Soft Sensing Method for Debutanizer Based on Deep Evolutionary Fuzzy Rule System
Through the deep evolution fuzzy rule system combining representation learning and prediction learning networks, a soft measurement model with strong adaptability in industrial process is constructed, which solves the shortcomings of traditional deep learning models in time-varying characteristics and improves prediction accuracy and stability.
Patent Information
- Application Number
- CN202310536760.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-12
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2043-05-12
AI Technical Summary
Traditional deep learning models perform poorly in dealing with time-varying characteristics in industrial processes, and the computational cost of online adjustment of deep learning models is high, making it difficult to optimize network structure in real time to follow the characteristics of time-varying process.
A deep evolution fuzzy rule system is adopted to build a deep evolution fuzzy rule system by combining the representation learning network and the predictive learning network to realize online prediction and modeling of nonlinear and time-varying processes. The system extracts features through stacking autoencoders and adapts to changes in industrial processes through the addition, deletion and parameter update of fuzzy rules.
It improves the prediction accuracy and stability of soft measurement modeling of industrial processes, has good adaptability, and can effectively deal with complex nonlinear and time-varying problems.
Smart Images

Figure CN116611525B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to an industrial process soft measurement method based on a deep evolution fuzzy rule system, and belongs to the field of industrial process control. Background Art
[0002] With the development of information technology and the increase in customer demand, industrial processes are becoming more and more complex, and a large number of key variables need to be accurately measured. At the same time, in order to monitor process performance and achieve the goals of green production, energy saving and profit maximization, key quality indicators must be obtained as quickly as possible. However, the harsh measurement environment, expensive analyzers, and large measurement delays make most of these variables difficult to measure online. Therefore, soft sensing technology estimates the difficult-to-measure quality variables by establishing mathematical models, thereby providing real-time information for process monitoring, optimization and control.
[0003] In recent years, data-driven soft sensor models have developed into a mainstream modeling method. Traditional data-driven soft sensor modeling methods include statistical inference techniques and machine learning techniques such as principal component regression, partial least squares, support vector regression, etc. Although these methods have been quite mature in many industrial processes, they do not perform well in dealing with complex nonlinear problems. In the past two decades, deep learning technology has been favored by scholars and technical developers because it is good at dealing with the nonlinearity of industrial processes and other more complex data patterns. Typical soft sensor deep learning networks can be divided into four categories: autoencoders or stacked autoencoders, restricted Boltzmann machines, convolutional neural networks, and recursive neural networks. Deep learning models complete prediction tasks by establishing regression models between abstract features and quality variables. However, due to problems such as operating condition changes, equipment aging, and catalyst deactivation caused by operating conditions, most industrial processes often exhibit strong time-varying process characteristics. This time-varying characteristic limits the prediction performance of the above traditional non-adaptive deep learning models. Since the computational cost of online adjustment of deep learning models is very high, it is difficult to optimize the structure of deep learning networks in real time to quickly follow the time-varying process characteristics. Summary of the invention
[0004] The present invention provides an industrial process soft measurement method based on a deep evolution fuzzy rule system, which is used to realize industrial process soft measurement by combining a representation learning network and a prediction learning network.
[0005] The technical solution of the present invention is: an industrial process soft measurement method based on a deep evolution fuzzy rule system, comprising: collecting historical industrial process data, constructing a pre-training sample set D pretrn ; The pre-training sample set D pretrn Includes pre-trained feature set X pretrn ={x1 ,x 2 ,…,xk,…,x N}、True label set y pretrn ={y 1 ,y 2 ,...,y k ,...,y N}, xk represents the kth pre-trained feature, each pre-trained feature consists of multiple auxiliary variables, y k represents the kth true label, N represents the total number of pre-training samples; using the pre-training feature set X pretrn Pre-train the representation learning network to obtain the pre-trained representation learning network; the pre-trained feature set X pretrn Re-input into the pre-trained representation learning network to obtain the initial feature sample set D init ; Based on the initial feature sample set D init , the true label y pretrn Pre-train the prediction learning network to obtain a deep evolutionary fuzzy rule system; perform online prediction on the samples to be predicted based on the pre-trained representation learning network and the deep evolutionary fuzzy rule system, and update the deep evolutionary fuzzy system when the labels of the samples to be tested arrive.
