the powder particle size PDF shape data;

[0100] Step 2: Using Gaussian radial basis function neural network (RBF-NN), the weight vector corresponding to the powder particle size PDF shape is decoupled and calculated, so that the powder particle size PDF is expressed as a product of a set of basis functions and corresponding weights;

[0101] Step 2.1, the grinding disc gap and feed amount are taken as the input variables of the system, and the powder particle size PDF shape is taken as the output variable of the system.

[0102] Step 2.2: Use a Gaussian radial basis function neural network (RBF-NN) to approximate the powder particle size PDF shape.

[0103] The approximation method is specifically:

[0104] For a non-Gaussian random distribution dynamic system such as a disc grinding system, the powder particle size is represented as a uniformly bounded random variable y∈[a,ζ], where a and ξ represent the value range of y;

[0105] Let u(k)∈R m R is the control variable input for controlling the powder particle size PDF shape. m represents an m-dimensional column vector;

[0106] The shape of the powder particle size PDF in the disc grinding process is represented by its probability density function γ(y,u(k)) at each sampling time k:

[0107]

[0108] Where P(a≤y<ξ,u(k)) represents the probability that the output powder particle size falls within the interval [a,ξ] when the control input u(k) is applied to the disc grinding system;

[0109] Assuming that the interval [a, b] is given in advance, where a and b represent the left and right endpoints of the random variable y, and the PDF is continuous and bounded, the powder particle size PDF shape is approximated by the following Gaussian radial basis function neural network RBF-NN:

[0110]

[0111] In the formula, w l (u(k)) is the basis function R controlled by the control input u(k) l (y), l represents the lth network node of the Gaussian radial basis function neural network RBF-NN, and n is the number of network nodes;

[0112] The basis function R defined in the interval [a,b] l (y) is as follows:

[0113]

[0114] In the formula, μ l and σ l Respectively represent the center and width of the lth basis function;

[0115] At this time, the powder particle size PDF shape is expressed as the product of the basis function and the corresponding weight

[0116]

[0117] In the formula, C(y) is the row vector composed of n basis functions, V(k) is the corresponding weight vector, and R n (y) and w n (u(k)) represents the nth basis function and the corresponding weight respectively.

[0118] Step 2.3, decoupling calculation of the weight vector corresponding to the PDF shape of the powder particle size;

[0119] The decoupling calculation is to solve the weight vector using matrix inverse operation, as follows:

[0120] V(k)=(C T (y)C(y)) -1 C T (y)γ(y,u(k))

[0121] It can be seen that when the powder particle size PDF shape data is measurable and the center and width of the RBF-NN basis function are determined, the corresponding weight vector can be easily obtained by matrix inversion operation.

[0122] Step 3: Using an iterative learning method to select a set of optimal Gaussian basis functions, so that the approximation error of the powder particle size PDF shape is minimized;

[0123] The iterative learning is to continuously adjust the center and width values ​​of the RBF-NN basis function in each iteration to minimize the error between the actual measured PDF of the powder particle size and the approximate PDF of the powder particle size, thereby selecting a set of optimal Gaussian basis functions.

[0124] First, let the vector form of the actual measurement PDF of M powder particle sizes be as follows:

[0125] g(y)=[g1(y),g2(y),...,g M (y)]

[0126] Where M represents the total number of sampling points, g M(y) represents the actual measurement PDF at the Mth sampling point;

[0127] Assume that the powder particle size approximation PDF obtained after the i-th iteration is:

[0128] γ i (y) = [γ i,1 (y),γ i,2 (y),...,γ i,M (y)]

[0129] In the formula, γ i,M (y) represents the approximate PDF at the Mth sampling point after the i-th iteration;

[0130] In order to evaluate the effect of powder particle size PDF shape approximation, the following performance index function is defined:

[0131]

[0132] In the formula, J i,m It represents the approximation error between the approximate PDF of the powder particle size at the mth sampling point after the ith iteration and the actual PDF.

[0133] Written in vector form:

[0134] E i =[J i,1 ,J i,2 ,...,J i,M ] T

[0135] In the formula, J i,M represents the approximation error between the approximate PDF of the Mth sampling point after the i-th iteration and the actual PDF, E i represents the error vector composed of the approximation errors of all sampling points after the i-th iteration, and E i All components are non-negative.

