CN102680646A - Method of soft measurement for concentration of reactant in unsaturated polyester resin reacting kettle - Google Patents

Abstract

The invention relates to a method of soft measurement for concentration of reactant in an unsaturated polyester resin reacting kettle. The method uses a data acquisition module based on OPC (object linking and embedding for process control) technology, an abnormal data eliminating and pretreating module, a data normalization pretreating module, a digital filtering and wavelet treating module, a main factor analyzing module, and a self-adapting genetic algorithm optimized BP (back propagation) neural network soft measurement module. The method with high measurement precision for the soft measurement of the concentration of reactant in the reacting kettle during a batch process is provided, and the difficulty in measuring the reactant concentration in a batch reacting kettle is solved.

Description

Soft measurement method for concentration of reactants in unsaturated polyester resin reaction kettle

Technical Field

The invention relates to the field of industrial process control soft measurement, in particular to a soft measurement method for the concentration of reactants in an unsaturated polyester resin reaction kettle in an intermittent reaction process.

Background

Unsaturated Polyester Resins (UPR) generally refer to polyesters having carbon-carbon Unsaturated bonds (e.g., double bonds) in the molecular chain, and are prepared by polycondensation of Unsaturated dibasic acids, saturated dibasic alcohols, and polyhydric alcohols. Due to the excellent technological properties, electrical properties, corrosion resistance, mechanical properties and the like of the UPR, the UPR is widely applied to various aspects in the fields of buildings, chemical engineering, electrical appliances, medicines and the like, wherein the UPR is most typically a glass fiber reinforced material (commonly called as glass fiber reinforced plastics), the polyester in the glass fiber reinforced plastics accounts for 70-80% of the total amount, and various commodity varieties such as general-purpose resin, heat-resistant resin, high-strength resin and the like are formed.

To accommodate the needs of the modern market, the process industry is gradually moving towards batch process development. The batch type reaction kettle has the advantages of low investment, flexible production and high added value, and is widely applied to the fields of medicine production, fine chemical engineering and the like. However, in the current application, the automation level of batch production is still slow compared with that of continuous production, and the main reasons are as follows: for some parameters such as reactant or product concentration, an economical and reliable sensor is lacked to detect and measure the parameters on line, and even some sensors are developed, the cost is high, the practicability and the reliability are very limited, and the actual industrial requirements cannot be met. Therefore, it is necessary to establish an online real-time soft measurement model for the reactant concentration of the reaction kettle in the production of unsaturated polyester resin to meet the actual industrial production requirements.

In recent years, research on soft measurement techniques has become a focus. The soft measurement model is the core of the soft measurement technology, and mainly carries out optimal precision estimation on a dominant variable through an auxiliary variable, and the current soft measurement modeling method mainly comprises the following methods: a mechanism analysis method, a regression analysis method, a state estimation method, and an artificial neural network method. The mechanism analysis method is established on the basis of deep understanding of the process mechanism of a process object, and the model has poor transportability due to the specificity; the modeling difficulty is high, and the intrinsic kinetics of reaction, a heat and mass transfer equation and the like need to be started; solving complex models composed of algebraic equations and differential equations is difficult. Regression analysis methods are generally used for those with few auxiliary variables and are less accurate.

A soft measurement technology based on an Artificial Neural Network (ANN) is a soft measurement method which is most studied in the recent years and has the widest application field. The ANN method takes simple nonlinear neurons as basic units, forms a distributed nonlinear network system through large-scale connection, and has certain advantages in solving serious uncertainty and nonlinear system control by means of the functional characteristics of associative memory, self-learning, nonlinear approximation and the like. Because the mechanism of the industrial process is very complex, the characteristics of high nonlinearity and uncertainty exist, and a more accurate soft measurement model is difficult to establish by using a traditional method. Under the condition that the ANN method does not have object prior knowledge, the auxiliary variable can be used as the input of the ANN, the main variable is used as the output, the soft measurement model can be established through network learning, and even if the characteristics of a modeling object are continuously changed, the ANN soft measurement model can be corrected in time through learning, so that the ANN becomes a main means and method for soft measurement, and an effective method is provided for solving the problem of the soft measurement difficulty of the complex industrial process parameters.