[0006] The representation learning network adopts a stacked autoencoder SAE, and the stacked autoencoder SAE is trained by L autoencoders AE.
[0007] The initial feature sample set D init , the true label y pretrn Pre-train the prediction learning network to obtain a deep evolutionary fuzzy rule system, including: using the initial feature sample set D init The first feature sample h 1 Initialize the deep evolutionary fuzzy rule system; and use the remaining feature samples D remain ={h 2 ,h 3 ,…,h N} is used to update the deep evolutionary fuzzy rule system antecedent, the true label y pretrn It is used to update the subsequent parts of the deep evolutionary fuzzy rule system, and the evolution process is carried out sample by sample, repeated m times to obtain the deep evolutionary fuzzy rule system.
[0008] The deep evolutionary fuzzy rule system consists of a series of fuzzy rules, each of which corresponds to a linear model:
[0009]
[0010] Among them, R i represents the i-th fuzzy rule, p i is the antecedent of the i-th fuzzy rule, ai is the least squares parameter of the consequent of the i-th fuzzy rule, h k is the kth pre-trained latent feature extracted by the learning network, h k =[h 1 ,h 2 ,..,h w ], w is the dimension of latent features, ~ is close to / similar to, so the prediction output of the entire deep evolutionary fuzzy rule system is for:
[0011]
[0012] Among them, λ i is the activation strength of the i-th fuzzy rule, is the predicted output value of the i-th fuzzy rule.
[0013] The activation strength calculation formula of the fuzzy rule is as follows:
[0014]
[0015] Among them, D i (h k ) means h k and the local density between the i-th fuzzy rule, and is a temporary variable, S i Because p i The cluster C i The cardinality, H i is the local norm of the i-th fuzzy rule.
[0016] The pre-trained prediction learning network includes:
[0017] S4.1. Extract x using the pre-trained representation learning network k h k ,k=1h k is the initial feature sample set D init The first feature sample h 1 , and then initialize the global meta parameters:
[0018] μ←h 1 ; χ←∥h 1 ∥ 2
[0019] Among them, μ is the global mean, χ is the square of its binary norm, representing the global norm; ← represents the update assignment operation, the parameter on the left of the arrow is the new update parameter value, and the parameter on the right is the current parameter value;
[0020] S4.2. Initialize the antecedent and consequent parameters of the first fuzzy rule, r = k = 1:
[0021] p r ←h k ; a r ←0 w+1 ; Θ r ←Ω 0 I (w+1)×(w+1)
[0022] Among them, Θ r is the covariance matrix of the rth fuzzy rule, p r is the antecedent of the rth fuzzy rule, a r is the least squares parameter of the consequent of the rth fuzzy rule, I (w+1)×(w+1) is the identity matrix of (w+1)×(w+1), 0 w+1 is a (w+1)-dimensional all-zero vector; Ω 0 is the initialization constant of the covariance matrix, used in the standard recursive least squares algorithm;
[0023] S4.3. Initialize local meta parameters:
[0024] C r ←{h k}; S r ←1;H r ←||h k || 2
[0025] Each fuzzy rule can form a cluster, represented by p r The cluster C r It can be approximated as a tessellated Voronoi distribution, S r C r The cardinality, H r is the local norm of the rth fuzzy rule; k←k+1;
[0026] S4.4, when k>1, the fuzzy rule addition, fuzzy rule deletion, and consequent parameter update processes are performed in sequence until the instance sample h k =h N ,h N D remain The last feature sample, at this time continue to repeat the whole process until it is completed m times to obtain the evolutionary fuzzy system after pre-training.
[0027] When k>1, the fuzzy rule adding, fuzzy rule deleting, and subsequent parameter updating processes are performed in sequence, including:
[0028] S4.4.1. Global meta-parameter update:
[0029]
[0030] S4.4.2. Calculate the global density D(z):
[0031]
[0032] S4.4.3. Determine whether to add fuzzy rules. If the first check condition is met, the rule is added; if the second check condition is met, the rule is deleted; if the third check condition is met, the rule consequence parameter is updated.
[0033] The inspection condition 1:
[0034]
[0035] Among them, D i (h k ) means h k The local density between the i-th fuzzy rule, D 0 is a constant;
[0036] If the check condition is met, a new fuzzy rule is added to the deep evolutionary fuzzy rule system, and the new antecedent and subsequent parameters are set to:
[0037]
[0038] Among them, the antecedent p r is initialized to a new h k , the posterior least squares parameter a r Initialized to the least squares parameter a of all other consequents in the deep evolutionary fuzzy rule system i average;
[0039] If the check condition 1 is not satisfied, then the sample h k If it is not initialized as a new rule, use h k Update the most recent cluster
[0040]
[0041] The nearest cluster is h k To the fuzzy rule antecedent p j The cluster corresponding to the minimum Euclidean distance value is:
[0042]
[0043] The parameters of the least squares antecedent and consequent of other rules remain unchanged.
[0044] The second inspection condition:
[0045] if(η i,k <η0 )then(R i ,C i are removed)
[0046] Among them, η i,k The calculation is as follows:
[0047]
[0048] Where i = 1, 2, ..., r, η i,k is the i-th fuzzy rule R i The corresponding availability of the kth sample, η 0 is an adjustable constant parameter, I i is the time when the i-th fuzzy rule is added; i,j represents the activation intensity of the jth sample corresponding to the i-th fuzzy rule;
[0049] If the i-th fuzzy rule satisfies the second check condition, then delete the i-th fuzzy rule R i and cluster C i ; otherwise, it will not be deleted.
[0050] The inspection condition three:
[0051] if(D(h k )≥D 1 )then(Θ i ,a i are updated)
[0052] If the third check condition is not met, no update is performed; otherwise, if the third check condition is met, the covariance matrix Θ i Updated to:
[0053]
[0054] The subsequent parameter a i Updated to:
[0055]
[0056] Among them, D 1 is a constant.
[0057] The beneficial effects of the present invention are: due to the complex network structure, large number of parameters, and difficulty in implementing the update process online, conventional soft sensor deep learning networks such as SAE cannot handle time-varying problems well, so they usually do not have adaptive capabilities, while shallow networks such as evolutionary fuzzy rule systems can update network structure and parameters at the same time, and have good adaptive capabilities, but due to their simple structure, they cannot handle nonlinear problems well. The present invention combines the advantages of both, takes their strengths and makes up for their weaknesses, constructs a deep evolutionary fuzzy rule system, and simultaneously performs online prediction and modeling of nonlinear and time-varying processes, greatly improving the prediction accuracy and stability of industrial process soft sensor modeling. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 is a flow chart of the present invention;
[0059] Figure 2 It is a structural block diagram of the present invention;
[0060] Figure 3 It is a process flow diagram of an industrial debutanizer;
[0061] Figure 4 It is a prediction curve diagram of different methods and the method of the present invention. DETAILED DESCRIPTION
[0062] The invention will be further described below in conjunction with the accompanying drawings and embodiments, but the content of the invention is not limited to the scope of the embodiments.
[0063] Example 1: Figure 1-4 As shown, according to one aspect of an embodiment of the present invention, a method for industrial process soft measurement based on a deep evolutionary fuzzy rule system is provided, comprising: collecting historical industrial process data, constructing a pre-training sample set D pretrn ; The pre-training sample set D pretrn Including pre-trained feature set X pretrn ={x 1 ,x 2 ,…,x k ,…,x N}、The true label set is y pretrn ={y 1 ,y 2 ,...,y k ,...,y N}, x k represents the kth pre-trained feature, each pre-trained feature consists of multiple auxiliary variables, y k represents the kth true label, N represents the total number of pre-training samples; using the pre-training feature set X pretrn Pre-train the representation learning network to obtain the pre-trained representation learning network; the pre-trained feature set X pretrnRe-input into the pre-trained representation learning network to obtain the initial feature sample set D init ; Based on the initial feature sample set D init , the true label y pretrn Pre-train the prediction learning network to obtain a deep evolutionary fuzzy rule system; perform online prediction on the samples to be predicted based on the pre-trained representation learning network and the deep evolutionary fuzzy rule system, and update the deep evolutionary fuzzy system when the labels of the samples to be tested arrive.
[0064] Furthermore, the optional specific implementation process of the present invention is described as follows;
[0065] S1. Collect historical industrial process data and construct pre-training sample set D pretrn ; The pre-training sample set D pretrn Including pre-trained feature set X pretrn ={x 1 ,x 2 ,…,x k ,…,x N}、The true label set is y pretrn ={y 1 ,y 2 ,...,y k ,...,y N}, x k represents the kth pre-trained feature, each pre-trained feature consists of multiple auxiliary variables, y k represents the kth true label, and N represents the total number of pre-training samples;
[0066] S2, using the pre-trained feature X in S1 pretrn Pre-training the representation learning network to obtain the pre-trained representation learning network;
[0067] The representation learning network adopts a stacked autoencoder SAE, and the stacked autoencoder SAE is trained by L autoencoders AE.
[0068] Each AE is composed of an encoder and a decoder; the encoder in this method is composed of a single-layer feedforward network, which can be expressed by the following formula:
[0069] h=f enc (x) = g(W enc x+b enc )
[0070] In the formula, x is the original sample, W enc is the encoder weight, b enc is the encoder bias, g is an optional activation function, generally the sigmoid function, and h is the hidden feature of the output; the decoder also uses the same architecture as the encoder:
[0071]
[0072] In the formula, h is the hidden feature, w dec is the decoder weight, b dec is the encoder bias, is the reconstructed sample of the output. The weights of the encoder and decoder can be determined by the following loss function:
[0073]
[0074] Where N is X pretrn The total number of samples, x i is the original input sample, The reconstructed sample is output after the codec, and the loss is used for backpropagation to determine the single AE pre-training parameters.
[0075] The pre-training process of the representation learning network SAE includes the following steps:
[0076] ①、Train the first layer of AE. The first layer of AE reconstructs the original sample. The process is as follows: X pretrn Input into the encoder and decoder in turn, and use the reconstruction loss of formula (3) to update the network and save the encoder parameters W enc ;
[0077] ②, the AEs of the 2nd to Lth layers reconstruct the hidden features of the previous layer in turn. The process can be as follows: the encoder of the trained AE of the first layer is used to obtain the feature sample H of the first layer pretrn , H pretrn Input into the encoder and decoder in turn, and use the reconstruction loss of formula (3) to update the network to complete the training of the second layer, and save the encoder parameters W of this layer enc ; Use the encoder of the trained AE in the second layer to obtain the feature samples of the second layer, complete the reconstruction of the feature samples in this layer, and then complete the training of the third layer, and so on until the training of the Lth layer is completed.
[0078] S3, the pre-trained feature X in S1 pretrn Re-input into the representation learning network after S2 pre-training to obtain the initial feature sample set D init ;
[0079] Specifically, we obtain the initial sample set D init as follows:
[0080]
[0081] In the formula, is the mapping of each encoder layer.
[0082] S4, based on the initial feature sample set Dinit , the true label y pretrn Pre-train the prediction learning network to obtain a deep evolutionary fuzzy rule system;
[0083] The initial feature sample set D init , the true label y pretrn Pre-train the prediction learning network to obtain a deep evolutionary fuzzy rule system, including: using the initial feature sample set D in S3 init The first feature sample h 1 Initialize the deep evolutionary fuzzy rule system; and use the remaining feature samples D remain ={h 2 ,h 3 ,…,h N} is used to update the deep evolutionary fuzzy rule system antecedent, the true label y pretrn It is used to update the subsequent parts of the deep evolutionary fuzzy rule system, and the evolution process is carried out sample by sample, repeated m times to obtain the deep evolutionary fuzzy rule system.
[0084] Furthermore, the deep evolutionary fuzzy rule system is composed of a series of fuzzy rules, each rule corresponding to a linear model:
[0085]
[0086] Among them, R i represents the i-th fuzzy rule, p i is the antecedent of the i-th fuzzy rule, a i is the least squares parameter of the consequent of the i-th fuzzy rule, h k is the kth pre-trained latent feature extracted by the learning network, h k ={h 1 ,h 2 ,..,h w}, w is the dimension of latent features, ~ is close to / similar to, so the prediction output of the entire deep evolutionary fuzzy rule system is for:
[0087]
[0088] Among them, λ i is the activation strength of the i-th fuzzy rule, is the predicted output value of the i-th fuzzy rule; D i (h k ) means h k and the local density between the i-th fuzzy rule, and A temporary variable.
[0089] Furthermore, the pre-training prediction learning network process includes:
[0090] S4.1. Extracting x using representation learning network k h k ,k=1h k is the initial feature sample set D init The first feature sample h 1 , and then initialize the global meta parameters:
[0091] μ←h 1 ; χ←∥h 1 ∥ 2
[0092] Among them, μ is the global mean, χ is the square of its binary norm, representing the global norm; ← represents the update assignment operation, the parameter on the left side of the arrow is the new update parameter value, and the parameter on the right side is the current parameter value (i.e., the old parameter value);
[0093] S4.2. Initialize the antecedent and consequent parameters of the first fuzzy rule, r = k = 1:
[0094] p r ←h k ; a r ←0 w+1 ; Θ r ←Ω 0 I (w+1)×(w+1)
[0095] Among them, Θ r is the covariance matrix of the rth fuzzy rule, p r is the prototype of the rth fuzzy rule, a r is the least squares parameter of the consequent of the rth fuzzy rule, I (w+1)×(w+1) is the identity matrix of (w+1)×(w+1), 0 w+1 is a (w+1)-dimensional all-zero vector; Ω 0 is the initialization constant of the covariance matrix (usually set to 100), used in the standard recursive least squares algorithm;
[0096] S4.3. Initialize local meta parameters:
[0097] C r ←{h k}; S r ←1;H r ←||h k || 2
[0098] Each fuzzy rule can form a cluster, represented by p r The cluster C rIt can be approximated as a tessellated Voronoi distribution, S r C r The cardinality, H r is the local norm of the rth fuzzy rule; k←k+1;
[0099] S4.4, when k>1, the fuzzy rule addition, fuzzy rule deletion, and consequent parameter update processes are performed in sequence:
[0100] S4.4.1. Global meta-parameter update:
[0101]
[0102] S4.4.2. Calculate the global density D(z): To test h k Can the potential energy of initialize a new rule? Let’s calculate its global density first:
[0103]
[0104] S4.4.3. Determine whether to add fuzzy rules. If the first check condition is met, the rule is added; if the second check condition is met, the rule is deleted; if the third check condition is met, the rule consequence parameter is updated.
[0105] Furthermore, the inspection condition 1:
[0106]
[0107] Among them, D i (h k ) means h k and the local density between the i-th fuzzy rule,
[0108] If the check condition is met, a new fuzzy rule is added to the deep evolutionary fuzzy rule system, and the new antecedent and subsequent parameters are set to:
[0109]
[0110] Among them, the antecedent p r is initialized to a new h k , the latter part a r Initialized to the least squares parameter a of all other consequents in the deep evolutionary fuzzy rule system i average;
[0111] If the check condition 1 is not satisfied, then the sample h k If it is not initialized as a new rule, use h k Update the most recent cluster
[0112]
[0113] The nearest cluster is h k To the fuzzy rule antecedent p j The cluster corresponding to the minimum Euclidean distance value is:
[0114]
[0115] The parameters of the least squares antecedent and consequent of other rules remain unchanged.
[0116] Furthermore, the rule deletion process: use the second check condition to determine whether to delete the rule in the rule base (i.e., if the i-th fuzzy rule satisfies the second check condition, then delete the i-th fuzzy rule R i and cluster C i ; otherwise not deleted):
[0117] Check condition 2: if(η i,k <η 0 )then(R i ,C i are removed)
[0118] Among them, η i,k The calculation is as follows:
[0119]
[0120] i=1,2,…,r,η i,k is the i-th fuzzy rule R i The corresponding availability of the kth sample, η 0 is an adjustable constant parameter (initialized to 0.05 in the embodiment of the present invention), I i is the time when the i-th fuzzy rule is added; i,j It represents the activation intensity of the jth sample corresponding to the i-th fuzzy rule.
[0121] Furthermore, the parameter update is mainly the update of the subsequent parameters: when the subsequent parameters meet the third check condition:
[0122] Check condition three: if(D(h k )≥D 1 )then(Θ i ,a i are updated)
[0123] Among them, D 1 =e -4 , the covariance matrix Θ i Updated to:
[0124]
[0125] The subsequent parameter a i Updated to:
[0126]
[0127] y k is the true label value corresponding to the sample.
[0128] The purpose of this condition is to reduce the computational burden. Instead of updating all rules, only the rule that satisfies the current input activation degree is updated, and the rest that do not meet condition three are not updated; k←k+1;
[0129] After completing all the above processes, the system returns to process 4.4.1 and waits for the next input until the instance sample h k =h N ,h N D remain The last feature sample, at this time continue to repeat the whole process until it is completed m times to obtain the evolutionary fuzzy system after pre-training.
[0130] S5, new sample x k When it arrives, the representation learning network SAE first extracts the feature sample h k , the feature sample h k Input into the deep evolution fuzzy rule system, inference and prediction output
[0131] S6, waiting for the true label y k When the real label arrives, the deep evolution fuzzy rule system is updated using the feature samples and the real label. The updating process is consistent with the rule adding, rule deleting and consequent parameter updating process in S4.4 of S4.
[0132] S7. Repeat S5 and S6 until the online prediction process is completed.
[0133] According to another aspect of an embodiment of the present invention, there is provided an industrial process soft measurement system based on a deep evolution fuzzy rule system, comprising any one of the above-mentioned industrial process soft measurement methods based on a deep evolution fuzzy rule system.
[0134] Example 2: The present invention provides a butane concentration prediction method, which is performed using any of the industrial process soft measurement methods based on deep evolution fuzzy rule systems described above.
[0135] The following is a detailed explanation using an industrial data debutanizer example.
[0136] Step 1: A total of 2394 raw debutanizer data were collected, which can be downloaded from http: / / www.springer.com / 1-84628-479-1. Figure 3 The following is a schematic diagram of the debutanizer structure. The auxiliary variables are selected as follows: top temperature (U1), top pressure (U2), top reflux (U3), top product outflow (U4), plate temperature (U5), bottom temperature 1 (U6), bottom temperature 2 (U7). Since the debutanizer process exhibits strong dynamics, it is necessary to delay the auxiliary variables. Here, the input variables are delayed by 6 sampling cycles. Therefore, the current data contains 7×(6+1)=49 auxiliary variables. The dominant variable is the concentration measurement value of butane. The total sample size after delay is 2388 data. The data is divided into pre-training sample set D in one step. pretrn (1000), online test dataset D test (1388).
[0137] Step 2: Pre-training dataset D pretrn Including pre-trained features for X pretrn (i.e. unsupervised training samples), the true label is y pretrn (i.e. supervised label y pretrn ), use unsupervised training samples to train the representation learning network SAE, SAE selects a three-layer architecture, which is trained by three separate AE architectures in sequence. After pre-training, the parameters of the pre-trained representation learning network remain unchanged. Table 1 gives the SAE pre-training hyperparameter settings. The specific process is as follows:
[0138] Table 1 SAE pre-training hyperparameter settings
[0139] Pre-training hyperparameters Setting Value Number of layers 3 Network Architecture 49-120-60-30-1 Batch size 20 Number of training rounds 200 Learning Rate 0.01
[0140] The third step is Figure 2 As shown in Figure 2, the entire deep evolutionary fuzzy rule system (DEFRS) consists of a representation learning network and a prediction learning network. The representation learning network is used for nonlinear latent feature extraction. After the representation learning network is pre-trained, the weights are fixed and no update process occurs. The prediction learning network is used for online prediction and update, and can update the structure and parameters. Based on this system, an online test data set D test Perform online prediction and use the root mean square error RMSE and determination coefficient R 2 As an evaluation index of prediction performance, the evaluation index calculation formula is as follows:
[0141]
[0142] Where N test , ytest , They represent the number of test samples, the predicted value of the test samples, the true value of the test samples, and the mean of the true value of the test samples. The smaller the RMSE value and the higher the R 2 The larger the value, the better the model prediction performance.
[0143] In order to verify the effectiveness and superiority of this method, two methods were used as comparison methods of the deep evolutionary fuzzy rule system (DEFRS) of the present invention: stacked autoencoder (SAE) and conventional evolutionary fuzzy rule system (EFRS). The stacked autoencoder has no adaptive updating capability, while EFRS has adaptive updating capability. However, due to its simpler system structure, EFRS usually cannot achieve the ability to handle high-dimensional nonlinear problems similar to the deep learning network (DNN).
[0144] Our approach combines the advantages of both.
[0145] Table 2 Comparison of prediction performance evaluation of different methods in debutanizer chemical process
[0146] method RMSE <![CDATA[R 2 ]]> Stacked Autoencoders (SAE) 0.2393 -0.536 Evolutionary Fuzzy Rule System (EFRS) 0.0686 0.874 Deep Evolutionary Fuzzy Rule System (DEFRS) 0.0630 0.893
[0147] From the prediction curve Figure 4 As can be seen from the performance index results in Table 2, the conventional deep feature extraction network SAE does not have online adaptive capabilities, cannot adapt to new environments, and significantly deteriorates under new test data. The conventional evolutionary fuzzy rule system (EFRS) has good prediction accuracy because it can update the structure and parameters. At the same time, the deep evolutionary fuzzy system of the method of the present invention combines the advantages of deep learning and has further improved the prediction accuracy. Therefore, compared with the two alone, the method of the present invention can simultaneously handle nonlinear and time-varying process characteristics and improve the effectiveness of modeling.
[0148] The specific implementation modes of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above implementation modes, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.
Claims
1. A soft sensing method for debutanizer based on deep evolution fuzzy rule system, It is characterized in that include: Collect historical debutanizer industrial process data and construct a pre-training sample set D pretrn , online test sample set; The pre-training sample set D pretrn Includes pre-trained feature set X pretrn ={x 1 ,x 2 ,…,x k ,…,x N }、True label set y pretrn ={y 1 ,y 2 ,...,y k ,...,y N }, x k represents the kth pre-trained feature, each pre-trained feature consists of multiple auxiliary variables, y k represents the kth true label, N represents the total number of pre-training samples; the multiple auxiliary variables include tower top temperature, tower top pressure, tower top reflux flow, tower top product outflow, tower plate temperature, tower bottom temperature; Using pre-trained feature set X pretrn Pre-training the representation learning network to obtain the pre-trained representation learning network; The pre-trained feature set X pretrn Re-input into the pre-trained representation learning network to obtain the initial feature sample set D init ; Based on the initial feature sample set D init , the true label y pretrn Pre-train the prediction learning network to obtain the pre-trained deep evolutionary fuzzy rule system; Based on the pre-trained representation learning network and deep evolutionary fuzzy rule system, online prediction is performed on the online test sample set, and the deep evolutionary fuzzy rule system is updated when the test sample label arrives; Butane concentration is predicted based on the pre-trained representation learning network and the updated deep evolution fuzzy rule system; The initial feature sample set D init , the true label y pretrn Pre-train the prediction learning network to obtain a deep evolutionary fuzzy rule system, including: using the initial feature sample set D init The first hidden feature sample h in 1 Initialize the deep evolutionary fuzzy rule system; and use the remaining latent feature samples D remain ={h 2 ,h 3 ,…,h N }Update the deep evolutionary fuzzy rule system antecedent, the true label y pretrn It is used to update the aftermath of the deep evolutionary fuzzy rule system, and the evolution process is performed sample by sample, and repeated m times to obtain the deep evolutionary fuzzy rule system; The pre-trained prediction learning network includes: S4.
1. Extract x using the pre-trained representation learning network k h k ,k=1h k is the initial feature sample set D init The first feature sample h 1 , and then initialize the global meta parameters: μ←h 1 ;χ←||h 1 || 2 Among them, μ is the global mean, χ is the square of its binary norm, representing the global norm; ← represents the update assignment operation, the parameter on the left of the arrow is the new update parameter value, and the parameter on the right is the current parameter value; S4.
2. Initialize the antecedent and consequent parameters of the first fuzzy rule, i=k=1: p i ←h k ;a i ←0 w+1 ;I i ←Oh 0 I (w+1)×(w+1) Among them, Θ i is the covariance matrix of the i-th fuzzy rule, p i is the antecedent of the i-th fuzzy rule, a i is the least squares parameter of the consequent of the i-th fuzzy rule, I (w+1)×(w+1) is the identity matrix of (w+1)×(w+1), 0 w+1 is a (w+1)-dimensional all-zero vector; Ω 0 is the initialization constant of the covariance matrix, used in the standard recursive least squares algorithm; S4.
3. Initialize local meta parameters: C i ←{h k };S i ←1;H i ←||h k || 2 Each fuzzy rule can form a cluster, represented by p i The cluster C i It can be approximated as a tessellated Voronoi distribution, S i C i The cardinality, H i is the local norm of the i-th fuzzy rule; k←k+1; S4.4, when k>1, the fuzzy rule addition, fuzzy rule deletion, and consequent parameter update processes are performed in sequence until the instance sample h k =h N ,h N D remain For the last feature sample, the process of adding fuzzy rules, deleting fuzzy rules, and updating consequent parameters is repeated until m times to obtain a deep evolutionary fuzzy rule system after pre-training. When k>1, the fuzzy rule adding, fuzzy rule deleting, and subsequent parameter updating processes are performed in sequence, including: S4.4.
1. Global meta-parameter update: S4.4.
2. Calculate the global density D(z): z∈{h k ,p 1 ,p 2 ,…,p r } S4.4.
3. Determine whether to add fuzzy rules. If the first check condition is met, the rule is added; if the second check condition is met, the rule is deleted; if the third check condition is met, the rule consequence parameter is updated.
2. The debutanizer soft sensing method based on the deep evolution fuzzy rule system according to claim 1, It is characterized in that The representation learning network adopts a stacked autoencoder SAE, and the stacked autoencoder SAE is trained by L autoencoders AE.
3. The debutanizer soft sensing method based on deep evolution fuzzy rule system according to claim 1, It is characterized in that The deep evolutionary fuzzy rule system consists of a series of fuzzy rules, each of which corresponds to a linear model: Among them, R i represents the i-th fuzzy rule, p i is the antecedent of the i-th fuzzy rule, a i is the least squares parameter of the consequent of the i-th fuzzy rule, h k is the kth pre-trained latent feature extracted by the representation learning network, h k =[t 1 ,t 2 ,...,t w ], w is the dimension of latent features, ~ is close to or similar to, so the prediction output of the entire deep evolution fuzzy rule system is for: Among them, λ i is the activation strength of the i-th fuzzy rule, is the predicted output value of the i-th fuzzy rule.
4. The debutanizer soft sensing method based on deep evolution fuzzy rule system according to claim 3, It is characterized in that The activation strength calculation formula of the fuzzy rule is as follows: Among them, D i (h k ) means h k and the local density between the i-th fuzzy rule, and is a temporary variable, S i Because p i The cluster C i The cardinality, H i is the local norm of the i-th fuzzy rule.
5. The debutanizer soft sensing method based on deep evolution fuzzy rule system according to claim 1, It is characterized in that The inspection condition 1: Among them, D i (h k ) means h k The local density between the i-th fuzzy rule, D 0 is a constant; If the check condition is met, a new fuzzy rule is added to the deep evolutionary fuzzy rule system, and the new antecedent and subsequent parameters are set to: Among them, the antecedent p r is initialized to a new h k , the posterior least squares parameter a r Initialized to the least squares parameter a of all other consequents in the deep evolutionary fuzzy rule system q average; If the check condition 1 is not satisfied, then the sample h k If it is not initialized as a new rule, use h k Update the most recent cluster The nearest cluster is h k To the fuzzy rule antecedent p j The cluster corresponding to the minimum Euclidean distance value is: The parameters of the least squares antecedent and consequent of other rules remain unchanged.
6. The debutanizer soft sensing method based on deep evolution fuzzy rule system according to claim 1, It is characterized in that The second inspection condition: if(η i,k <η 0 )then(R i ,C i are removed) Among them, η i,k The calculation is as follows: Where i = 1, 2, ..., r, η i,k is the i-th fuzzy rule R i The corresponding availability of the kth sample, η 0 is an adjustable constant parameter, I i is the time when the i-th fuzzy rule is added; i,j represents the activation intensity of the jth sample corresponding to the i-th fuzzy rule; If the i-th fuzzy rule satisfies the second check condition, then delete the i-th fuzzy rule R i and cluster C i ; otherwise, it will not be deleted.
7. The debutanizer soft sensing method based on deep evolution fuzzy rule system according to claim 1, It is characterized in that The inspection condition three: if(D(h k )≥D 1 )then(Θ i ,a i are updated) If the third check condition is not met, no update is performed; otherwise, if the third check condition is met, the covariance matrix Θ i Updated to: The subsequent parameter a i Updated to: Among them, D 1 is a constant.
Citation Information
Patent Citations
Industrial process soft measurement method based on layer-by-layer data expansion deep learning
CN110096810A
MSWI process dioxin emission soft measurement method based on integrated T-S fuzzy regression tree
CN114943151A
Deep TSK fuzzy classifier based on multilevel feature fusion
CN115186731A
An evolutionary neuro-fuzzy system using evolutionary fuzzy clustering and syndicated neural networks
DE202022104771U1
Fuzzy curve analysis based soft sensor modeling method using time difference Gaussian process regression
US20170061305A1
