Process index lightweight sensing method based on Bayesian random incremental learning

Abstract

The invention discloses a process index lightweight sensing method based on Bayesian random incremental learning, and the method comprises the following steps: 1, obtaining historical data of an industrial process, and carrying out the preprocessing of the data; 2, constructing a global constraint mechanism through a Grevile iteration method, and screening the generated hidden layer nodes based on the global constraint mechanism to form a candidate node pool; 3, selecting a node parameter which contributes to the accuracy of the random increment model most from the candidate node pool, calculating an output matrix of the node, and reconstructing an output matrix of a hidden layer; 4, optimizing the output weight of the hidden layer through a Bayesian statistical method; 5, hidden layer nodes are added in a point increment construction mode, and the steps 2-4 are repeated until the random increment model is constructed; and 6, sensing the process indexes by using the constructed random increment model. According to the method, the compactness and prediction precision of the model structure are improved, and the requirements of lightweight soft measurement modeling in the industrial process can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of process indicator perception, and in particular to a lightweight process indicator perception method based on Bayesian random incremental learning. Background Art

[0002] The operating mechanism and dynamic characteristics of industrial processes are usually very complex, making it difficult to build an accurate process model. At the same time, due to economic and technical limitations, as well as the inability to perform timely maintenance when the detection device is damaged, it is difficult to achieve online detection of process indicators. Therefore, it is necessary to use data-driven soft measurement technology to establish an intelligent perception model between process auxiliary variables that are easy to measure and process indicators.

[0003] Due to the limitations of storage and computing power, distributed DCS control systems for industrial processes often find it difficult to effectively cope with complex data-driven models and data analysis tasks. Therefore, establishing an efficient and lightweight data-driven model to perceive process indicators has become an urgent problem to be solved in the industrial field. In recent years, the incremental model based on random learning, namely the random incremental model, has become the preferred tool for lightweight modeling of industrial processes due to its high construction efficiency and lightweight model. In order to improve the compactness of the model, the random incremental model adds hidden layer nodes to match the modeling task requirements by constructing a constraint mechanism. However, the existing constraint mechanism mainly relies on the current network residual, and does not consider the impact of the new nodes on the existing model, resulting in the inability to effectively constrain the new nodes, resulting in a non-lightweight model. In addition, due to the complexity of the measurement environment, it is almost impossible to obtain completely clean measurement data, and these data are often full of uncertainty, which brings great challenges to the accurate estimation of the model. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a lightweight perception method for process indicators based on Bayesian random incremental learning to solve the technical problems that the prior art cannot accurately handle the uncertainty problem in the process and the model prediction accuracy is low.

[0005] The present invention provides a lightweight perception method for process indicators based on Bayesian random incremental learning, comprising the following steps:

[0006] Step 1: Obtain historical data of industrial processes and preprocess the data;

[0007] Step 2: Construct a global constraint mechanism through the Greville iteration method, and screen the generated hidden layer nodes based on the global constraint mechanism to form a candidate node pool;

[0008] Step 3: Select the node parameters that contribute most to the accuracy of the random incremental model from the candidate node pool, calculate the output matrix of the node, and reconstruct the output matrix of the hidden layer;

[0009] Step 4: Optimize the output weights of the hidden layer using Bayesian statistical methods;

[0010] Step 5: Add hidden layer nodes by point-increment construction, and repeat steps 2-4 until the random incremental model is built;

[0011] Step 6: Use the constructed random incremental model to perceive the process indicators.

[0012] Furthermore, in step 2, the expression of the global constraint constructed is: L ≥0,

[0013] in,

[0014] In the formula, e before is the residual of the random increment model;

[0015]

[0016] H L-1 Moore-Penrose generalized inverse of (X);

[0017]

[0018] T q Output data for the qth process indicator of T; s=1,2,...,T max .

