Complex industrial process soft measurement modeling method based on double-layer dynamic random configuration width learning

Abstract

The invention designs a complex industrial process soft measurement modeling method based on double-layer dynamic random configuration width learning, and aims to better extract dynamic characteristics from industrial data and improve the performance of a soft sensor. According to the method, a double-layer dynamic feature analysis model is established in a mapping feature layer, a slow feature analysis method is used in the first layer to extract intrinsic essential features in industrial data, and the influence of noise in the data on the model is reduced; in the second layer, dynamic features in data are further extracted from features extracted by a slow feature analysis method by adopting a deterministic jump cyclic reserve pool, and meanwhile, nodes of an enhanced feature layer are obtained by adopting a random configuration method in the enhanced feature layer, so that the quality of the enhanced nodes is improved. The method designed by the invention not only has good dynamic feature extraction capability, but also improves the prediction performance of the soft sensor.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial process soft sensing, and relates to a soft sensing modeling method for complex industrial processes based on a dual-layer dynamic stochastic configuration broad learning system (abbreviation: DDSCBLS). Background Art

[0002] With the gradual expansion of modern industrial production scale, the process flow has become gradually complex, and the requirements for product quality have been gradually improved. However, due to the lack of sensors or the harsh industrial field environment, it is difficult to measure some key quality variables online. To solve the problem of difficult measurement of key quality variables and improve the real-time performance and effectiveness of key quality measurement, soft sensing technology has emerged. Soft sensing technology establishes a system model based on the mathematical relationship between easily measurable auxiliary variables and difficult-to-measure quality variables, and indirectly realizes the real-time and effective estimation of quality variables. It has the advantages of low development cost, flexible configuration, and rapid response. Soft sensing models now play an increasingly important role in many fields such as petrochemical, biochemical, metallurgical, and pharmaceutical industries.

[0003] Broad learning system (BLS) is a simple and efficient neural network structure. Due to its advantages such as simple structure, fast operation speed, and independence from deep structure, it has been successfully applied to the field of soft sensing modeling. However, although the weights of the nodes in the mapping feature layer of the BLS algorithm are fine-tuned through a sparse autoencoder, the enhanced nodes still use random parameters that affect performance, resulting in low performance of the BLS model. Stochastic configuration BLS uses a stochastic configuration method to obtain the nodes of the enhanced feature layer on the basis of BLS, improving the performance of the BLS algorithm. However, stochastic configuration BLS does not consider the fact that industrial data often has certain dynamic characteristics. To address the above problems, a soft sensing modeling method for complex industrial processes based on dual-layer dynamic stochastic configuration broad learning is proposed. A dual-layer dynamic feature analysis model is established in the mapping feature layer to extract the dynamic features in the data, improving the prediction performance of the soft sensor, and a stochastic configuration method is used to obtain the nodes of the enhanced feature layer, improving the quality of the enhanced nodes. Summary of the Invention

[0004] In view of the problem that random configuration BLS cannot extract the dynamic features in industrial data, the present invention proposes a soft sensor modeling method for complex industrial processes based on double-layer dynamic random configuration width learning. This method establishes a double-layer dynamic feature analysis model in the mapping feature layer. The first layer uses the slow feature analysis method to extract the inherent essential features in industrial data, and the second layer uses a deterministic jump recurrent reservoir to further extract the dynamic features in the data from the features extracted by the slow feature analysis method, improving the prediction performance of the soft sensor. And a random configuration method is used to obtain the nodes of the enhanced feature layer, improving the quality of the enhanced nodes.

[0005] To achieve the above object, the present invention provides a soft sensor modeling method for complex industrial processes based on double-layer dynamic random configuration width learning, comprising the following steps:

[0006] (1) Collect the training data set D s ={X s , Y s} and the test data set D v ={X v , Y v} from the actual industrial process, where D s has N s samples for establishing the soft sensor model, and D v has N v samples for testing the model generalization ability. X s , X v respectively represent the input variable data in D s , D v , and Y s , Y v respectively represent the output variable data in D s , D v ; preprocess the training data set D s and the test data set D v to obtain the preprocessed training data set and the test data set