The most commonly used in soft measurement modeling is the contemporary BP network, and in particular, the function approximation function, which is one of the strongest application functions of the ANN, is mainly applied. Although BP networks have been widely used in many ways, they also have disadvantages: firstly, since the learning rate is constant, it takes a long training time and the convergence speed is slow; secondly, the BP algorithm only can obtain a local optimum value by adopting a gradient descent method, and cannot obtain a global minimum value by an error; again, the network topology can only be determined empirically or based on repeated experiments. The Genetic Algorithm (GA) is a global optimization Algorithm, which can effectively overcome the above-mentioned deficiency of BP, and from this point, the Genetic Algorithm is just a star of the disadvantage of the BP neural network, so combining the two algorithms, and optimizing the neural network by the Genetic Algorithm becomes a natural choice. The invention provides a soft measurement model established based on an Improved Adaptive Genetic Algorithm (IAGA) by using a method for optimizing BP neural network weight by using a Genetic Algorithm, which has higher precision and higher speed and can be completely and better applied to soft measurement work.

Disclosure of Invention

In order to overcome the defects of the BP network in soft measurement: the learning rate is constant, and the convergence speed is slow due to the fact that long training time is spent; only local optimum values can be obtained; the network topology can only be determined empirically. The invention optimizes BP neural network weight by genetic algorithm, and provides a soft measurement model method established based on improved adaptive genetic algorithm, which is used for measuring the concentration of reactants in an unsaturated polyester resin reaction kettle in the intermittent reaction process.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a soft measurement method for reactant concentration in an unsaturated polyester resin reaction kettle comprises an on-site intelligent instrument, a data storage device and an upper computer, wherein the on-site intelligent instrument, the data storage device and the upper computer are connected with an object in a production process of the intermittent reaction kettle and are used for storing historical data, the on-site intelligent instrument, the data storage device and the upper computer are sequentially connected, the upper computer is a soft measurement intelligent processor, and the soft measurement intelligent processor comprises a data acquisition module, an abnormal data rejection preprocessing module, a data normalization preprocessing module, a digital filtering wavelet processing module, a main factor analysis module and a self-adaptive genetic algorithm optimization BP neural network soft measurement module based on an OPC technology.

The data acquisition module of the OPC technology is realized by the following method:

the method comprises the steps of taking MATLAB as an OPC client, taking WinCC as an OPC server, and collecting field process data of an intermittent reaction kettle in real time, wherein a communication structure block diagram of WinCC and MATLAB based on OPC is shown in figure 1, WinCC has the function of reading and writing MATLAB internal variables, MATLAB provides a special interface for OPC, an OPC object is configured, and real-time data collection is started.

The abnormal data elimination preprocessing module is realized by the following method:

for a set of data samples repeatedly measured, the size difference between individual data and other data of the same group is obvious, the probability of the data containing gross errors is large, and whether the data is suspicious abnormal data needs to be judged through a criterion. Adopting 3 sigma criterion (three-sigma code criterion) as criterion for judging gross error, i.e. adopting 3 sigma as criterion

<math> <mrow> <mo>|</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <mover> <mi>X</mi> <mo>&OverBar;</mo> </mover> <mo>|</mo> <mo>></mo> <mn>3</mn> <mi>&sigma;</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.1</mn> <mo>)</mo> </mrow> </mrow> </math>

Determine data x_i(i ═ 1, 2, …, n) is anomalous data that must be culled, where

Is the arithmetic mean of n data and σ is the mean square error or standard deviation. In order to ensure the continuity of the data, the removed vacancy must be filled after the data is removed. The change of the physical quantity at any site is continuous and smooth from small to large or from large to small, and can be calculated and filled by using a first-order difference equation, wherein the calculation equation is as follows:

x_k′＝x_k-1+(x_k-1-x_k-2) (1.2)

in the formula, x_k-1Is the sampled value at time k-1, x_k-2Is the sampled value at time k-2, x_k' is an estimated value at time k.