[0019] Furthermore, in the step 2, the process of screening the generated hidden layer nodes by the global constraint mechanism also includes:

[0020] When no node in the generated hidden layer nodes satisfies the global constraint condition, the selection interval of the hidden layer nodes is adjusted. The specific method of adjusting the selection interval is as follows:

[0021] By adding a dynamic change value to the original selection interval, the adjusted selection interval is:

[0022] [-λ min +jΔλ λ min +jΔλ] d and [-λ min +jΔλ λ min +jΔλ]

[0023] Among them, [-λ min λmin ] d and [-λ min λ min ] is the original selection interval; jΔλ is the dynamic change value, j∈(1,M), M is the number of changes, Δλ is the interval value of the selected interval change.

[0024] Furthermore, the step 2 also includes scaling the learning parameters in the global constraint mechanism so that the learning parameters approach 1.

[0025] Furthermore, in step 3, the output matrix of the reconstructed hidden layer is:

[0026]

[0027] In the formula, is the output matrix of the node that contributes most to the accuracy of the random incremental model,

[0028]

[0029] H L-1 (X) is the hidden layer output matrix of L-1 nodes.

[0030] Furthermore, the specific process of step 4 is as follows:

[0031] Step 41: Set 4 minimal hyperparameters;

[0032] Step 42: Update the posterior mean and covariance. The update formula is:

[0033]

[0034] Where γ is a hyperparameter; σ 2 is the variance;

[0035] Step 43: Update the hyperparameters and variance. The update formula is:

[0036]

[0037] Step 44: Repeat steps 42-43 until convergence, and use the obtained posterior mean as the output weight of the hidden layer to complete the optimization process.

[0038] Beneficial effects of the present invention:

[0039] The present invention updates the output weights of the random incremental model through Bayesian reasoning, effectively suppressing the uncertainty in the process, and designs a global constraint mechanism that considers the impact of newly added nodes on the random incremental model, thereby improving the reliability and stability of random parameter learning, thereby improving the structural compactness and prediction accuracy of the model, and can better meet the requirements of lightweight soft measurement modeling of industrial processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:

[0041] Figure 1 is a flow chart of a specific embodiment of the present invention;

[0042] Figure 2 This is a soft measurement effect diagram of the grinding particle size during the grinding process in a specific embodiment of the present invention;

[0043] Figure 3 It is a scatter plot of the estimated grinding particle size during the grinding process in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0045] The present invention is further illustrated below in conjunction with specific embodiments. Those skilled in the art should understand that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention, and modifications to various equivalent forms of the present invention fall within the scope defined by the appended claims of this application.

[0046] The present invention provides a lightweight perception method for process indicators based on Bayesian random incremental learning, comprising the following steps:

[0047] Step S1: Obtain historical data of the industrial process, preprocess the data, and initialize the random increment model:

[0048] Provide a set of industrial process historical data, whose input is defined as X = {x1, x2, ..., x N},x i ∈R d , the output is defined as T = {t1, t2, ..., t N},t i∈R m , where N is the number of historical data, x i is the input of the i-th group of data, d is the number of process auxiliary variables, t i is the output of the i-th group of data, and m is the number of process indicators.

[0049] The initial random incremental model includes: input layer, hidden layer and output layer, where the number of nodes in the hidden layer is L-1. Other parameters of the random incremental model include: the pre-set maximum model size L max ; Regularization parameter λ; The selection range of hidden layer random learning parameters Υ, Υ: = {λ min :Δλ:λ max}, where λ min is the minimum setting value of the selected range, λ max is the maximum setting value of the selection range, Δλ is the interval value of the range change; error tolerance ε; learning parameter r; adjustment parameter K; the number of nodes T of the hidden layer generated by each parameter selection range max ; Residual e before , e before =TH L-1 (X)β before , T is the hidden layer node, is the output weight of the current random incremental model, H L-1 The Moore-Penrose generalized inverse of (X), H L-1 (X) is the hidden layer output matrix of L-1 hidden layer nodes:

[0050] H L-1 (X)=[h1(X),h2(X),...,h i (X),...,h L-1 (X)]

[0051] Among them, h i (X) = [g T (x1w i +b i ),g T (x2w i +b i ),...,g T (x N w i +b i )] T is the activation of the i-th hidden layer node, g(·) is the activation function, and bias are the input weights and bias in the activation function respectively.