[0136] An algorithm for updating the center and width values ​​based on an iterative learning rule is used to adjust the basis function parameters:

[0137]

[0138] In the formula, μ l,i and σ l,i Respectively represent the center and width of the lth basis function after the i-th iteration, and the learning parameter α u and β σ They are defined as follows:

[0139]

[0140] In the formula, λ i and λi ′ is the weight coefficient, ζ u and σ is the learning rate.

[0141] Step 4: Use subspace identification to establish a linear dynamic model of weights.

[0142] The method for constructing the linear dynamic model of the weights is:

[0143] Ignoring process and measurement noise, the linear dynamic model of the weights is described by the state space model of the following deterministic system:

[0144]

[0145] In the formula, u k ∈R m 、v k ∈R l and x k ∈R n Respectively represent the input vector, output vector and state vector at the kth sampling moment, R m , R l and R n They represent m-dimensional, l-dimensional and n-dimensional column vectors respectively, and A, B, C, D are system matrices of corresponding dimensions.

[0146] Step 4.1, determining the input and output variables of subspace modeling; wherein the input variables are the control data of the grinding disc gap and the feeding amount, and the output data are the weight vector data corresponding to the PDF shape of the powder particle size;

[0147] Step 4.2, construct the Hankel data matrix:

[0148]

[0149]

[0150]

[0151]

[0152] Wherein, the Hankel matrix contains i blocks of rows and j blocks of columns, that is, each element in the above matrix is ​​a column vector;

[0153] Among them, u k ∈R m represents the input vector at the kth sampling moment, v k ∈R l Represents the output vector at the kth sampling moment, and the value range of k is [0,2i+j-2].

[0154] Step 4.3: Calculate the oblique projection O using the oblique projection formula i and O i-1 :

[0155]

[0156]

[0157] Where W p =[U p V p ] T , W p + =[U p + V p + ] T , the matrix U p + , U f - 、V p + and V f - is defined as follows:

[0158]

[0159]

[0160]

[0161]

[0162] Wherein, the superscripts + and - respectively indicate that the number of row blocks of the corresponding matrix increases by one row and decreases by one row.

[0163] Step 4.4, oblique projection O i Perform singular value decomposition to obtain the generalized energy matrix Γ i and the estimated value of the state vector sequence At the same time, determine the order n of the system:

[0164]

[0165] Where U1, U2, S1, S2, V1, V2 are all matrices O i The decomposed submatrix.

[0166] Step 4.5: Oblique projection O i-1 Perform singular value decomposition and calculate Γ i-1 and the estimated value of the state vector sequence at the next moment

[0167]

[0168] Where U a1 , U a2 , S a1 , S a2 、V a1 、V a2 Both are matrices O i-1 The sub-matrix obtained by decomposition; Γ i-1 Represents Γ i Remove the submatrix obtained by removing the last l rows.

[0169] Step 4.6: Extract the system matrices A, B, C, and D by solving the following least squares problem:

[0170]

[0171] Where U i =[u i u i+1 ... u i+j-1 ],V i =[v i v i+1 ... v i+j-1 ].

[0172] Step 5, using the linear dynamic model of weights established in step 4, combined with the RBF-NN approximation of the PDF shape of the powder particle size in steps 2 and 3, a random distribution model of the disc grinding system, i.e., a dynamic model of the PDF shape of the powder particle size, is obtained;

[0173] The dynamic model of the powder particle size PDF shape is:

[0174]

[0175] At this point, a random distribution model of the disc grinding system can be established with the disc gap and feed amount as the control input and the powder particle size PDF shape as the output.

[0176] In order to verify the effectiveness of the proposed powder particle size PDF shape modeling method based on iterative learning and subspace identification, an experiment was carried out using 55 sets of on-site input and output data collected when the disc grinding system was running smoothly.

[0177] The input data are the control quantity data of the grinding disc gap and the feeding amount, which are collected by the lower computer PLC.

[0178] The output data is powder particle size PDF shape data, which is obtained by laser particle size analyzer detection.

[0179] The laser particle size analyzer is a German Sympatec laser particle size analyzer. The obtained powder can be intermittently sampled through a sampling device and then sent to the laser particle size analyzer for testing to obtain the powder particle size PDF shape data.

[0180] Figure 3 is the actual measured value of the powder particle size PDF shape, which shows obvious non-Gaussian characteristics. Combined with the characteristics of the powder particle size PDF shape, the present invention selects RBF-NN with three Gaussian basis functions to verify the modeling effect.