[0007] (2) Establish a double-layer dynamic feature analysis model to extract the primary mapping features of the width learning network. In the first layer, the input variables in the training data set are subjected to slow feature analysis to obtain the slow feature matrix S and the linear transformation matrix W. In the second layer, the dynamic feature matrix Z is extracted from the slow feature matrix S through a deterministic jump recurrent reservoir network;

[0008] (3) Construct an enhanced feature layer with the primary mapping feature Z as the input. The weight matrix W h and the bias matrix β hObtained by using the random configuration method, and the output weight matrix W is obtained simultaneously after the random configuration is completed. out ;

[0009] (4) Use the preprocessed test sample set as the input data of the soft sensor model, calculate the corresponding output prediction, and analyze the prediction effect of the double-layer dynamic random configuration BLS model.

[0010] In step (1), the training data set D s and the test data set D v are preprocessed respectively through formulas (1)-(4), and the expressions of formulas (1)-(4) are:

[0011]

[0012] Among them, X s,min is the minimum value of the input variable X s in the training data set, and X s,max is the maximum value of the input variable X s in the training data set; Y s,min is the minimum value of the output variable Y s in the training data set, and Y s,max is the maximum value of the output variable Y s in the training data set.

[0013] In step (2), each sample in the training data set is subjected to slow feature analysis through formula (5) to obtain the slow feature matrix S, and the expression of formula (5) is:

[0014]

[0015] Among them, W = [w1, w2,..., w l T represents the linear transformation matrix, and w l represents the l-th vector in the linear transformation matrix;

[0016] W can be solved according to formula (6), and the expression of formula (6) is:

[0017] Aw j = λ j Bw j (6)

[0018] Among them, the matrix A represents the covariance matrix of the derivative of and the matrix B represents the covariance matrix of;

[0019] ​Then, dynamic features Z are extracted from the slow feature matrix S through formulas (7) and (8), and the expressions of formulas (7) and (8) are as follows:

[0020] z(t) = Φ(aW e s(t)+β e +(1 - a)W x z(t - 1)) (7)

[0021]

[0022] where a is a hyperparameter that balances s(t) and z(t - 1), s(t) represents the slow feature at time t in S, z(t - 1) and z(t) represent the deterministic jump recurrent reservoir states at time t - 1 and time t respectively, the matrix W e , β e , W x are the input weight matrix, bias matrix, and state weight matrix of the deterministic jump recurrent reservoir nodes respectively, all initialized with fixed parameters.

[0023] In step (iii), first, the enhanced layer feature nodes h j are configured according to the inequality constraint conditions of formula (9), and the expression of formula (9) is as follows:

[0024]

[0025] where represents the model output with j - 1 enhanced nodes. For b g ∈R + , there is 0 < ‖h j ‖ < b g , and h j is the output of the jth enhanced node;

[0026] δ j is calculated according to formula (10), and the expression of formula (10) is as follows:

[0027] δ j = (1 - r - u j )‖e j-1 ‖ 2 (10)

[0028] where 0 < r < 1, {u j} is a non - negative real - number sequence and satisfies and 0 < u j ≤1 - r;

[0029] The output H of the enhanced layer feature nodes is calculated through equations (11) and (12), and the expressions of equations (11) and (12) are as follows:

[0030] h j = ξ j (ZW hj + β hj ), j = 1, 2, …, m (11)

[0031]

[0032] where W hj , β hj are the weight matrix and bias matrix of the j-th boosting node respectively; ξ j is the activation function of the j-th boosting node;

[0033] After updating the boosting nodes according to the inequality constraint conditions, the final output is calculated by formula (13), and the expression of formula (13) is:

[0034]

[0035] where W out is the output matrix;

[0036] W out is calculated by formula (14), and the expression of formula (14) is:

[0037]

[0038] where C is the regularization coefficient.

[0039] In step (iv), for the preprocessed new test data set the slow feature matrix S of the test data set is obtained by calculating through formula (15) v , and the expression of formula (15) is:

[0040]

[0041] The dynamic feature matrix is obtained by calculating through formulas (16) and (17), and the expressions of formulas (16) and (17) are:

[0042] z v (t) = Φ(aW e s v (t) + β e + (1 - a)W x z v (t - 1)) (16)

[0043]

[0044] The output H of the boosting node is calculated through formulas (18) and (19) v , and the expressions of formulas (18) and (19) are:

[0045] h v,j = ξ j (Z v W hj + β hj ), j = 1, 2, …, m (18)

[0046]

[0047] The predicted output of the final test set is calculated by formula (20), and the expression of formula (20) is:

[0048]

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] The complex industrial process soft sensor modeling method based on double-layer dynamic random configuration width learning provided by the present invention establishes a double-layer dynamic feature analysis model in the mapping feature layer. The first layer uses the slow feature analysis method to extract the inherent essential features in industrial data, and the second layer uses the deterministic jump recurrent reservoir to further extract the dynamic features in the data from the features extracted by the slow feature analysis method, improving the prediction performance of the soft sensor. And the method of random configuration is adopted to obtain the nodes of the enhanced feature layer, improving the quality of the enhanced nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a structural diagram of the complex industrial process soft sensor modeling method based on double-layer dynamic random configuration BLS of the present invention;

[0052] Figure 2a is a comparison schematic diagram of the predicted value and the true value curves of the test sample set by using the traditional BLS method in the embodiment of the present invention;

[0053] Figure 2b is a comparison schematic diagram of the predicted value and the true value curves of the test sample set by using the random configuration BLS method in the embodiment of the present invention;

[0054] Figure 2c is a comparison schematic diagram of the predicted value and the true value curves of the test sample set by using the complex industrial process soft sensor modeling method based on double-layer dynamic random configuration BLS of the present invention in the embodiment of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] Hereinafter, the present invention will be specifically described by way of exemplary embodiments. However, it should be understood that, without further narration, the elements, structures and features in one embodiment can also be beneficially combined into other embodiments.

[0056] See Figure 1, the present invention discloses a soft sensor modeling method for complex industrial processes based on double-layer dynamic random configuration width learning, comprising the following steps:

[0057] (1) Collect the training data set D s ={X s , Y s} and the test data set D v ={X v , Y v}, where D s has N s samples for establishing the soft sensor model, and D v has N v samples for testing the model generalization ability. X s , X v respectively represent the input variable data in D s , D v , and Y s , Y v respectively represent the output variable data in D s , D v . Preprocess the training data set D s and the test data set D v to obtain the preprocessed training data set and the test data set . The specific steps are as follows:

[0058] Preprocess the training data set D s and the test data set D v respectively through formulas (21)-(24). The expressions of formulas (21)-(24) are:

[0059]

[0060] where X s,min is the minimum value of the input variable X s in the training data set, and X s,max is the maximum value of the input variable X s in the training data set; Y s,min is the minimum value of the output variable Y s in the training data set, and Y s,max is the maximum value of the output variable Y s in the training data set.

[0061] (2) Establish a double-layer dynamic feature parsing model to extract the primary mapping features of the width learning network. The first layer maps the input variables in the training data set Slow feature analysis is performed to obtain the slow feature matrix S and the linear transformation matrix W. In the second layer, the dynamic feature matrix Z is extracted from the slow feature matrix S through a deterministic jump recurrent reservoir network. The specific steps are as follows:

[0062] Each sample in the training dataset is subjected to slow feature analysis through formula (25) to obtain the slow feature matrix S. The expression of formula (25) is:

[0063]

[0064] where W = [w1, w2, …, w l T represents the linear transformation matrix, and w l represents the l-th vector in the linear transformation matrix;

[0065] W can be obtained by solving according to formula (26). The expression of formula (26) is:

[0066] Aw j = λ j Bw j (26)

[0067] where the matrix A represents the covariance matrix of the derivative of and the matrix B represents the covariance matrix of ;

[0068] Then, the dynamic feature Z is extracted from the slow feature matrix S through formulas (27) and (28). The expressions of formulas (27) and (28) are:

[0069] z(t) = Φ(aW e s(t) + β e + (1 - a)W x z(t - 1)) (27)

[0070]

[0071] where a is a hyperparameter that balances s(t) and z(t - 1), s(t) represents the slow feature at time t in S, z(t - 1) and z(t) represent the deterministic jump recurrent reservoir states at time t - 1 and time t respectively, the matrix W e , β e , W x are the input weight matrix, bias matrix, and state weight matrix of the deterministic jump recurrent reservoir nodes respectively, all initialized with fixed parameters.

[0072] (3) Using the primary mapped feature Z as the input, an enhanced feature layer is constructed. The weight matrix W h and the bias matrix β​​h Obtained by using the random configuration method, and the output weight matrix W is obtained simultaneously after the random configuration is completed out ; The specific steps are as follows:

[0073] First, configure the enhanced layer feature node h according to the inequality constraint condition of formula (29) j , and the expression of formula (29) is:

[0074]

[0075] Among them, represents the model output with j - 1 enhanced nodes. For b g ∈R + , there is 0 < ‖h j ‖ < b g , and h j is the output of the jth enhanced node;

[0076] δ j is calculated according to formula (30), and the expression of formula (30) is:

[0077] δ j =(1 - r - u j )‖e j-1 ‖ 2 (30)

[0078] Among them, 0 < r < 1, {u j} is a non - negative real number sequence and satisfies and 0 < u j ≤1 - r;

[0079] Calculate the output H of the enhanced layer feature node through equations (31)(32), and the expressions of equations (31)(32) are:

[0080] h j =ξ j (ZW hj +β hj ), j = 1, 2, …, m (31)

[0081]

[0082] Among them, W hj , β hj are the weight matrix and bias matrix of the jth enhanced node respectively; ξ j is the activation function of the jth enhanced node;

[0083] After updating the enhanced node according to the inequality constraint condition, the final output is calculated by formula (33), and the expression of formula (33) is:

[0084]

[0085] Among them, W out is the output matrix;

[0086] W out is calculated by formula (34), and the expression of formula (34) is:

[0087]

[0088] Among them, C is the regularization coefficient.

[0089] (IV) Use the preprocessed test sample set as the input data of the soft measurement model, calculate the corresponding output prediction, and analyze the prediction effect of the double-layer dynamic random configuration BLS model; the specific steps are as follows:

[0090] For the preprocessed new test data set calculate the slow feature matrix S of the test data set through formula (35) v , and the expression of formula (35) is:

[0091]

[0092] Calculate the dynamic feature matrix through formulas (36) and (37), and the expressions of formulas (36) and (37) are:

[0093] z v (t) = Φ(aW e s v (t) + β e + (1 - a)W x z v (t - 1)) (36)

[0094]

[0095] Calculate the enhanced node output H through formulas (38) and (39) v , and the expressions of formulas (38) and (39) are;

[0096] h v,j = ξ j (Z v W hj + β hj ), j = 1, 2,..., m (38)

[0097]

[0098] Finally, the predicted output of the test set is calculated by formula (40), and the expression of formula (40) is:

[0099]

[0100] The above soft sensor method of the present invention establishes a two - layer dynamic feature analysis model in the mapped feature layer. The first layer uses the slow feature analysis method to extract the intrinsic essential features in industrial data, reducing the impact of noise in the data on the model performance. The second layer uses a deterministic jump recurrent reservoir to further extract the dynamic features in the data from the features extracted by the slow feature analysis method, improving the prediction performance of the soft sensor. And a random configuration method is adopted to obtain the nodes of the enhanced feature layer, improving the quality of the enhanced nodes.

[0101] In order to more clearly illustrate the beneficial effects of the above soft sensor method of the present invention, the following further describes the above soft sensor method of the present invention in combination with specific embodiments.

[0102] Embodiment:

[0103] The sulfur recovery unit (SRU) is an important process unit in a refinery. It can remove acidic gases in environmental pollutants and prevent their emission into the atmosphere. At the same time, it can also recover elemental sulfur as a valuable by - product. The input stream of the SRU contains two acidic gases, monoethanolamine (MEA) and sour water stripping gas (GAS). The tail gas usually contains a certain amount of H2S and SO2, and their content is an important quality variable, but it is difficult to measure online. Therefore, it is necessary to use soft sensor technology for measurement.

[0104] Select 5 conventional measurement variables as prediction variables, and use the SO2 concentration as the output variable for simulation. The description of relevant variables is shown in Table 1.

[0105] Table 1

[0106] Variable label Variable description 1 MEA gas flow rate 2 First air flow 3 Second air flow 4 Gas flow in SWS area 5 Air flow in SWS area 6 <![CDATA[SO2 concentration]]>

[0107] In this embodiment, three methods, namely the traditional BLS method, the random configuration BLS method, and the DDSCBLS method of the present invention, are used for simulation comparison. The root mean square error (RMSE) is used as the performance index to evaluate different soft sensor methods. Specifically, RMSE is defined as the square root of the ratio of the sum of the squares of the deviations between the predicted values and the true values of the test data in the target domain to the number of samples. In the DDSCRBLS method of the present invention, the model structure is that the mapped layer reservoir contains 500 nodes, and the enhanced layer has 20 nodes. The training data is 3000 training samples, and the test data is 2000 new test samples.

[0108] For the simulation comparison using the traditional BLS method, the random configuration BLS method, and the DDSCBLS method of the present invention, the curve comparison diagram of the predicted values and the true values of the new samples in the test set is shown in Figure 2a 、 Figure 2b andFigure 2c See Figure 2a , since the sulfur recovery industrial process has certain dynamic characteristics, and the traditional BLS method is a static algorithm with poor prediction effect, and its average root mean square error is 0.0507. See Figure 2b , although the randomly configured BLS improves the quality of enhanced nodes through inequality constraints and the prediction effect is improved, it still cannot extract dynamic features, and its average root mean square error is 0.0497. The DDSCBLS provided by the present invention will effectively extract dynamic features through slow feature analysis of the mapping layer and the reservoir structure. See Figure 2c , the fitting effect of the curve is greatly improved, and the average root mean square error is reduced to 0.0403. Therefore, the DDSCBLS method provided by the present invention significantly improves the prediction performance for new test sample data.

[0109] Table 2 shows the root mean square errors of the traditional BLS method, the randomly configured BLS method, and the DDSCBLS method of the present invention for the new sample data of the test set.

[0110] Table 2

[0111] Model Root mean square error BLS 0.0507 Randomly configured BLS 0.0497 DDSCBLS 0.0403

[0112] As can be seen from Table 2, the DDSCBLS method provided by the present invention generally obtains good prediction results and has the best prediction performance.

[0113] Based on the above analysis, the DDSCBLS method provided by the present invention establishes a double-layer dynamic feature analysis model in the mapping feature layer, which can effectively extract dynamic features in the data, improve the prediction performance of the soft sensor, and uses the randomly configured method to obtain the nodes in the enhanced feature layer, improving the quality of the enhanced nodes. It is proved that its prediction effect is better than that of the traditional BLS method and the randomly configured BLS method.

[0114] The above-mentioned embodiments are only used to conveniently illustrate the present invention and are not limitations on the protection scope of the present invention. Within the scope of the technical solution described in the present invention, all kinds of simple deformations and modifications made by those skilled in the art should be included in the scope of the above patent application.

Claims

1. A soft sensor modeling method for complex industrial processes based on a double-layer dynamic random configuration width learning network, characterized in that, It includes the following steps: ( (1) Collect the training dataset D from the actual industrial process s ={X s ,Y s} and the test dataset D v ={X v ,Y v}, where D s has N s samples for establishing the soft sensor model, and D v has N v samples for testing the model generalization ability. X s , X v represent the input variable data in D s , D v respectively, and Y s , Y v represent the output variable data in D s , D v respectively. Preprocess the training dataset D s and the test dataset D v to obtain the preprocessed training dataset and the test dataset (2) Establish a double-layer dynamic feature parsing model to extract the primary mapping features of the width learning network. The first layer performs slow feature analysis on the input variables in the training data set to obtain the slow feature matrix S and the linear transformation matrix W. The second layer extracts the dynamic feature matrix Z from the slow feature matrix S through a deterministic jump recurrent reservoir network; (3) Using the primary mapping feature Z as the input to construct an enhanced feature layer, the weight matrix W of the enhanced feature layer h and the bias matrix β h are obtained by using the random configuration method. After the random configuration is completed, the output weight matrix W is obtained simultaneously out ; (4) Use the preprocessed test sample set as the input data of the soft sensor model, calculate the corresponding output prediction, and analyze the prediction effect of the double-layer dynamic random configuration BLS model.

