CN113191092A - Industrial process product quality soft measurement method based on orthogonal increment random configuration network - Google Patents

Abstract

The invention discloses an industrial process product quality soft measurement method based on an orthogonal increment random configuration network, which comprises the following steps: acquiring historical industrial process auxiliary data and corresponding product quality data, initializing parameters required by orthogonal increment random configuration network learning, and constructing an orthogonal supervision mechanism to select candidate nodes; and selecting an optimal hidden layer node from the newly-added candidate nodes, finding an optimal hidden layer parameter corresponding to the optimal hidden layer node, converting the output of the optimal node into an orthogonal vector, constructively evaluating a model output weight, and establishing an industrial process product quality soft measurement model based on an orthogonal incremental random configuration network. The invention can quickly establish the industrial process product quality soft measurement model with optimal parameters in the form of a construction method, avoids redundant nodes of an original random configuration network by utilizing an orthogonal technology, and ensures the lightness and the accuracy of the soft measurement model, namely, the soft measurement model has good compactness and generalization performance.

Description

Industrial process product quality soft measurement method based on orthogonal increment random configuration network

Technical Field

The invention relates to the technical field of soft measurement of industrial process product quality indexes, in particular to a soft measurement method of industrial process product quality indexes based on an orthogonal increment random configuration network.

Background

The soft measurement technology which aims at reducing the production cost, improving the production efficiency and quality and accurately forecasting the quality index of the produced product in real time is an important research direction in the field of current complex industrial process control and has profound significance and practical application value. With the development of sensors, internet of things and data storage technologies, the soft measurement technology for product quality prediction of complex industrial processes based on machine learning is widely applied to various industrial processes. Meanwhile, product quality soft measurement is essentially a regression fitting problem, but in this field, a traditional fitting model is not ideal in approximating effect performance. The random configuration network is an advanced single-hidden-layer random weight network with infinite approximation characteristics in recent years, and a large number of regression and classification tests prove that the random configuration network has obvious advantages in the aspects of compactness, quick learning, generalization performance and the like. However, nodes with small output still exist in the model structure, the redundant nodes are poor in quality, do not contribute much to reducing residual errors, reduce compactness of the network, and have risks of ill-conditioned matrixes, so that generalization performance is reduced, and prediction quality is affected.

Disclosure of Invention

Aiming at the fitting defect of the existing industrial process product quality soft measurement method, the invention provides an industrial process product quality soft measurement method based on an orthogonal increment random configuration network, which comprises the following steps:

step 1, acquiring historical industrial process auxiliary data and corresponding product quality data, initializing parameters required by learning of an orthogonal incremental random configuration network soft measurement model, constructing an orthogonal form supervision mechanism, and selecting hidden nodes meeting constraints as newly-added candidate nodes;

and 2, finding out optimal hidden layer parameters from the newly added candidate nodes, converting the optimal node output into the corresponding optimal orthogonal vector, constructively evaluating the model output weight, and obtaining the industrial process product quality soft measurement model based on the orthogonal increment random configuration network.

The step 1 comprises the following steps:

step 1-1, establishing an orthogonal increment random configuration network: acquiring N groups of historical industrial process auxiliary data and corresponding product quality data { X, T }, and ith group of historical process auxiliary data X_iContaining d auxiliary process variables, which correspond to product quality data t_iContaining m product quality data, i being 1-N, then inputting sample matrix X ═ X₁,x₂,…,x_i,…,x_ND auxiliary process variable sets of the ith group are marked as { x }_i1,x_i2,…,x_id}，x_idRepresents the ith set of the mth auxiliary process variables; and taking the product quality index T as an output variable of the orthogonal increment random configuration network, and then outputting a sample T ═ T₁,t₂,…,t_i,…,t_N},t_i＝{t_i1,t_i2,…,t_imIs the ith set of process assistance data x_iM sets of product quality data in corresponding product quality data, where t_imRepresents the ith set of mth product data values;