[0181] Wherein, the initial center and width values ​​of the basis function when the iterative learning method is not used are:

[0182] μ1=0.3, μ2=0.5, μ3=0.8,

[0183] After adopting the iterative learning method, the center and width values ​​of the basis function are finally obtained as follows:

[0184] μ1=0.302, μ2=0.5002, μ3=0.784,

[0185] Then, the powder particle size PDF shape modeling method is used to establish a random distribution model of the disc grinding system, wherein: Figure 4 The PDF shape modeling effect of powder particle size when iterative learning is not used. Figure 5 The PDF shape modeling error of powder size when iterative learning is not used. Figure 6 The PDF shape modeling effect of powder particle size when iterative learning is used. Figure 7 It is the powder particle size PDF shape modeling error when using iterative learning.

[0186] It can be seen that compared with the modeling effect when iterative learning is not used, a set of Gaussian basis functions with better parameters are selected after iterative learning, which greatly improves the powder particle size PDF shape modeling effect; and the powder particle size PDF shape modeling error presents a flat feature as a whole, indicating that the linear dynamic model of weights established by using the subspace identification can ensure that the model is simple and easy to use while also having high accuracy, and can well approximate the actual measurement value of the powder particle size PDF shape.

[0187] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. A powder particle size PDF shape modeling method based on iterative learning and subspace identification, characterized in that: The following steps are involved: Step 1, collecting the control data of the grinding disc gap and the feeding amount during the operation of the grinding system, as well as the powder particle size PDF shape data; Step 2: Using Gaussian radial basis function neural network (RBF-NN), the weight vector corresponding to the powder particle size PDF shape is decoupled and calculated, so that the powder particle size PDF is expressed as a product of a set of basis functions and corresponding weights; Step 3: Using an iterative learning method to select a set of optimal Gaussian basis functions, so that the approximation error of the powder particle size PDF shape is minimized; The iterative learning is to continuously adjust the center and width values ​​of the RBF-NN basis function in each iteration to minimize the error between the actual measured PDF of the powder particle size and the approximate PDF of the powder particle size, thereby selecting a set of optimal Gaussian basis functions; First, let the vector form of the actual measurement PDF of M powder particle sizes be as follows: g(y)=[g1(y),g2(y),...,g M (y)] Where M represents the total number of sampling points, g M (y) represents the actual measurement PDF at the Mth sampling point; Assume that the powder particle size approximation PDF obtained after the i-th iteration is: c i (y)=[γ i,1 (y),γ i,2 (y),...,γ i,M (y)] In the formula, γ i,M (y) represents the approximate PDF at the Mth sampling point after the i-th iteration; In order to evaluate the effect of powder particle size PDF shape approximation, the following performance index function is defined: In the formula, J i,m It represents the approximation error between the approximate PDF and the actual PDF of the powder particle size at the mth sampling point after the i-th iteration; Written in vector form: E i =[J i,1 ,J i,2 ,...,J i,M ] T In the formula, J i,M represents the approximation error between the approximate PDF of the Mth sampling point after the i-th iteration and the actual PDF, E i represents the error vector composed of the approximation errors of all sampling points after the i-th iteration, and E i All components are non-negative; An algorithm for updating the center and width values ​​based on an iterative learning rule is used to adjust the basis function parameters: In the formula, μ l,i and σ l,i Respectively represent the center and width of the lth basis function after the i-th iteration, and the learning parameter α u and β σ They are defined as follows: In the formula, λ i and λ i ′ is the weight coefficient, ζ u and σ is the learning rate; Step 4: Use subspace identification to establish a linear dynamic model of weights; The method for constructing the linear dynamic model of the weights is: Ignoring process and measurement noise, the linear dynamic model of the weights is described by the state space model of the following deterministic system: In the formula, u k ∈R m 、v k ∈R l and x k ∈R n Respectively represent the input vector, output vector and state vector at the kth sampling moment, R m , R l and R n denote m-dimensional, l-dimensional and n-dimensional column vectors respectively, and A, B, C, D are system matrices of corresponding dimensions; The step 4 specifically comprises the following steps: Step 4.1, determine the input and output variables of subspace modeling; the input variables are the control data of the grinding disc gap and the feeding amount, and the output variables are the weight vector data corresponding to the PDF shape of the powder particle size; Step 4.2, construct the Hankel data matrix: Wherein, the Hankel matrix contains i blocks of rows and j blocks of columns, that is, each element in the above matrix is ​​a column vector; Among them, u k ∈R m represents the input vector at the kth sampling moment, v k ∈R l Represents the output vector at the kth sampling moment, and the value range of k is [0, 2i+j-2]; Step 4.3: Calculate the oblique projection O using the oblique projection formula i and O i-1 : Where W p =[U p V p ] T , W p + =[U p + V p + ] T , the matrix U p + , U f - 、V p + and V f - is defined as follows: In the formula, the superscript + and - respectively indicate that the number of rows of the corresponding matrix increases or decreases by one row; Step 4.4, oblique projection O i Perform singular value decomposition to obtain the generalized energy matrix Γ i and the estimated value of the state vector sequence At the same time, determine the order n of the system: Where U1, U2, S1, S2, V1, V2 are all matrices O i The decomposed sub-matrix; Step 4.5: Oblique projection O i-1 Perform singular value decomposition and calculate Γ i-1 and the estimated value of the state vector sequence at the next moment Where U a1 , U a2 , S a1 , S a2 、V a1 、V a2 Both are matrices O i-1 The sub-matrix obtained by decomposition; Γ i-1 Represents Γ i Remove the submatrix obtained by removing the last l rows; Step 4.6: Extract the system matrices A, B, C, and D by solving the following least squares problem: where U i = [u i u i+1 …u i+j-1 , V i = [v i v i+1 …v i+j-1 ; Step 5: Using the linear dynamic model of weights established in step 4, combined with the RBF-NN approximation of the powder particle size PDF shape in steps 2 and 3, a random distribution model of the disc grinding system, i.e., a dynamic model of the powder particle size PDF shape, is obtained.