2. A soft sensor modeling method for complex industrial processes based on double-layer dynamic random configuration width learning as claimed in claim 1, wherein In the step (i), the training data set D and the test data set D are preprocessed respectively by the formulas (1)-(4), and the expressions of the formulas (1)-(4) are as follows: s and the test data set D v are preprocessed, and the expressions of the formulas (1)-(4) are: Among them, X s,min is the minimum value of the input variable X in the training dataset s , and X s,max is the maximum value of the input variable X in the training dataset s ; Y s,min is the minimum value of the output variable Y in the training dataset s , and Y s,max is the maximum value of the output variable Y in the training dataset s .

3. A soft sensor modeling method for complex industrial processes based on double-layer dynamic random configuration width learning, as described in claim 1, wherein In the step (ii), each sample in the training data set is subjected to slow feature analysis through the formula (5) to obtain a slow feature matrix S. The expression of the formula (5) is as follows: Among them, W = [w1, w2, …, w l T represents a linear transformation matrix, and w l represents the l-th vector in the linear transformation matrix;​ W can be solved according to formula (6), and the expression of formula (6) is: Aw j = λ j Bw j (6) where the matrix A represents the covariance matrix of the derivative and the matrix B represents the covariance matrix; Then, the dynamic feature Z is extracted from the slow feature matrix S through formulas (7) and (8), and the expressions of formulas (7) and (8) are: z(t) = Φ(aW e s(t) + β e + (1 - a)W x z(t - 1)) (7) Among them, a is a hyperparameter that balances s(t) and z(t - 1), s(t) represents the slow feature at time t in S, z(t - 1) and z(t) represent the deterministic jump recurrent reservoir states at time t - 1 and time t respectively, and the matrix W e , β e , W x are the input weight matrix, bias matrix, and state weight matrix of the deterministic jump recurrent reservoir nodes respectively, all initialized with fixed parameters.

4. A soft sensor modeling method for complex industrial processes based on double-layer dynamic random configuration width learning, as described in claim 1, wherein In step (iii), first, the enhanced layer feature node h is configured according to the inequality constraint condition of formula (9). j The expression of formula (9) is as follows: Among them, represents the model output with j - 1 boosting nodes. For b g ∈R + , there is 0 < ‖h j ‖ < b g , and h j is the output of the j-th boosting node; δ j Calculated according to formula (10), the expression of formula (10) is: δ j =(1 - r - u j )‖e j-1 ‖ 2 (10) where \(0 < r < 1\), \(\{u j \}\) is a sequence of non - negative real numbers and satisfies and \(0 < u j \leq1 - r\); The output H of the enhanced layer feature nodes is calculated through formulas (11) and (12), and the expressions of formulas (11) and (12) are: h j = ξ j (ZW hj + β hj ), j = 1, 2, …, m (11) Among them, W hj , β hj are respectively the weight matrix and bias matrix of the j-th boosting node; ξ j is the activation function of the j-th boosting node; After updating the enhanced nodes according to the inequality constraint conditions, the final output is calculated by formula (13), and the expression of formula (13) is: Among them, W out is the output matrix; W out Calculated by formula (14), the expression of formula (14) is: Among them, C is the regularization coefficient.

5. A soft sensor modeling method for complex industrial processes based on double-layer dynamic random configuration width learning, characterized in that In step (iv), for the preprocessed new test data set calculate the slow feature matrix S of the test data set through formula (15) v , and the expression of formula (15) is: The dynamic feature matrix is calculated through formulas (16) and (17), and the expressions of formulas (16) and (17) are: z v (t) = Φ(aW e s v (t) + β e + (1 - a)W x z v (t - 1)) (16) Calculate the enhanced node output H through formulas (18) and (19). v , and the expressions of formulas (18) and (19) are as follows; h v,j = ξ j (Z v W hj + β hj ), j = 1, 2, …, m (18) Finally, the predicted output of the test set is calculated by formula (20), and the expression of formula (20) is:

Images