parameters required for initializing orthogonal increment random configuration network learning, including maximum hidden node number L_maxMaximum number of times of configuration T_maxDesired tolerance ε, range parameter λ,Range of random configuration of hidden layer parameters γ: ═ λ_min:Δλ:λ_maxA very small positive number σ (usually 1e-6), an initial residual e₀T, activation function g (x); wherein λ_minFor a given minimum range parameter, λ_maxΔ λ is a given interval value for a given maximum range parameter;

step 1-2, in the process of constructing the orthogonal increment random configuration network, when the L-th node is added, giving the 0 th round non-negative sequence r of the L-th iteration_L,0L/(L +1) and the sequence μ_L,0＝(1-r_L,0) V (L + 1); within an adjustable range [ - λ [ ]_min,λ_min]Internal random generation of T_maxFor hidden layer parameters, i.e. T of Lth node_maxFor the input weight w_LAnd bias b_L；

Step 1-3, constructing an orthogonal supervision mechanism of hidden layer nodes:

ξ_L,q＝<e_L-1,q,v_L>²/||v_L||²-(1-r_L,0-μ_L,0)||e_L-1,q||²,q＝1,2,…,m；

wherein, the number of m as the leading variable is also the dimension of the leading variable to be predicted, ξ_L,qFor the q-th output in the L-th iteration, the corresponding orthogonal supervised value, e_L-1,qNetwork residual, v, corresponding to the q-th output after the L-1-th node is added_LIs that the L < th > iteration satisfies | | | v_L||>The calculation form of the newly added node orthogonal vector of the sigma is as follows:

node hidden layer output

Superscript T is the transpose of a matrix or vector;

calculate the node h_LOrthogonal vector of (2):

if L is 1, v₁＝h₁(ii) a Wherein h is₁Is the output of the first hidden node;

if L is>1，

Wherein

Orthogonal vectors for the 1 st to L-1 st best nodes;

if the node satisfies min { ξ_L,1,ξ_L,2,…,ξ_L,q,…,ξ_L,mIs more than or equal to 0, wherein ξ is_L,qAnd the orthogonal supervision value which represents the q output in the L iteration is selected as a candidate node.

The step 2 comprises the following steps:

step 2-1, substituting the candidate nodes obtained in the step 1-3 into the orthogonal supervision values in the L iteration

Obtaining a candidate node group:

k≤T_maxwherein

The node orthogonal supervision value of the kth random configuration in the Lth iteration is represented;

in a candidate node group

Find out the orthogonal supervision mechanism xi in the L-th iteration_LMaximum set of hidden layer parameters, i.e. xi_LThe maximum set of hidden layer parameters is the best hidden layer parameters satisfying the orthogonal supervision mechanism, including the best input weight w_L ^*Optimum bias b_L ^*And the best orthogonal vector v_L ^*V at this time_L ^*Randomly configuring subsequent output parameters of a network hidden layer as orthogonal increments;

if front j wheels T_maxNo optimal hidden layer parameter is found that satisfies the condition, j ═ 1,2, …, ((λ @)_max-λ_min) And 1, automatically updating the next round of non-negative sequence by adopting the following formula, and then returning to the step 1-2 to update the next round of sequence mu_L,j＝(1-r_L,j) /(L +1), in the next configuration scope

Inner regeneration candidate nodes:

wherein the auxiliary variable τ_j∈(1/2(1-r_L,j-1),1-r_L,j-1)；

If it is (lambda)_max-λ_min) [ Delta ] lambda wheel T_maxI.e., [ - λ ]_max,λ_max]Within the range, the optimal hidden layer parameter satisfying the condition is not found yet, and a larger lambda needs to be given_maxReturning to the step 1-1 to reestablish the orthogonal increment random configuration network;

randomly configuring a hidden layer output vector of the network according to the current orthogonal increment to calculate the network output weight as follows:

β_L＝[β_L,1,β_L,2,…,β_L,q,…,β_L,m]_1×m，

wherein

Orthogonal vector of current best hidden layer node, e_L-1Adding the model residual error for the previous node;

the output weight beta of the current soft measurement model is:

wherein beta is_LThe output weight of the L-th node;

model residual e at this time_LComprises the following steps: e.g. of the type_L＝e_L-1-v_L ^*β_L；

When the number of hidden nodes of the orthogonal incremental random configuration network exceeds the given maximum number of hidden nodes or the residual error in the current iteration meets the expected tolerance, no new node is added, and an industrial process product quality soft measurement model based on the orthogonal incremental random configuration network is obtained; otherwise, returning to the step 1-2.