[0052] Step S2: construct a global constraint mechanism through the Greville iteration method, and screen the generated hidden layer nodes based on the global constraint mechanism to form a candidate node pool;

[0053] When adding the Lth node, first in the interval [-λ min λ min ] d and [-λ min λ min ] Randomly generate T max nodes, then the T max The output of the hidden layer nodes can be expressed in matrix form as:

[0054]

[0055] The global constraint mechanism is constructed by Greville iteration method. Screen and store the nodes that meet the constraints in the candidate node pool. The expression of the global constraint is: L ≥0,

[0056] In the formula, e before is the residual of the random increment model;

[0057]

[0058] H L-1 Moore-Penrose generalized inverse of (X);

[0059]

[0060] T q Output data for the qth process indicator of T; s=1,2,...,T max .

[0061] Put the nodes that satisfy the global constraint mechanism into the candidate pool to form a new set of candidate nodes At the same time, a set of numerical values ​​is generated according to the global constraints z≤T max .

[0062] When no node in the generated hidden layer nodes satisfies the global constraint condition, the selection interval of the hidden layer nodes is adjusted. The specific method of adjusting the selection interval is as follows:

[0063] By adding a dynamic change value to the original selection interval, the adjusted selection interval is:

[0064] [-λ min +jΔλ λ min+jΔλ] d and [-λ min +jΔλ λ min +jΔλ]

[0065] Among them, [-λ min λ min ] d and [-λ min λ min ] is the original selection interval; jΔλ is the dynamic change value, j∈(1,M), M is the number of changes, Δλ is the interval value of the selected interval change.

[0066] At the same time, in order to more easily search for hidden layer nodes that meet the global constraints, the learning parameter r in the global constraint mechanism is scaled to be close to 1. The specific steps are as follows:

[0067] r:=r+τ

[0068] Wherein, τ is a random number in the interval (0.5*(1-r), 1-r). Repeat step S2 until the candidate node pool contains hidden layer nodes.

[0069] Step S3: Select the node parameters that contribute most to the accuracy of the random incremental model from the candidate node pool, calculate the output matrix of the node, and reconstruct the output matrix of the hidden layer:

[0070] Find the hidden layer node that contributes most to the accuracy of the random incremental model in the candidate node pool, that is, the node corresponding to the set Through this process, a set of optimal input weights can be obtained. and bias Thus, the output matrix of the hidden layer node is calculated:

[0071]

[0072] Finally, this output matrix is ​​combined with H L-1 (X) combined to form a new hidden layer output matrix:

[0073]

[0074] Step S4: Optimize the output weights of the hidden layer using the Bayesian statistical method:

[0075] Let β new Output weights for the model after adding the node. Assume that the data T are independent of each other and have a mean of H L (X)β new , with variance σ 2 The Gaussian white noise distribution of is, and its likelihood function can be expressed as:

[0076]

[0077] For β new , assuming that it obeys a mean of 0 and a variance of γ -1 The Gaussian distribution of I is expressed as:

[0078] p(β new |γ)=N(β new |0,γ 1 I)

[0079] Among them, γ is a hyperparameter, which is assumed to follow the gamma distribution, and σ 2 It also obeys the gamma distribution, and its expression is:

[0080] p(γ)=Gamma(γ|α1,α2)

[0081] p(σ 2 )=Gamma(σ -2 |α3,α4)

[0082] Among them, based on the flat prior, α1, α2, α3 and α4 are extremely small parameter given values, and the mathematical expression of the Gamma function is:

[0083] Gamma(α|a,b)=Γ(a) -1 b a α a-1 exp(-bα)

[0084]

[0085] According to Bayesian reasoning, β new The posterior estimate of follows a Gaussian distribution with mean m and covariance ∑, and its expression is:

[0086]

[0087] Based on the maximum likelihood estimation method, we can get γ and σ 2 Estimated value:

[0088]

[0089] in, λ i for The i-th eigenvalue of .

[0090] Based on the above, the specific process of step S4 is as follows:

[0091] Step S41: setting 4 minimum hyperparameters;

[0092] Step S42: Update the posterior mean and covariance, the update formula is:

[0093]

[0094] Where γ is a hyperparameter; σ 2 is the variance;

[0095] Step S43: Update hyperparameters and variance. The update formula is:

[0096]

[0097] Step S44: repeat steps S42-S43 until convergence, and use the obtained posterior mean as the output weight of the hidden layer to complete the optimization process;

[0098] At this time, the residual error of the model is L for:

[0099] e L =TH L β new

[0100] Step S5: Add hidden layer nodes through point-increment construction, and repeat steps 3-4 until the number of nodes L of the model reaches the preset maximum model size, or the accuracy of the model reaches the set error tolerance ε. At this time, the random incremental model is constructed, and the constructed random incremental model is used to perceive the process indicators.

[0101] The specific implementation mode of the present invention is described in detail below in conjunction with the specific prediction of the grinding particle size in the specific grinding process:

[0102] Step A1: Select 1500 sets of historical data of grinding measured on site from the historical data source of grinding process, where the input variables of each set of data include ball mill current c1, spiral classifier overflow concentration c2, spiral classifier current c3, ball mill feed c4 and ball mill inlet water c5, and the output variable is grinding particle size. Normalize all data and let the input data be X={x1,x2,...,x i ,...,x 1400}, The output data is T = {t1, t2, ..., t i ,...,t 1400}.

[0103] In the process of building a random incremental model, some important parameters need to be set, as follows: the pre-set maximum model size L max =100, regularization parameter λ = 2 -30, the selection range of the hidden layer random learning parameters Υ: = {1:0.1:10}, error tolerance ε = 0.001, adjustment parameter K = 100, the number of hidden nodes generated by each parameter selection range T max =200, the learning parameter r is in the following form:

[0104]

[0105] The initial number of hidden layer nodes in the model is L = 1, and the activation function of the hidden layer nodes is Then, the initial residual of the model is:

[0106] e before =T-H1(X)β before

[0107] Where H1(X)=h1(X)=[g T (x1w1+b1),g T (x2w1+b1),...,g T (x N w1+b1)] T , the input weight w1 and bias b1 are in the interval [-1 1] 5 and [-1 1] are randomly generated.

[0108] Step A2: When adding the Lth hidden layer node and performing the first search, first search in the interval [-1 1] 5 and [-1 1] randomly generate 200 sets of hidden layer learning parameters, that is, input weights w L and bias b L , and introduce it into the activation function. Therefore, the output of 200 sets of hidden layer nodes is obtained:

[0109]

[0110] Build a global constraint mechanism:

[0111]

[0112] in,

[0113]

[0114] Will The hidden layer node activations are substituted into the above constraint mechanism one by one, and the nodes that meet the constraint conditions are stored in the candidate node pool. At this time, the elements in the candidate node pool are represented as a set:

[0115]

[0116] When the first search fails to find any node, a second search is performed. When performing the jth search, the constraint degree is scaled, that is, r: = r + τ, where τ is a random number in the interval (0.5*(1-r), 1-r). Return to step A2 and in the interval [-(λ min +j0.1)(λ min +j0.1)] to reconstruct the candidate node pool of the Bayesian stochastic incremental model.

[0117] Step A3: In the candidate node set Find the largest value and use its corresponding random learning parameter, i.e., input weight w L and bias b L , defined as the optimal random learning parameter, that is, the optimal input weight and the optimal bias At this point, the optimal output of the newly added node can be expressed as:

[0118]

[0119] The best output matrix obtained is combined with H L-1 (X) combined to form a new hidden layer output matrix

[0120] Step A4: Use Bayesian reasoning to calculate the optimal output weight β new :

[0121] Let α1=1e-5, α2=1e-5, α3=1e-5 and α4=1e-5, let σ 2 =1 and γ=1.

[0122] Step 1: Utilization Update the posterior mean m and covariance Σ;

[0123] Step 2: Utilization Update γ and σ 2 ;

[0124] Repeat Step 1 and Step 2 until the algorithm converges, and define the m obtained at this time as the output weight β of the model new .