The data normalization preprocessing module is realized by the following method:

the normalization (normalization) of the data is to convert all the data into the same range, and the mean value and standard deviation of each variable, which are respectively 0 and 1, can be obtained, thereby improving the precision of data processing. Firstly, the two methods are assumed to be in a simple linear relationship, that is, the change of the actual value of the index can correspondingly cause a linear change of the normalized value, and the following methods are several common linear normalization methods:

(1) standard deviation normalization

The standard deviation normalization can obtain the mean value and the standard deviation of each variable respectively being 0 and 1, and the corresponding formula is:

<math> <mrow> <msub> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>k</mi> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>-</mo> <mover> <mi>x</mi> <mo>&OverBar;</mo> </mover> </mrow> <mi>s</mi> </mfrac> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>1,2</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.3</mn> <mo>)</mo> </mrow> </mrow> </math>

wherein, <math> <mrow> <mover> <mi>x</mi> <mo>&OverBar;</mo> </mover> <mo>=</mo> <mfrac> <mn>1</mn> <mi>n</mi> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> </mrow> </math> <math> <mrow> <mi>s</mi> <mo>=</mo> <msqrt> <mfrac> <mn>1</mn> <mrow> <mi>n</mi> <mo>-</mo> <mn>1</mn> </mrow> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msup> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <mi>x</mi> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> </mrow> </math>

(2) polar difference normalization

The maximum value and the minimum value of each variable obtained by normalizing the range are respectively 1 and 0, and the corresponding formula is as follows:

<math> <mrow> <msub> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>k</mi> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>-</mo> <munder> <mrow> <mi>min</mi> <mo>{</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>}</mo> </mrow> <mrow> <mn>1</mn> <mo>&le;</mo> <mi>i</mi> <mo>&le;</mo> <mi>n</mi> </mrow> </munder> </mrow> <mrow> <munder> <mrow> <mi>max</mi> <mo>{</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>}</mo> </mrow> <mrow> <mn>1</mn> <mo>&le;</mo> <mi>i</mi> <mo>&le;</mo> <mi>n</mi> </mrow> </munder> <mo>-</mo> <munder> <mrow> <mi>min</mi> <mo>{</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>}</mo> </mrow> <mrow> <mn>1</mn> <mo>&le;</mo> <mi>i</mi> <mo>&le;</mo> <mi>n</mi> </mrow> </munder> </mrow> </mfrac> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>1,2</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.4</mn> <mo>)</mo> </mrow> </mrow> </math>

(3) data hundred differentiation

The data hundred differentiation is to calculate the percentage of each variable, the sum of the processed variables is 1, and the corresponding formula is as follows:

<math> <mrow> <msub> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>k</mi> </msub> <mo>=</mo> <mfrac> <msub> <mi>x</mi> <mi>k</mi> </msub> <mrow> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msub> <mi>x</mi> <mi>i</mi> </msub> </mrow> </mfrac> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>1,2</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.5</mn> <mo>)</mo> </mrow> </mrow> </math>

(4) vector normalization

The data hundred differentiation is to calculate the percentage of each variable, the sum of the processed variables is 1, and the corresponding formula is as follows:

<math> <mrow> <msub> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>k</mi> </msub> <mo>=</mo> <mfrac> <msub> <mi>x</mi> <mi>k</mi> </msub> <msqrt> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msubsup> <mi>x</mi> <mi>i</mi> <mn>2</mn> </msubsup> </msqrt> </mfrac> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>1,2</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.6</mn> <mo>)</mo> </mrow> </mrow> </math>

the digital filtering wavelet processing module is realized by the following method:

in consideration of the characteristics of wavelet filtering, the filtering of the detection signals of the batch process reaction kettle system is real-time filtering, so the invention selects wavelet threshold filtering for analysis and processing. The method mainly comprises the following steps:

step 1: wavelet decomposition is performed.

The proper wavelet and the decomposition layer number are selected, and the orthogonal wavelet transform is carried out on the signal containing the noise by using a Mallat fast algorithm, namely, the decomposition is carried out by using an equation (1.7) and an equation (1.8). Fig. 2 shows a wavelet decomposition process with a layer number of three, and noise components are contained in cD1, cD2, and cD 3.