Has the advantages that: compared with the prior art, the method has the advantages that the method can establish the industrial process product quality soft measurement model with the infinite approximation characteristic and the optimal parameters in a structural method mode, does not need a complex retraining process, can ensure the accuracy of the model, and has good compactness and generalization performance.

Drawings

Fig. 1 is a schematic diagram of an orthogonal incremental random configuration network model.

Fig. 2 is a graph showing the variation trend of the training error of the model in the soft measurement of the ore grinding process.

Fig. 3 is an estimated scatter plot of a soft measurement of ore grind size.

Detailed Description

The following describes in detail the embodiments of the present invention in conjunction with the prediction of quality indicators for a particular grinding process.

The fitting model structure used by the invention is shown in FIG. 1 and comprises an input layer, an orthogonal hidden layer and an output layer; d is 5 and m is 1. The modeling method mainly comprises the following steps:

step one, selecting 500 groups of historical data measured in the traditional hematite grinding process from a historical database in the grinding process, namely each group contains the current c of the ball mill₁And spiral classifier current c₂Mill feed c₃Mill inlet feed water flow c₄And the overflow concentration c of the classifier₅Five auxiliary process variable data, with x_i＝{c_i1,c_i2,…,c_i5Denotes the input data after homogenization, and the correspondingProduct quality data, i.e. grinding particle size t_i. Wherein 350 groups are used as training set and 150 groups are used as testing set. Then the input sample is X ═ X₁,x₂,…,x_i,…,x₃₅₀Therein of

The output sample is T ═ T₁,t₂,…,t_i,…,t₃₅₀}。

Initializing parameters required by learning of an orthogonal increment random configuration network soft measurement model, wherein the maximum hidden node number L_maxMaximum number of dispositions T100_max10, desired tolerance ∈ 0.05, hidden layer parameter random configuration range γ: {1:1:10}, one very small positive number σ ═ 1e-6, initial residual e₀The activation function selects Sigmoid (S-curve) function g (x) 1/(1+ exp (-x));

in the process of constructing the orthogonal increment random configuration network, when the L-th node is added:

given round 0 non-negative sequence r_L,0L/(L +1) and the sequence μ_L,0＝(1-r _L,0)/(L+1)；

Within the adjustable range of [ -1,1 [)]Generating 10 pairs of hidden layer parameters, namely input weight w at internal random_LAnd bias b_LSubstituting it into the activation function g (x);

constructing an orthogonal supervision mechanism:

ξ_L＝<e_L-1,v_L>²/||v_L||²-(1-r_L,1-μ_L,1)||e_L-1||²

wherein e is_L-1For the residual after the last iteration, v_LIs that the L < th > iteration satisfies | | | v_L||>1e-6, the calculation form of the added node orthogonal vector is as follows:

hidden layer output of current node

Wherein x_i＝{c_i1,c_i2,…,c_i5T denotes a transposition operation;

calculating the orthogonal vector corresponding to the node:

if L is 1, v₁＝h₁；

If L is>1,

Wherein

Orthogonal vectors for the 1 st to L-1 st best nodes;

if the node satisfies min { ξ_LAnd if the rate is larger than or equal to 0, selecting the node as a candidate node.

The second step comprises the following steps:

obtaining a candidate node group:

in which ξ_L＝<e_L-1,v_L>²/||v_L||²-(1-r_L,0-μ_L,0)||e_L-1||²。

Find the order xi_LThe maximum set of hidden layer parameters is the best hidden layer parameter w satisfying the orthogonal supervision mechanism_L ^*，b_L ^*And v_L ^*At this time, the hidden node output is replaced by v_L ^*；

If the front j (j ═ 1,2, …,8) wheel T_maxIf the optimal hidden layer parameter meeting the condition is not found, automatically updating the next round of sequence:

wherein the auxiliary variable τ_j∈(1/2(1-r_L,j-1),1-r_L,j-1) Returning to the step one, updating the sequence mu_L,j＝(1-r_L,j) V (L +1), in the next configurationRange of placement