2. The powder particle size PDF shape modeling method based on iterative learning and subspace identification according to claim 1 is characterized in that: The step 2 specifically includes the following steps: Step 2.1, the grinding disc gap and feed amount are used as the input variables of the system, and the powder particle size PDF shape is used as the output variable of the system; Step 2.2, using a Gaussian radial basis function neural network RBF-NN to approximate the powder particle size PDF shape; Step 2.3, perform decoupling calculation on the weight vector corresponding to the powder particle size PDF shape.

3. The powder particle size PDF shape modeling method based on iterative learning and subspace identification according to claim 2 is characterized in that: The approximation method described in step 2.2 is specifically: For a dynamic system with non-Gaussian random distribution, the powder particle size is represented as a uniformly bounded random variable y∈[a,ζ], where a and ξ represent the value range of y; Let u(k)∈R m R is the control variable input for controlling the powder particle size PDF shape. m represents an m-dimensional column vector; The shape of the powder particle size PDF in the disc grinding process is represented by its probability density function γ(y,u(k)) at each sampling time k: Where P(a≤y<ξ,u(k)) represents the probability that the output powder particle size falls within the interval [a,ξ] when the control input u(k) is applied to the disc grinding system; Assuming that the interval [a, b] is given in advance, where a and b represent the left and right endpoints of the random variable y, and the PDF is continuous and bounded, the powder particle size PDF shape is approximated by the following Gaussian radial basis function neural network RBF-NN: In the formula, w l (u(k)) is the basis function R controlled by the control input u(k) l (y), l represents the lth network node of the Gaussian radial basis function neural network RBF-NN, and n is the number of network nodes; The basis function R defined in the interval [a,b] l (y) is as follows: In the formula, μ l and σ l Respectively represent the center and width of the lth basis function; At this time, the powder particle size PDF shape is expressed as the product of the basis function and the corresponding weight In the formula, C(y) is the row vector composed of n basis functions, V(k) is the corresponding weight vector, and R n (y) and w n (u(k)) represents the nth basis function and the corresponding weight respectively.

4. The powder particle size PDF shape modeling method based on iterative learning and subspace identification according to claim 2 is characterized in that: The decoupling calculation described in step 2.3 is to solve the weight vector using matrix inversion operation, as follows: V(k)=(C T (y)C(y)) -1 C T (y)γ(y,u(k))。 5. The powder particle size PDF shape modeling method based on iterative learning and subspace identification according to claim 1 is characterized in that: The dynamic model of the powder particle size PDF shape in step 5 is: