Health assessment method of benzene-olefin ratio control system based on class imbalance data enhancement

Abstract

The present invention discloses a benzene-olefin ratio control system health assessment method based on class imbalance data enhancement, including: obtaining the operating data of the target benzene-olefin ratio control system under preset working conditions and the system benzene-olefin ratio output value; obtaining multiple characteristic factors therefrom after preprocessing; inputting the multiple characteristic factor data of the real class imbalance into the C‑WGAN‑GP data enhancement model, and outputting the generated data; inputting the real class imbalance data and the generated data into the ACNN health assessment model, and outputting the system benzene-olefin ratio output prediction value to achieve the health assessment of the target benzene-olefin ratio control system; tracing the non-optimal factors that cause the health level of the target benzene-olefin ratio control system to decline through the attention mechanism of the ACNN health assessment model. The method uses the C‑WGAN‑GP model to generate new samples to provide reliable data support for the assessment; by constructing an ACNN model that integrates both assessment and tracing, the accuracy of the assessment model and the interpretability of the decision-making process are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of control system performance evaluation, and more specifically to a benzene-olefin ratio control system health evaluation method based on class imbalance data enhancement. Background Art

[0002] Styrene is an important basic organic chemical raw material, widely used in the production and preparation of rubber, synthetic plastics, pesticides and other products. In the actual chemical process, the alkylation process section is the core unit of styrene production. The alkylation reaction process flow chart of the alkylation unit is as follows: Figure 1 As shown, R-101 and R-102 both represent alkylation reactors; M-101 and M-102 both represent mixers; E-102 and E-103 both represent coolers; C-101 represents a pre-fractionator; P-101 represents an alkylation circulating benzene pump; V-101 represents an ethylene buffer tank; FIC-1003, FIC-1008, FIC-1009, FIC-1012 and FIC-1013 all represent regulating valves; benzene and ethylene undergo alkylation reaction under the action of an alkylation catalyst to prepare ethylbenzene. In order to extend the service life of the catalyst and avoid the generation of by-products, the healthy operation of the benzene-olefin ratio control system becomes the key to ensuring product quality and improving comprehensive economic benefits. The system adopts a two-layer architecture based on upper-layer model predictive control (MPC)-bottom-layer PID loop control. The bottom layer uses 1 benzene main loop and 4 ethylene sub-loops to form a PID closed-loop ratio control, so that the benzene-olefin ratio reaches the desired control range. However, due to the influence of uncertain factors such as changes in raw materials, fluctuations in working conditions, and various interferences, the system will inevitably deviate from the good working state at the beginning of operation, making it difficult for the company to achieve production expectations. Therefore, for the benzene-olefin ratio control system with a double-layer architecture, how to timely and accurately evaluate the system health status and trace the cause of abnormalities is of great significance for system intelligent regulation, optimized decision-making and safe operation and maintenance, and to achieve stable and efficient production operation.

[0003] The existing control system performance evaluation methods mainly include those based on multivariate statistics and neural networks. The evaluation method based on multivariate statistics mainly uses various statistical tools such as variance, covariance, and entropy to analyze the system operation data. It is more interpretable and easy to implement, but the processing results for complex data types are not ideal. The evaluation method based on neural networks uses neural networks to learn and predict the operating status data, and evaluates the system performance based on the prediction results. It can effectively capture the nonlinear relationship between the input variables of complex systems, but the model has weak interpretability. After entering the era of big data, deep learning can use multi-hidden layer networks to automatically learn from data, extract the inherent laws and representation levels that reflect the essence of the problem, and use them for benzene-olefin ratio control system health assessment. It can better reflect the complex correlation characteristics between system levels and loop units, but it has poor model interpretability and it is difficult to trace and analyze what factors cause the decline in system health level.

[0004] In addition, the quality of the evaluation data and the number of samples are key factors in determining whether the control system health assessment model can achieve accurate assessment. However, in the actual operation of the system, the system is in a good health state and sub-health state far more often than an unhealthy state. The obtained evaluation data set has an imbalance problem in the status of each category, which makes it difficult for the established evaluation model to make an accurate and reliable evaluation of the system health.

[0005] Therefore, for the MPC-PID two-layer architecture benzene-olefin ratio complex control system with interactive correlations between levels and coupling between loops, when there is a class imbalance problem in the evaluation data set, how to obtain a complete evaluation data set to improve the accuracy of system health status assessment and solve the problems of insufficient system nonlinear attribute extraction ability of traditional multivariate statistical evaluation methods and poor interpretability of deep learning models are issues that technical personnel in this field urgently need to solve. Summary of the invention

[0006] In view of the above problems, the present invention provides a benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement to at least solve some of the technical problems mentioned in the above background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A health assessment method for a benzene-ethylene ratio control system based on class imbalance data enhancement includes the following steps:

[0009] S1. Obtaining operation data and system benzene-olefin ratio output values ​​of a target benzene-olefin ratio control system under a plurality of preset working conditions, and preprocessing the operation data and the system benzene-olefin ratio output values ​​respectively;

[0010] S2. Based on the pre-processed operating data and the system benzene-olefin ratio output value, a plurality of characteristic factors related to the health status of the target benzene-olefin ratio control system are obtained;

[0011] S3, selecting characteristic quantity data corresponding to the plurality of characteristic factors from the operation data as real class imbalance data showing uneven distribution of health status of the target benzene-olefin ratio control system; inputting the real class imbalance data into the C-WGAN-GP data enhancement model, and outputting generated data;

[0012] S4, taking the real class imbalance data and the generated data as a complete evaluation data set, inputting the complete evaluation data set into the constructed ACNN health assessment model, and outputting the system benzene-olefin ratio output prediction value;

[0013] S5. According to the predicted value of the benzene-olefin ratio output of the system, a health assessment of the target benzene-olefin ratio control system is performed; and the non-optimal factors causing the decline in the health level of the target benzene-olefin ratio control system are traced back through the attention mechanism in the ACNN health assessment model.

[0014] Furthermore, in the step S1, the operation data and the system benzene-olefin ratio output value are preprocessed respectively, specifically including:

[0015] Performing denoising processing on the operation data;

[0016] The Min-Max interval mapping normalization method is adopted to map the benzene-olefin ratio output value of the system between [0,1].

[0017] Furthermore, the step S2 specifically includes:

[0018] S21, based on the pre-processed operation data, pre-evaluate each hierarchical unit in the benzene-olefin ratio control system according to preset evaluation indicators: the preset evaluation indicators include the hierarchical unit evaluation indicators of the benzene-olefin ratio control system and the follow-up performance evaluation indicators of each loop in the benzene-olefin ratio control system;

[0019] S22. According to the preliminary evaluation result, a plurality of characteristic factors whose correlation with the system benzene-olefin ratio output value exceeds a preset value are determined.

[0020] Furthermore, the multiple characteristic factors include: a benzene loop, four ethylene loops, a performance evaluation index of MPC, and a follow-up index characterizing loop coupling properties.

[0021] Furthermore, the step S2 further includes:

[0022] S23, numbering the multiple characteristic factors, and using a mutual information heat map method to analyze the nonlinear relationship between every two characteristic quantities.

[0023] Furthermore, in the C-WGAN-GP data augmentation model:

[0024] The generator adopts a deconvolution network structure, which consists of 1 fully connected layer and 3 deconvolution layers; the input of the generator is random noise that conforms to the Gaussian distribution; the fully connected layer samples from the random noise; the deconvolution layer learns the distribution of characteristic quantities in the benzene-olefin ratio control process; the output of the generator is 1025*7-dimensional data;

[0025] The discriminator adopts a convolution structure, which is composed of 3 convolution layers and 1 fully connected layer; the convolution layer is used to extract the dynamic characteristic information of the characteristic quantity curve in the benzene-olefin ratio control process; the fully connected layer is used to discriminate the probability.

[0026] Furthermore, the generator loss function in the C-WGAN-GP data augmentation model is expressed as:

[0027] L(G)=1-D(G(z))

[0028] Where L(G) represents the generator loss function of the C-WGAN-GP data augmentation model; D(·) and G(·) represent the discriminator function and generator function, respectively;

[0029] The discriminator loss function in the C-WGAN-GP data augmentation model is expressed as:

[0030]

[0031] Where L(D) represents the discriminator loss function of the C-WGAN-GP data augmentation model; represents the gradient operator; λ represents the penalty term coefficient; represents the random interpolation between the real class imbalanced data and the generated data, ζ follows a uniform distribution within the range of

[01] ; Represents the imbalanced data distribution P from the real class data (.) and generate data distribution P g (.) Straight line uniform sampling between sampling points.

[0032] Furthermore, the step S4 specifically includes:

[0033] S41, inputting the complete assessment data set into the ACNN health assessment model through the input layer;

[0034] S42, extracting features from the complete evaluation data set through a CNN convolutional layer to obtain convolutional features;

[0035] S43, sampling the convolutional features through a CNN pooling layer to obtain a CNN pooling layer output;

[0036] S44, processing and calculating the output of the CNN pooling layer through the hidden layer to obtain the intermediate state of the feature factor; obtaining the similarity between the intermediate state feature vector and the hidden state through the Attention mechanism layer;

[0037] S45, obtaining the attention weight of the feature factor hidden layer vector through the softmax function, and performing weighted sum processing on the attention weight and the intermediate state feature vector to obtain the output of the Attention mechanism layer;

[0038] S46, passing the output of the attention mechanism layer through a fully connected layer to obtain a system benzene-olefin ratio output prediction value representing a health state of the system;

[0039] S47, outputting the system benzene-olefin ratio output prediction value through the output layer.

[0040] Furthermore, the step S5 further includes:

[0041] According to the weight adaptively assigned to each characteristic factor by the Attention mechanism layer in the ACNN health assessment model, a contribution diagram is drawn, and the nonlinear relationship between the health status of the target benzene-olefin ratio control system and the characteristic factors is quantitatively analyzed, so as to effectively trace the non-optimal factors that cause the decline in the health level of the target benzene-olefin ratio control system.

[0042] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a health assessment method for a benzene-to-ethylene ratio control system based on class imbalance data enhancement, which has the following beneficial effects:

[0043] The present invention focuses on the problem that the actual system operation health status is unevenly distributed, resulting in incomplete evaluation data sets, which affects the reliable evaluation of the system health status. The C-WGAN-GP adversarial network is used to learn the distribution law of the system state data and generate new samples to enhance the original unbalanced data, thereby obtaining a complete evaluation data set that can characterize the system's full health status attributes, providing reliable data support for system health evaluation.

[0044] The present invention embeds the attention mechanism into CNN to construct a system health assessment model that integrates evaluation and traceability. It can not only use CNN to mine the nonlinear and strongly coupled correlation properties of system subsystems and loops to reliably evaluate the health status of the system; but also measure the contribution of each feature quantity in the attention mechanism module to the predicted value of the benzene-olefin ratio output according to its weight, quantify the factors that cause the decline in the health level of the traceability system, enhance the interpretability of the deep learning assessment model to a certain extent, reduce the amount of computation brought by the reconstruction of the traceability network, and provide technical support for the intelligent evaluation and maintenance of the health status of the benzene-olefin ratio control system throughout its life cycle.

[0045] Other features and advantages of the present invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the present invention.

[0046] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0048] Figure 1 A schematic diagram of the alkylation reaction process flow of the hydrocarbonation unit provided for the relevant prior art.

[0049] Figure 2 A schematic flow chart of a benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement provided in an embodiment of the present invention.

[0050] Figure 3 A schematic diagram of normalization of benzene-olefin ratio output interval mapping provided in an embodiment of the present invention.

[0051] Figure 4 A heat diagram of the mutual information of feature factor vectors provided in an embodiment of the present invention.

[0052] Figure 5 Schematic diagram of the generator and discriminator loss functions provided for an embodiment of the present invention.

[0053] Figure 6 A schematic diagram of the structure of the ACNN health assessment model provided in an embodiment of the present invention.

[0054] Figure 7 A schematic diagram comparing the health assessment results of the benzene-to-olefin ratio control system under different data sets provided in an embodiment of the present invention.

[0055] Figure 8 A schematic diagram of the contribution of various characteristic factors to the health level of the benzene-olefin ratio control system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only 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 ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] The embodiment of the present invention provides a health assessment method for a benzene-olefin ratio control system based on quasi-unbalanced data enhancement, based on a styrene alkylation unit alkylation reaction process production line and combined with the process requirements and control mechanism of the benzene-olefin ratio control system. Figure 2 As shown; specifically includes the following steps:

[0058] S1. Obtaining the operating data and system benzene-olefin ratio output value of the target benzene-olefin ratio control system under various preset working conditions, and preprocessing the operating data and the system benzene-olefin ratio output value respectively;

[0059] S2. Based on the pre-processed operating data and the system benzene-olefin ratio output value, a plurality of characteristic factors related to the health status of the target benzene-olefin ratio control system are obtained;

[0060] S3, selecting characteristic quantity data corresponding to multiple characteristic factors from the operating data as real class imbalanced data with uneven distribution of health status of the target benzene-olefin ratio control system; inputting the real class imbalanced data into the C-WGAN-GP data enhancement model, and outputting generated data;

[0061] S4, taking the real class imbalanced data and the generated data as the complete evaluation data set, inputting the complete evaluation data set into the constructed ACNN health assessment model, and outputting the system benzene-olefin ratio output prediction value;

[0062] S5. According to the predicted value of the system benzene-olefin ratio output, the health assessment of the target benzene-olefin ratio control system is realized; and the non-optimal factors that cause the health level of the target benzene-olefin ratio control system to decline are traced through the attention mechanism in the ACNN health assessment model.

[0063] The method of the present invention focuses on the problem that the uneven distribution of the actual system operation health status leads to incomplete evaluation data sets, which affects the reliable evaluation of the system health status. The C-WGAN-GP adversarial network is used to learn the distribution law of the system state data and generate new samples to enhance the original unbalanced data, so as to obtain a complete evaluation data set that can characterize the full health status attributes of the system, and provide reliable data support for the accurate establishment of the health evaluation model; in addition, an ACNN health evaluation model of the benzene-olefin ratio control system that integrates evaluation and traceability is established. While accurately evaluating the health status of the system, the weight of each feature quantity is adaptively assigned through the attention mechanism, and the non-optimal factors that cause the health level of the target benzene-olefin ratio control system to decline are traced, so as to enhance the interpretability of the deep learning evaluation model;

[0064] Next, each of the above steps will be described in detail.

[0065] In the above step S1, the operation data and the system benzene-olefin ratio output value of the target benzene-olefin ratio control system under a plurality of preset operation conditions are obtained, and the operation data and the system benzene-olefin ratio output value are preprocessed respectively; specifically, the process includes:

[0066] Combined with the sampling frequency of the on-site sensor equipment and the historical database, under the five preset operating conditions of normal system operation T1, system output decreases, benzene-olefin ratio output is basically unaffected T2, benzene-olefin ratio output is too high T3, benzene-olefin ratio output is too low T4, and benzene-olefin ratio output is uncontrollable T5, the operating data of the target benzene-olefin ratio control system (i.e., the output and control quantity of each level unit loop) and the system benzene-olefin ratio output value at the corresponding time are collected;

[0067] The collected operation data and the system benzene-olefin ratio output value are preprocessed; specifically including:

[0068] (1) De-noising the operating data of the benzene-olefin ratio control system (i.e., the output and control quantity of each hierarchical unit loop) to serve as the basic data for obtaining the evaluation results between hierarchical units in the subsequent step S2;

[0069] (2) In combination with the system process control requirements and expert experience and knowledge, the system benzene-olefin ratio output value is used as the key indicator for evaluating the overall operating health status of the benzene-olefin ratio control system. The system benzene-olefin ratio output value is de-noised and interval-mapped and normalized. For details, see Figure 3 As shown, Figure 3Figure (a) shows the actual output value of the system benzene-olefin ratio, which is mostly around 10.5. Due to the limitations of process conditions in the on-site production process, the output of the benzene-olefin ratio is mostly high. In order to effectively compare with the prediction results of the evaluation model and ensure the consistency of data types, the Min-Max interval mapping normalization method is adopted to map the system benzene-olefin ratio output value between [0,1] as a reference benchmark for the health assessment of the benzene-olefin ratio control system. The interval mapping normalization process can be expressed as:

[0070]

[0071] Among them, Y min-max is the normalized system benzene-olefin ratio output value; Y is the current system benzene-olefin ratio output value of the sequence; Y min is the minimum value of the system benzene-olefin ratio output; Y max It is the maximum value of the system benzene-olefin ratio output value;

[0072] The output histogram after interval mapping normalization is as follows Figure 3 As shown in Figure (b); Figure 3 As can be seen in Figure (b), during the actual operation of the benzene-olefin ratio control system, there are more cases in a sub-healthy state and fewer in an unhealthy state. The distribution of each health status data set is uneven, which makes it difficult for machine learning methods to make an accurate and reliable evaluation of the health status of the benzene-olefin ratio control system.

[0073] In the above step S2, based on the pre-processed operating data and the system benzene-olefin ratio output value, combined with the system control mechanism and multivariate statistical methods, multiple characteristic factors related to the health status of the target benzene-olefin ratio control system are obtained; specifically including:

[0074] S21. According to the process and control mechanism of the benzene-olefin ratio control system, a preset evaluation index is determined; based on the pre-processed operation data (i.e., the output and control amount of each level unit loop), each level unit in the benzene-olefin ratio control system is pre-evaluated according to the preset evaluation index:

[0075]

[0076] η s =1-max(p 1min ,p 2min ) (3)

[0077]

[0078]

[0079] Among them, η h is the hierarchical unit evaluation index of the benzene-olefin ratio control system; H2(·) GMETo estimate the hierarchical unit GME benchmark using Renyi entropy H2; is the generalized output feedback invariant term of each level unit in the benzene-olefin ratio control system; f is the actual generalized output; η s is the evaluation index of the tracking performance of each loop in the benzene-olefin ratio control system; p1 is the distance that the system benzene-olefin ratio output value deviates from the control target interval; p 1min is the minimum value of p1, which is used to constrain the degree to which the benzene-olefin ratio output value deviates from the midline of the control target interval; p2 is the peak value of the system benzene-olefin ratio output value; p 2min is the minimum value of p2, which is used to constrain the influence of the peak value in the benzene-olefin ratio output value on the control quality; p1, p2∈[0,0.1,0.2,L,1]; L is the length of the evaluation sliding data window; Y i is the system benzene-olefin ratio output value at time i;

[0080] S22, based on the prior knowledge of the process mechanism and control requirements of the benzene-ethylene ratio control system, and according to the pre-evaluation results of each level unit in the benzene-ethylene ratio control system, multiple characteristic factors that significantly affect the output value of the system benzene-ethylene ratio are determined; in the embodiment of the present invention, the selected multiple characteristic factors include: the performance evaluation index of the benzene loop, the four ethylene loops, the MPC, and the follow-up index characterizing the loop coupling property;

[0081] S23, the selected multiple characteristic factors are numbered and analyzed by mutual information heat map method; specifically, the seven characteristic quantities are numbered according to R1 to R7, and due to the correlation between levels and coupling between loops in the benzene-olefin ratio control system, the mutual information heat map method is used to perform relevant data analysis on the selected seven characteristic quantities, indicating the nonlinear and strong correlation characteristics between the variables, such as Figure 4 shown.

[0082] As an indicator to measure the amount of shared information between two variables, the mutual information heat map includes not only linear relationships but also more complex nonlinear relationships. Its value ranges from 0 to 1. The larger the value, the greater the dependence between the variables. Figure 4 The mutual information values ​​between the variables are mostly above 0.5, and only a few variables show low dependence. It can be seen that there is no purely linear relationship between the determined characteristic factor vectors, and there are nonlinear relationships and strong correlation characteristics between the variables.

[0083] In the above step S3, data corresponding to the above multiple characteristic factors are selected from the above operation data, i.e., the output and control quantities of the unit loops at each level, as real data with unbalanced distribution of the system health status class, recorded as real class unbalanced data; the real class unbalanced data is input into the C-WGAN-GP data enhancement model, and the generated data is output to enhance the original data set;

[0084] The C-WGAN-GP data enhancement model is a Wasserstein generative adversarial network with gradient penalty (C-WGAN-GP) data enhancement model based on the gradient penalty term. The construction and training content of the C-WGAN-GP data enhancement model specifically include:

[0085] (1) Composition of the C-WGAN-GP data augmentation model:

[0086] Aiming at the characteristics of chemical production of benzene-olefin ratio control system, a data generation network based on C-WGAN-GP is proposed. The generator adopts a deconvolution network structure, which consists of 1 fully connected layer and 3 deconvolution layers. The input of the generator is noise that conforms to the Gaussian distribution. The fully connected layer samples from the random noise, and the deconvolution layer learns the distribution of characteristic quantities in the benzene-olefin ratio control process. The number of filters is 36, 72 and 1 respectively. The output of the generator is 1025*7-dimensional data. The discriminator adopts a convolution structure, which consists of 3 convolution layers and 1 fully connected layer. The convolution layer is used to extract the dynamic feature information of the characteristic quantity data curve for evaluating the benzene-olefin ratio control process, and the fully connected layer is used to discriminate the probability. The number of filters is 36, 72 and 144 respectively.

[0087] (2) Training process of C-WGAN-GP data augmentation model:

[0088] In the actual operation of the benzene-to-olefin ratio control system, the number of healthy and sub-healthy states is far greater than that of unhealthy states, which results in imbalance in the acquired operation and evaluation data, making it difficult for the machine learning-based evaluation method to make an accurate and reliable evaluation of the health status of the system. In the embodiment of the present invention, a CGAN generative model with adversarial thinking is used to enable the generator to generate virtual samples in a specified direction or type. The CGAN model loss function is:

[0089]

[0090] Where L(G,D) represents the CGAN model loss function; x represents the real class imbalanced data of the input; z represents the generator input noise; y1 represents the generation condition, which includes the four data attributes of the mean, variance, standard deviation and range of the training data; p data (x), p z (z) represents the real class imbalanced data distribution and random noise distribution respectively; D(·) and G(·) represent the discriminator function and generator function respectively, where the generator is training the data distribution given the additional information y1;

[0091] For the above CGAN model loss function, the Wasserstein distance, which can reflect the distribution distance of the real class imbalanced data and the generated data used for evaluation and changes smoothly, is used to measure the difference in the probability distribution between the two samples. The CGAN cost function based on the Wasserstein distance is expressed as:

[0092]

[0093] Among them, W(p g ,p data ) represents the CGAN cost function; ‖D‖ L represents the K-order Lipschitz norm; Indicates the generated data. Represents the generated data distribution;

[0094] In order to meet the Lipschitz continuity constraint and avoid overfitting of the discriminator D, a gradient penalty term is introduced on the basis of the cost function. The gradient norm of the discriminator output is directly constrained according to the input of the discriminator D, and the discriminator loss function of the improved C-WGAN-GP data augmentation model is obtained; the discriminator loss function is expressed as:

[0095]

[0096] Where L(D) represents the discriminator loss function; represents the gradient operator; λ represents the penalty term coefficient; represents the random interpolation between the real class imbalanced data and the generated data, ζ follows a uniform distribution within the range of

[01] ; Represents the imbalanced data distribution P from the real class data (.) and generate data distribution P g (.) Straight line uniform sampling between sampling points;

[0097] The generator loss function in the C-WGAN-GP data augmentation model is expressed as:

[0098] L(G)=1-D(G(z)) (9)

[0099] Where L(G) represents the generator loss function of the C-WGAN-GP data augmentation model; D(·) and G(·) represent the discriminator function and generator function, respectively;

[0100] Set initialization parameters, initialize the network structure of the generator and discriminator in the C-WGAN-GP data augmentation model, load the real class imbalanced dataset of the benzene-olefin ratio control system, iterate the training network, check the changes in the loss function and gradient norm, and ensure the stability of the training process; when the values ​​of the discriminator loss function and the generator loss function are both minimized, stop the training and save the trained model; mix the real class imbalanced dataset and the generated data as a complete evaluation dataset for system health assessment;

[0101] In the embodiment of the present invention, a total of 7 items of evaluation indicators of benzene loop, 4 ethylene loops, MPC and follow-up performance indicators are used as basic feature quantities in the C-WGAN-GP data enhancement model, and 475 groups of feature quantity data with uneven health status distribution under five common working conditions of benzene-ethylene ratio control system T1~T5 are collected, and the organized 475*7 data are sent to the C-WGAN-GP data enhancement model; by adjusting appropriate parameters, the generator loss function curve is obtained as shown in Figure 5 As shown in Figure (a), the discriminator loss function curve is as follows Figure 5 As shown in Figure (b), from Figure 5 As can be seen in the figure, with the increase of the number of iterations, the training gradually stabilizes, the generator and the discriminator are in a Nash equilibrium state, and the generator can stably generate data that is approximately distributed with the 7 basic feature quantities training data.

[0102] Table 1 is a comparison of KL divergence values ​​under different data augmentation models. It can be seen from the table that, compared with the GAN, WGAN and WGAN-GP basic models, the C-WGAN-GP data augmentation model used in the present invention has a smaller KL divergence value, and can generate high-quality new samples that are more similar to the real data distribution, which is used to evaluate data augmentation and expansion to reduce the impact of data state category imbalance on the accuracy of system health assessment;

[0103] Table 1 KL divergence metrics under different models

[0104] Model GAN WGAN WGAN-GP C-WGAN-GP KL divergence value 0.4729 0.4054 0.3671 0.3524

[0105] In the above step S4, an ACNN system health assessment model integrating assessment and traceability is constructed, and the above real class imbalance data and generated data are used as a complete assessment data set for system health assessment; the complete assessment data set is input into the constructed ACNN health assessment model, and the system benzene-olefin ratio output prediction value is output;

[0106] The composition of the ACNN health assessment model is as follows: Figure 6As shown in the figure, it includes input layer, CNN layer and Attention mechanism layer, fully connected layer and output layer; based on this, the correlation attributes between the levels of the benzene-olefin ratio control system and the coupling characteristics between the internal loops of the levels are mined, and the ACNN health assessment model is trained offline based on the complete evaluation data set; specifically, the training data set can be divided into training set, validation set and test set according to 8:1:1. According to the process characteristics of the benzene-olefin ratio system, a one-dimensional convolution layer is used, the convolution kernel size is 3×3, the sliding step is 1, the activation function is LeakyReLU, the pooling layer uses the maximum pooling, the Adam optimizer is used to optimize the network parameters, the initial value of the learning rate is set to 0.001, the number of training iterations is set to 1000, the Attention layer Dropout rate is set to 0.3, the output of the Attention layer is used as the input of the fully connected layer, and the fully connected layer uses Sigmoid as the activation function. The ACNN health assessment model is established using the training set.

[0107] In the specific implementation process, the working principle of the ACNN health assessment model includes:

[0108] (1) Input layer: The complete assessment data set is input into the ACNN health assessment model through the input layer;

[0109] (2) CNN convolutional layer and pooling layer, specifically:

[0110] ① The CNN convolutional layer is used to extract features from the complete evaluation data set to explore the strong coupling characteristics between the system MPC-PID hierarchical control and the loops; the complete evaluation data set is divided into d parts, each of which includes n data, each of which corresponds to m basic system health states, forming an n*m*d matrix as the input of the CNN convolutional layer; the convolution features of the output of the CNN convolutional layer are as follows:

[0111]

[0112] Among them, f cov (.) represents the activation function; k represents the sliding window size; w n,m represents the weight of the convolution kernel n rows and m columns; X i+n,j+m is the value of the nth row and mth column of the input feature matrix; b n,m is the convolution kernel bias;

[0113] ② The above convolution features are sampled through the CNN pooling layer using a 2*2 filter and a sliding window with a step size of 1 to reduce the data feature size and network parameters. The output of the CNN pooling layer is as follows:

[0114] y[i]=max(x[i:i+k-1]) (11)

[0115] (3) Attention mechanism layer

[0116] ① The output of the CNN pooling layer is processed through the hidden layer to obtain the intermediate state of the feature factors; through the Attention mechanism layer, the function score([h t,i ,h t ]) Calculate the eigenvector h of the intermediate state t,i With the hidden state h t The similarity between them, the expression of the Attention mechanism layer is as follows:

[0117] score([h t,i ,h t ])=W s h t T +b s (12)

[0118] Among them, W s represents the weight matrix; b s represents the bias vector;

[0119] ② After the softmax function, the attention weight α of the hidden layer vector of the feature factor in the Attention mechanism layer is obtained i , and the feature vector h of the intermediate state t,i The weighted summation is used to obtain the output h of the Attention mechanism layer. t · , whose expression is as follows:

[0120]

[0121]

[0122] Where τ is the output node of the fully connected layer;

[0123] (4) Fully connected layer: h t · Input to the fully connected layer to obtain the system benzene-olefin ratio output prediction value that represents the health status of the system;

[0124] (5) Output layer: outputs the predicted value of the system benzene-olefin ratio output through the output layer.

[0125] In order to verify the effectiveness of the complete evaluation data set in the present invention, in an embodiment of the present invention, the real class imbalanced data and the complete evaluation data set are respectively sent to the ACNN health assessment model, such as Figure 7 As shown in the figure, the actual health assessment index value corresponding to the real class imbalance data is obtained (i.e. Figure 7The ACNN+original dataset in the above figure) and the system benzene-olefin ratio output prediction value corresponding to the complete evaluation dataset (i.e. Figure 7 ACNN+ perfect dataset in );

[0126] In addition, the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R 2 )Calculate the assessment accuracy of the benzene-olefin ratio control system health assessment model:

[0127]

[0128]

[0129]

[0130] Among them, y t represents the system benzene-olefin ratio output value after normalization of the benzene-olefin ratio interval mapping at time t; y′ t represents the system benzene-olefin ratio output prediction value after normalization of the benzene-olefin ratio interval mapping at time t; Represents y t The average value of; N represents the number of test samples.

[0131] The comparison of the evaluation accuracy indicators of the corresponding evaluation models under the two data sets is shown in Table 2:

[0132] Table 2 Comparison of prediction evaluation indicators under different data sets

[0133] category RMSE MAE <![CDATA[R 2 ]]> ACNN+Original Dataset 0.0117 0.0104 0.8677 ACNN+ complete dataset 0.0048 0.0031 0.9984

[0134] from Figure 7 It can be seen intuitively that, compared with the real class imbalanced data (i.e., the original data set), the system health assessment results based on the enhanced complete data set are closer to the real health status of the system; combined with the comparison of the evaluation accuracy indicators in Table 2, it is further explained that the use of the enhanced data set can obtain more accurate prediction and evaluation results, which verifies that the introduction of the data enhancement method in this paper can well improve the imbalance of evaluation data quality and its impact on the accuracy of system health assessment.

[0135] In order to further verify the accuracy of the ACNN health assessment model in the present invention, in an embodiment of the present invention, four models, namely ACNN and CNN, deep belief network DBN, and deep belief ADBN with attention mechanism, are respectively used. Based on the complete evaluation data set obtained after enhancement, a comparative experiment on the health assessment of the double-layer architecture benzene-olefin ratio control system is carried out. The comparison of evaluation results and evaluation indicators under the four different models is shown in Table 3.

[0136] Table 3 Comparison of evaluation index values ​​of different health assessment models

[0137] Model RMSE MAE <![CDATA[R 2 ]]> CNN 0.0053 0.0037 0.9715 ACNN 0.0048 0.0031 0.9984 DBN 0.0067 0.0041 0.9436 ADBN 0.0059 0.0039 0.9612

[0138] From the four evaluation model accuracy evaluation index values ​​shown in Table 3, it can be seen that the evaluation index R based on ACNN of the present invention is 2 The value is higher than that of the other three models, with higher evaluation accuracy, while the smaller RMSE and MAE indicate lower errors than other models. Overall, the effective combination of CNN and attention mechanism makes up for its inability to better capture and retain important feature information in the data sequence, effectively improves the accuracy of system health assessment, and verifies the effectiveness and superiority of the health assessment method of benzene ratio control system based on C-WGAN-GP and ACNN deep learning model mentioned in this paper.

[0139] In the above step S5, according to the system benzene-olefin ratio output prediction value obtained in the above step S4, the health assessment of the target benzene-olefin ratio control system is achieved; specifically, the system benzene-olefin ratio output value obtained in the above step S1 and the system benzene-olefin ratio output prediction value obtained in the above step S4 are compared to achieve a reliable assessment of the health status of the target benzene-olefin ratio control system;

[0140] In addition, it also includes: drawing a contribution graph based on the attention weight of the feature vector in the Attention mechanism layer in the ACNN health assessment model, quantitatively analyzing the nonlinear relationship between the health status of the target benzene-olefin ratio control system and the characteristic factors, and effectively tracing the non-optimal factors that cause the health level of the target benzene-olefin ratio control system to decline.

[0141] Figure 8 The contribution diagram of 7 characteristic factors to the health status of the system under 5 common operating conditions is given. Among them, the X-axis represents the 7 types of characteristic factors that affect the health status of the system, the Y-axis represents the 5 common operating conditions of the system, and the Z-axis represents the weight value of each characteristic factor under different operating conditions.

[0142] from Figure 8It can be seen that the weight of each characteristic factor can intuitively characterize its contribution rate to the health status of the system. Taking the T2 working condition as an example, among the contribution rates of each characteristic under this working condition, the contribution rate of the benzene loop performance to the health status of the system is the highest, that is, the benzene loop performance is considered to be the main factor causing the change in the health level of the system. Combined with the mechanism knowledge, benzene material is the main material of the ratio control system. When the performance of the benzene loop decreases, it directly affects the production capacity of the underlying ratio control system, so that the top-level MPC regulation performance is affected, which is consistent with the second highest contribution rate of MPC performance under the T2 working condition in the figure, thereby causing changes in the health level of the system; in addition, when the health status of the system deteriorates with the change of the working condition, the contribution rate of the follow-up index and the MPC performance index characterizing the coupling performance of the system gradually increases, which is also consistent with the control mechanism in the actual system. In summary, the present invention introduces the ACNN health assessment model of the attention mechanism. When the current system health status is given, due to the special setting of the attention module, the cause of the fluctuation of the health level of the system can be traced according to the contribution of the weight coefficients of each characteristic quantity of the attention module, and then the decision-making process of the evaluation model is explained, which provides a scientific basis for system intelligent decision-making and maintenance.

[0143] In summary, the present invention aims at a type of complex control system of MPC-PID double-layer architecture of benzene-olefin ratio in chemical production process, and proposes a health assessment method of benzene-olefin ratio control system based on class imbalanced data enhancement; for the first time, the original imbalanced data is expanded and enhanced with the help of C-WGAN-GP data enhancement model to obtain a complete data set for evaluation, thereby improving the accuracy of system health assessment; the deep learning assessment method of ACNN based on the pre-assessment results of traditional multivariate statistical methods determines the characteristic factors that significantly affect the health status of the system, and uses CNN to deeply mine the correlation characteristics between system levels and between units within the level, avoiding the shortcomings of traditional linear weighted fusion on the coupling, nonlinearity and being subject to prior knowledge of the health status of hierarchical subsystems; at the same time, the health assessment model based on ACNN can not only reliably assess the health status of the system, but also trace the cause of the change in the health level of the system through the influence of the weights of each feature in the attention layer in the model on the health level, providing a scientific basis for system health maintenance.

[0144] The benzene-olefin ratio control system health assessment method based on class-imbalanced data enhancement provided by the present invention has been verified to be effective on an operating data set under actual operating conditions of the benzene-olefin ratio control system. Experiments show that for a complex MPC-PID two-layer architecture control system, the method provided by the present invention effectively improves the accuracy of system health status assessment when there is class-imbalance in the health status of the assessment data, and has obvious advantages over other algorithms.

[0145] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the embodiments can be referred to each other. The above description of the disclosed embodiments enables professionals and technicians in this field to implement or use the present invention. Various modifications to these embodiments will be obvious to professionals and technicians in this field, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown in this article, but will comply with the widest range consistent with the principles and novel features disclosed herein.

Claims

1. A health assessment method for benzene-olefin ratio control system based on class imbalance data enhancement, characterized in that: The steps include: S1. Obtaining operation data and system benzene-olefin ratio output values ​​of a target benzene-olefin ratio control system under a plurality of preset working conditions, and preprocessing the operation data and the system benzene-olefin ratio output values ​​respectively; S2. Based on the pre-processed operating data and the system benzene-olefin ratio output value, a plurality of characteristic factors related to the health status of the target benzene-olefin ratio control system are obtained; S3, selecting characteristic quantity data corresponding to the plurality of characteristic factors from the operation data as real class imbalance data indicating uneven distribution of health status of the target benzene-olefin ratio control system; Inputting the real class imbalanced data into the C-WGAN-GP data augmentation model, and outputting generated data; S4, taking the real class imbalance data and the generated data as a complete evaluation data set, inputting the complete evaluation data set into the constructed ACNN health assessment model, and outputting the system benzene-olefin ratio output prediction value; S5. Implement health assessment of the target benzene-olefin ratio control system according to the system benzene-olefin ratio output prediction value; and trace the non-optimal factors that cause the health level of the target benzene-olefin ratio control system to decline through the attention mechanism in the ACNN health assessment model; The step S4 specifically includes: S41, inputting the complete assessment data set into the ACNN health assessment model through the input layer; S42, extracting features from the complete evaluation data set through a CNN convolutional layer to obtain convolutional features; S43, sampling the convolutional features through a CNN pooling layer to obtain a CNN pooling layer output; S44, processing and calculating the output of the CNN pooling layer through the hidden layer to obtain the intermediate state of the feature factor; obtaining the similarity between the intermediate state feature vector and the hidden state through the Attention mechanism layer; specifically: The output results of the CNN pooling layer are processed through the hidden layer to obtain the intermediate state of the feature factors; through the Attention mechanism layer, the function score([h t,i ,h t ]) Calculate the intermediate state feature vector ht,i and the hidden state h t The similarity between them, the expression of the Attention mechanism layer is as follows: Among them, W s represents the weight matrix; b s represents the bias vector; S45, obtaining the attention weight of the feature factor hidden layer vector through the softmax function, and performing weighted sum processing on the attention weight and the intermediate state feature vector to obtain the output of the Attention mechanism layer; specifically: After the softmax function, the attention weight α of the hidden layer vector of the feature factor in the Attention mechanism layer is obtained i , and the feature vector h of the intermediate state t,i Weighted summation to get the output of the Attention mechanism layer Its expression is as follows: Where, τ is the output node of the fully connected layer; S46, passing the output of the attention mechanism layer through a fully connected layer to obtain a system benzene-olefin ratio output prediction value representing a health state of the system; S47, outputting the system benzene-olefin ratio output prediction value through an output layer; The step S5 further comprises: According to the weight adaptively assigned to each characteristic factor by the Attention mechanism layer in the ACNN health assessment model, a contribution diagram is drawn, and the nonlinear relationship between the health status of the target benzene-olefin ratio control system and the characteristic factors is quantitatively analyzed, so as to effectively trace the non-optimal factors that cause the decline in the health level of the target benzene-olefin ratio control system.

2. The benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement according to claim 1 is characterized in that: In the step S1, the operation data and the system benzene-olefin ratio output value are preprocessed respectively, specifically including: Performing denoising processing on the operation data; The Min-Max interval mapping normalization method is adopted to map the benzene-olefin ratio output value of the system between [0,1].

3. The benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement according to claim 1 is characterized in that: The step S2 specifically includes: S21, based on the pre-processed operation data, pre-evaluate each hierarchical unit in the benzene-olefin ratio control system according to preset evaluation indicators: the preset evaluation indicators include the hierarchical unit evaluation indicators of the benzene-olefin ratio control system and the follow-up performance evaluation indicators of each loop in the benzene-olefin ratio control system; S22. According to the preliminary evaluation result, a plurality of characteristic factors whose correlation with the system benzene-olefin ratio output value exceeds a preset value are determined.

4. The benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement according to claim 3 is characterized in that: The multiple characteristic factors include: a benzene loop, four ethylene loops, a performance evaluation index of MPC, and a follow-up index characterizing loop coupling properties.

5. The benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement according to claim 3 is characterized in that: The step S2 further comprises: S23, numbering the multiple characteristic factors, and using a mutual information heat map method to analyze the nonlinear relationship between every two characteristic quantities.

6. The benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement according to claim 1 is characterized in that: In the C-WGAN-GP data augmentation model: The generator adopts a deconvolution network structure, which consists of 1 fully connected layer and 3 deconvolution layers; the input of the generator is random noise that conforms to the Gaussian distribution; the fully connected layer samples from the random noise; the deconvolution layer learns the distribution of characteristic quantities in the benzene-olefin ratio control process; the output of the generator is 1025*7-dimensional data; The discriminator adopts a convolution structure, which is composed of 3 convolution layers and 1 fully connected layer; the convolution layer is used to extract the dynamic characteristic information of the characteristic quantity curve in the benzene-olefin ratio control process; the fully connected layer is used to discriminate the probability.

7. The benzene-ethylene ratio control system health assessment method based on class imbalance data enhancement according to claim 6 is characterized in that: The generator loss function in the C-WGAN-GP data augmentation model is expressed as: L(G)=1-D(G(z)) Where L(G) represents the generator loss function of the C-WGAN-GP data augmentation model; D(·) and G(·) represent the discriminator function and generator function, respectively; The discriminator loss function in the C-WGAN-GP data augmentation model is expressed as: Where L(D) represents the discriminator loss function of the C-WGAN-GP data augmentation model; represents the gradient operator; λ represents the penalty term coefficient; represents the random interpolation between the real class imbalanced data and the generated data, ζ follows a uniform distribution in the range [0, 1]; Represents the imbalanced data distribution P from the real class data (.) and generate data distribution P g (.) Straight line uniform sampling between sampling points.

Images