[0125] At this time, the e of the Bayesian random increment model L for:

[0126] e L =e L-1 -H L (X)β new .

[0127] When the number of hidden nodes in the constructed random incremental model reaches the preset maximum model size of 100, or the residual of the model is within the set acceptable error range of 0.001, the hidden nodes will be stopped from being added, and finally a lightweight perception model of process indicators based on Bayesian random incremental learning for grinding particle size will be formed. Otherwise, it will return to step A1 to continue to build the model.

[0128] Step A5: Use the test set to evaluate the performance of the trained Bayesian random incremental learning grinding particle size lightweight perception model, that is, use the model to predict the grinding particle size in the test data.

[0129] In order to demonstrate the performance advantage of the present invention, a test effect diagram of the random increment model with the number of hidden layer nodes L=50 is given. Figure 2 As shown in Figure 2, the estimated value of the grinding particle size is very close to the actual value. Figure 3 As shown, it can be seen that the predicted value of the grinding particle size is well matched with the true value, and the error is significantly smaller, and the accuracy is higher. The feedback information from the two figures can prove that the present invention has high modeling accuracy and good generalization ability, indicating that the model can achieve excellent prediction effect in industrial process applications.

[0130] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A lightweight perception method for process indicators based on Bayesian random incremental learning, characterized in that: The steps include: Step 1: Obtain historical data of industrial processes and preprocess the data; Step 2: Construct a global constraint mechanism through the Greville iteration method, and screen the generated hidden layer nodes based on the global constraint mechanism to form a candidate node pool; Step 3: Select the node parameters that contribute most to the accuracy of the random incremental model from the candidate node pool, calculate the output matrix of the node, and reconstruct the output matrix of the hidden layer; Step 4: Optimize the output weights of the hidden layer using Bayesian statistical methods; Step 5: Add hidden layer nodes by point-increment construction, and repeat steps 2-4 until the random incremental model is built; Step 6: Use the constructed random incremental model to perceive the process indicators.

2. The lightweight perception method of process indicators based on Bayesian random incremental learning according to claim 1 is characterized in that: In step 2, the expression of the constructed global constraint is: L ≥0, in, In the formula, e before is the residual of the random increment model; H L-1 Moore-Penrose generalized inverse of (X); T q Output data for the qth process indicator of T; s=1,2,...,T max .

3. The lightweight perception method of process indicators based on Bayesian random incremental learning as described in claim 1 or 2, characterized in that: In the step 2, the process of screening the generated hidden layer nodes by the global constraint mechanism also includes: When no node in the generated hidden layer nodes satisfies the global constraint condition, the selection interval of the hidden layer nodes is adjusted. The specific method of adjusting the selection interval is as follows: By adding a dynamic change value to the original selection interval, the adjusted selection interval is: [-λ min +jΔλ λ min +jΔλ] d and [-λ min +jΔλ λ min +jΔλ] Among them, [-λ min λ min ] d and [-λ min λ min ] is the original selection interval; jΔλ is the dynamic change value, j∈(1,M), M is the number of changes, Δλ is the interval value of the selected interval change.

4. The lightweight perception method of process indicators based on Bayesian random incremental learning according to claim 1 or 2, characterized in that: The step 2 also includes scaling the learning parameters in the global constraint mechanism so that the learning parameters approach 1.

5. The lightweight perception method of process indicators based on Bayesian random incremental learning according to claim 1 is characterized in that: In step 3, the output matrix of the reconstructed hidden layer is: In the formula, is the output matrix of the node that contributes most to the accuracy of the random incremental model, H L-1 (X) is the hidden layer output matrix of L-1 nodes.

6. The lightweight perception method of process indicators based on Bayesian random incremental learning according to claim 1 is characterized in that: The specific process of step 4 is as follows: Step 41: Set 4 minimal hyperparameters; Step 42: Update the posterior mean and covariance. The update formula is: Where γ is a hyperparameter; σ 2 is the variance; Step 43: Update the hyperparameters and variance. The update formula is: Step 44: Repeat steps 42-43 until convergence, and use the obtained posterior mean as the output weight of the hidden layer to complete the optimization process.

Images