<math> <mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>n</mi> </munder> <mi>h</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>0,1</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.7</mn> <mo>)</mo> </mrow> </mrow> </math>

<math> <mrow> <msub> <mi>d</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>n</mi> </munder> <mi>g</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>0,1</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.8</mn> <mo>)</mo> </mrow> </mrow> </math>

In the formula, c_j，kIs a scale coefficient, j is the number of decomposition layers, n is the number of discrete sampling points, d_j，kH (-) g (-) is a pairwise quadrature mirror filter bank.

Step 2: the wavelet coefficients are processed.

For small after decompositionWave coefficient d_j，kPerforming nonlinear processing, only keeping wavelet coefficient with absolute value greater than set threshold, eliminating others to set as 0, and finally obtaining processed wavelet coefficient d'_j，k。

And 3, step 3: and (5) signal reconstruction.

And (3) reconstructing the wavelet coefficient after threshold processing according to the formula (1.9) to obtain an estimated value of the original signal, namely the optimal estimated value corresponding to the original signal.

<math> <mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>k</mi> </munder> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mi>h</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <mo>+</mo> <munder> <mi>&Sigma;</mi> <mi>k</mi> </munder> <msub> <msup> <mi>d</mi> <mo>&prime;</mo> </msup> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mi>g</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.9</mn> <mo>)</mo> </mrow> </mrow> </math>

The main factor analysis module is realized by the following method:

the method for converting a plurality of variables with correlation into a few independent variables by the aid of the main factor analysis aims to reduce the dimension of a high-dimensional space into a low-dimensional space or find information capable of representing an original data matrix in a modeling data matrix, and achieves the purposes of data simplification compression, singular value detection, variable screening and the like. The specific process is as follows:

(1) for the input data matrix X ═ X₁，x₂，…x_n)^TAnd standardizing it

<math> <mrow> <mover> <mi>X</mi> <mo>~</mo> </mover> <mo>=</mo> <mrow> <mo>(</mo> <mi>X</mi> <mo>-</mo> <mover> <mi>X</mi> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> <mo>/</mo> <msqrt> <msub> <mi>D</mi> <mi>&sigma;</mi> </msub> </msqrt> <mo>,</mo> </mrow> </math> <math> <mrow> <msub> <mi>D</mi> <mi>&sigma;</mi> </msub> <mo>=</mo> <mi>diag</mi> <mrow> <mo>(</mo> <msubsup> <mi>&sigma;</mi> <mn>1</mn> <mn>2</mn> </msubsup> <mo>,</mo> <msubsup> <mi>&sigma;</mi> <mn>2</mn> <mn>2</mn> </msubsup> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <msubsup> <mi>&sigma;</mi> <mi>n</mi> <mn>2</mn> </msubsup> <mo>)</mo> </mrow> <mo>,</mo> </mrow> </math> Is the mean value of X;

is x_iThe variance of (c).

(2) Computing

Covariance matrix of

(3) For matrix

Performing feature decomposition

Wherein L ═ L₁，l₂，…，l_n)，l₁，l₂，…，l_nIs the corresponding regularized feature vector;

Λ＝diag(λ₁，λ₂，…，λ_n)，λ₁≥λ₂≥…≥λ_nnot less than 0 isThe characteristic value of (2).

(4) Calculating the cumulative variance contribution rate n of the first p principal elements_pWherein

p≤n

The general approach is to remove p smaller so that the cumulative variance contribution of p principal components does not fall below a certain level (n is generally required_p＞0.85)。

The self-adaptive genetic algorithm optimized BP neural network soft measurement module is realized by the following method:

for p in adaptive genetic algorithm_cAnd p_mThe selection of the adaptive genetic algorithm is not simply changed linearly according to the fitness value f, but a nonlinear design is added, and an improved adaptive genetic algorithm based on an adaptive crossover and mutation algorithm is proposed and designed, and the method comprises the following specific steps:

$p_c=\left\{\begin{array}{c}{p}_{c1}-\frac{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">P</figure-callout>}}{1+\exp (-P)}(p_{c1}-p_{c2})\\ p_{c2}-\frac{\mathrm{<figure-callout\; id="1"\; label="Q"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">Q</figure-callout>}}{1+\exp (-Q)}(p_{c2}-p_{c3})\end{array}\right.---\left(1.10\right)$

$p_m=\left\{\begin{array}{c}{p}_{m1}-\frac{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">R</figure-callout>}}{1+\exp (-R)}(p_{m1}-p_{m2})\\ p_{m2}-\frac{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">S</figure-callout>}}{1+\exp (-S)}(p_{m2}-p_{m3})\end{array}\right.---\left(1.11\right)$

in the formula, <math> <mrow> <mi>P</mi> <mo>=</mo> <mfrac> <mrow> <msup> <mi>f</mi> <mo>&prime;</mo> </msup> <mo>-</mo> <msub> <mi>f</mi> <mi>avg</mi> </msub> </mrow> <mrow> <msub> <mi>f</mi> <mi>max</mi> </msub> <mo>-</mo> <msub> <mi>f</mi> <mi>avg</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math> <math> <mrow> <mi>Q</mi> <mo>=</mo> <mfrac> <mrow> <msup> <mi>f</mi> <mo>&prime;</mo> </msup> <mo>-</mo> <msub> <mi>f</mi> <mi>min</mi> </msub> </mrow> <mrow> <msub> <mi>f</mi> <mi>avg</mi> </msub> <mo>-</mo> <msub> <mi>f</mi> <mi>min</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math> $R=\frac{f-f_{\mathrm{avg}}}{f_{\max }-f_{\mathrm{avg}}};$ $S=\frac{f-f_{\mathrm{avg}}}{f_{\max }-f_{\mathrm{avg}}};$

f_maxthe maximum individual fitness in the population; f. of_avgThe average individual fitness of the population; f' is the greater fitness of the two individuals to be crossed; p is a radical of_c1＝0.9，p_c2＝0.6，p_c3＝0.5，p_m1＝0.1，p_m2＝0.01，p_m2＝0.008。

FIG. 3 is a flow chart of an improved adaptive genetic algorithm IAGA optimization BP neural network algorithm. The genetic algorithm optimization BP neural network process is divided into three parts, namely determination of a BP neural network structure, optimization of an improved adaptive genetic algorithm IAGA (inertial navigation algorithm) and prediction of soft measurement modeling. The BP neural network structure determination is to determine a network topological structure according to the number of input and output parameters of soft measurement modeling data, and further determine the individual length of a soft color body of an IAGA algorithm; the genetic algorithm IAGA optimizing module optimizes the connection weight and threshold of the BP neural network by using GA, all the weights and thresholds of the network are contained in individuals in a population, and an individual fitness value is calculated through a fitness function, and then the individuals with the optimal fitness value are found through selection, intersection and variation operations to provide the optimal weight and threshold for a third module, so that the second module is an intermediate transition module; the soft measurement modeling prediction is to assign optimal individuals obtained by a genetic algorithm to an initial neural network weight and a threshold value, and finally predict target output after the network is trained.

(1) The specific flow of the BP neural network (BPNN) algorithm refers to FIG. 3.

1) The reaction kettle concentration soft measurement model has 4 input parameters (Q)_c，Q_o，T，T_c) 1 output parameter (C). The BP neural network consists of an input layer, a hidden layer and an output layer, wherein the hidden layer adopts a double hidden layer structure, and the double hidden layer has better normalization performance and higher prediction precision than a single hidden layer. Namely, the neurons of the BPNN algorithm input layer, hidden layer 1, hidden layer 2 and output layer are 4: 3: 1, namely inputnum is 4, hidnum1 is 4, hidnum 2 is 3, outputnum is 1;

2) transfer function created by neural network

The hidden layer transfer functions of the first layer and the second layer are hyperbolic tangent sigmoid transfer functions (namely tansig), and the transfer function of the output layer is a linear transfer function (namely purelin);