Internally regenerating candidate nodes;

calculating the output weight according to the current hidden layer output vector as follows:

the output weight of the current orthogonal increment random configuration network is as follows:

at this time, the network residual error randomly configured by the orthogonal increment is as follows: e.g. of the type_L＝e_L-1-v_L ^*β_L。

When the number of hidden nodes of the orthogonal increment random configuration network exceeds 100 or the residual error in the current iteration meets the expected tolerance of 0.05, no new node is added, modeling is completed, and the ore grinding granularity soft measurement model based on the orthogonal increment random configuration network is obtained; otherwise, returning to the first step and continuing to construct the network.

And step three, testing the trained ore grinding granularity soft measurement model based on the orthogonal increment random configuration network by using the test data set, namely estimating the ore grinding granularity value of the test data by using the soft measurement model.

In order to illustrate the superiority of the method, a quality index of the model method in the ore grinding process, namely a change trend graph of a training error in an ore grinding granularity soft measurement model is provided, as shown in fig. 2, as can be seen from fig. 2, the error of the model is continuously reduced along with the increase of nodes, and 7 nodes can meet the given precision requirement, so that the compactness of the model is ensured. In order to further illustrate the prediction estimation accuracy of the model established by the invention, a test result scatter diagram of ore grinding quality index ore grinding granularity soft measurement is provided, as shown in fig. 3, the fitting effect of the estimation result of 150 groups of test data and actual data is good, the estimation error is small, and the accuracy is high. The model provided by the invention is used for estimating the ore grinding quality index ore grinding granularity, has the advantages of high convergence rate and compact model structure, effectively solves the problem of interference of overfitting on the model caused by the risk of the ill-conditioned matrix brought by redundant nodes, and has the advantages of high estimation precision, low cost and high practical value.

It will be appreciated by those of ordinary skill in the art that the examples described herein are intended to assist those of ordinary skill in the art in understanding the manner in which the invention is practiced, and it is to be understood that the scope of the invention is not to be limited to such specific statements and examples. The present invention is not limited to the above embodiments, but is also applicable to other technical fields.

Claims (3)

1. A soft measurement method for product quality of industrial process based on orthogonal increment random configuration network is characterized by comprising the following steps:

step 1, acquiring historical industrial process auxiliary data and corresponding product quality data, initializing parameters required by orthogonal increment random configuration network learning, constructing an orthogonal form supervision mechanism, and selecting hidden nodes meeting constraints as newly-added candidate nodes;

and 2, finding out optimal hidden layer parameters from the newly added candidate nodes, converting the optimal node output into the corresponding optimal orthogonal vector, constructively evaluating the model output weight, and obtaining the industrial process product quality soft measurement model based on the orthogonal increment random configuration network.

2. The method of claim 1, wherein step 1 comprises:

step 1-1, establishing an orthogonal increment random configuration network: acquiring N groups of process auxiliary data of historical industrial processes and corresponding product quality data { X, T }, and ith group of historical process auxiliary data X_iContaining d auxiliary process variables, which correspond to product quality data t_iContaining m product quality data, i being 1-N, then inputting sample matrix X ═ X₁,x₂,…,x_i,…,x_ND auxiliary process variable sets of the ith group are marked as { x }_i1,x_i2,…,x_id}，x_idRepresents the ith set of the mth auxiliary process variables; and taking the product quality index T as an output variable of the orthogonal increment random configuration network, and then outputting a sample T ═ T₁,t₂,…,t_i,…,t_N},t_i＝{t_i1,t_i2,…,t_imIs the ith set of process assistance data x_iM sets of product quality data in corresponding product quality data, where t_imRepresents the ith set of mth product data values;

parameters required for initializing orthogonal increment random configuration network learning, including maximum hidden node number L_maxMaximum number of times of configuration T_maxDesired tolerance epsilon, range parameter lambda, hidden layer parameter random configuration range gamma: ═ lambda_min:Δλ:λ_maxA positive number sigma, an initial residual e₀T, activation function g (x); wherein λ_minFor a given minimum range parameter, λ_maxΔ λ is a given interval value for a given maximum range parameter;

step 1-2, in the process of constructing the orthogonal increment random configuration network, when the L-th node is added, giving the 0 th round non-negative sequence r of the L-th iteration_L,0L/(L +1) and the sequence μ_L,0＝(1-r_L,0) V (L + 1); within an adjustable range [ - λ [ ]_min,λ_min]Internal random generation of T_maxFor hidden layer parameters, i.e. T of Lth node_maxFor the input weight w_LAnd bias b_L；

Step 1-3, constructing an orthogonal supervision mechanism of hidden layer nodes:

ξ_L,q＝<e_L-1,q,v_L>²/||v_L||²-(1-r_L,0-μ_L,0)||e_L-1,q||²,q＝1,2,…,m；

wherein ξ_L,qFor the q-th output in the L-th iteration, the corresponding orthogonal supervised value, e_L-1,qNetwork residual, v, corresponding to the q-th output after the L-1-th node is added_LIs satisfied with/in the L th iteration_L∥>The calculation form of the newly added node orthogonal vector of the sigma is as follows:

node hidden layer output

Superscript T is the transpose of a matrix or vector;

calculate the node h_LOrthogonal vector of (2):

if L is 1, v₁＝h₁(ii) a Wherein h is₁Is the output of the first hidden node;

if L is>1，

Wherein

Orthogonal vectors for the 1 st to L-1 st best nodes;

if the node satisfies min { ξ_L,1,ξ_L,2,…,ξ_L,q,…,ξ_L,mIs more than or equal to 0, wherein ξ is_L,qAnd the orthogonal supervision value which represents the q output in the L iteration is selected as a candidate node.

3. The method of claim 2, wherein step 2 comprises:

step 2-1, substituting the candidate nodes obtained in the step 1-3 into the orthogonal supervision values in the L iteration

Obtaining a candidate node group:

wherein

The node orthogonal supervision value of the kth random configuration in the Lth iteration is represented;

in a candidate node group

Finding out the orthogonal supervision value xi of the L-th node_LMaximum set of hidden layer parameters, i.e. xi_LThe maximum set of hidden layer parameters is the optimal hidden layer parameters meeting the orthogonal supervision mechanism, and comprises the optimal input weight w_L ^*Optimum bias b_L ^*And the best orthogonal vector v_L ^*V at this time_L ^*Randomly configuring subsequent output parameters of a network hidden layer as orthogonal increments;

if front j wheels T_maxNo optimal hidden layer parameter is found that satisfies the condition, j ═ 1,2, …, ((λ @)_max-λ_min) And 1, automatically updating the next round of non-negative sequence by adopting the following formula, and then returning to the step 1-2 to update the next round of sequence mu_L,j＝(1-r_L,j) /(L +1), in the next configuration scope

Inner regeneration candidate nodes:

wherein the auxiliary variable τ_j∈(1/2(1-r_L,j-1),1-r_L,j-1)；

If it is (lambda)_max-λ_min) [ Delta ] lambda wheel T_maxI.e., [ - λ ]_max,λ_max]Within the range, the optimal hidden layer parameter satisfying the condition is not found yet, and a larger lambda needs to be given_maxReturning to the step 1-1 to reestablish the orthogonal increment random configuration network;

calculating the output weight of the network according to the hidden layer output vector of the current orthogonal increment random configuration network as follows:

β_L＝[β_L,1,β_L,2,…,β_L,q,…,β_L,m]_1×m，

wherein

Orthogonal vector of current best hidden layer node, e_L-1,qAdding the q-th output residual error into the L-1-th node of the model;

the output weight β of the current orthogonal incremental random configuration network is:

wherein beta is_LThe output weight of the L-th node;

model residual e at this time_LComprises the following steps: e.g. of the type_L＝e_L-1-v_L ^*β_L；

When the number of hidden nodes of the orthogonal incremental random configuration network exceeds the given maximum number of hidden nodes or the residual error in the current iteration meets the expected tolerance, no new node is added, and an industrial process product quality soft measurement model based on the orthogonal incremental random configuration network is obtained; otherwise, returning to the step 1-2.

Images