3) BP network training function and learning function of weight and threshold

The BP network training function is selected as a Levenberg-Marquardt back propagation algorithm training function (Trainlm), and the network weight and threshold learning function is a gradient descent weight/threshold learning function (Learndm) with an additional momentum factor;

4) the maximum number of iterations epochs is 100;

5) the learning rate lr is 0.1;

6) target error E_goalIs 10^-5。

(2) The genetic algorithm optimization neural network (GA-BPNN) algorithm simulation is shown in a specific flow chart in FIG. 4, and simulation parameters are set as follows:

1) using real number coding

The individual codes adopt real number codes, each individual is a real number string and consists of a connection weight value of an input layer and a hidden layer 1, a connection weight value of the hidden layer 1 and a hidden layer 2, a connection weight value of the hidden layer 2 and an output layer and a threshold value of each network layer.

2) The total number of nodes numsum is calculated as follows:

numsum＝inputnum*hiddnum1+hiddnum1+hidnum1*hidnum2+hidnum2

+hidnum2*outputnum+outputnum＝(4*4+4)+(4*3+3)+(3*1+1)＝39

3) fitness function

Selecting the sum E of the absolute values of the errors of the predicted output and the expected output as an individual fitness value F, wherein the calculation formula is as follows:

<math> <mrow> <mi>F</mi> <mo>=</mo> <mi>k</mi> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <mi>abs</mi> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>o</mi> <mi>i</mi> </msub> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.12</mn> <mo>)</mo> </mrow> </mrow> </math>

where k is the gain coefficient, n is the number of network output nodes, y_iIs the expected output of the ith node of the BP neural network, o_iIs the predicted output of the ith node.

4) Selection operation

The example uses the roulette method, i.e. a selection strategy based on fitness scale, with a probability of selection p for each individual i_iIs composed of

f_i＝k/F_i (1.13)

<math> <mrow> <msub> <mi>p</mi> <mi>i</mi> </msub> <mo>=</mo> <mi>f</mi> <mo>/</mo> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <msub> <mi>f</mi> <mi>i</mi> </msub> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.14</mn> <mo>)</mo> </mrow> </mrow> </math>

In the formula, F_iFor the fitness value of the individual i, the fitness value is reciprocal before individual selection, k is a coefficient, and N is the number of population individuals, in view of the principle that the smaller the fitness value is, the better the fitness value is.

5) Crossover operation

Because real number coding is adopted, the crossing operation also adopts a real number crossing method, and the kth soft color body a_kAnd the first soft color body a_lThe operation method of interleaving at j bit is shown as (1.15), and the interleaving probabilities are respectively selected according to the formula (1.10).

$\left.\begin{array}{c}{a}_{\mathrm{kj}}=a_{\mathrm{kj}}*(1-b)+a_{\mathrm{lj}}*b\\ a_{\mathrm{lj}}=a_{\mathrm{lj}}*(1-b)+a_{\mathrm{kj}}*b\end{array}\right\}---\left(1.15\right)$

6) Mutation operation

Selecting the jth gene a of the ith individual_ijCarrying out mutation, wherein the mutation operation methods are shown as (1.16) and (1.17), and the mutation probabilities are respectively selected according to the formula (1.11).

<math> <mrow> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>=</mo> <mfenced open='{' close=''> <mtable> <mtr> <mtd> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>+</mo> <mrow> <mo>(</mo> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mi>max</mi> </msub> <mo>)</mo> </mrow> <mo>*</mo> <mi>f</mi> <mrow> <mo>(</mo> <mi>g</mi> <mo>)</mo> </mrow> </mtd> <mtd> <mi>r</mi> <mo>&GreaterEqual;</mo> <mn>0.5</mn> </mtd> </mtr> <mtr> <mtd> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>+</mo> <mrow> <mo>(</mo> <msub> <mi>a</mi> <mi>min</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mi>ij</mi> </msub> <mo>)</mo> </mrow> <mo>*</mo> <mi>f</mi> <mrow> <mo>(</mo> <mi>g</mi> <mo>)</mo> </mrow> </mtd> <mtd> <mi>r</mi> <mo>&lt;</mo> <mn>0.5</mn> </mtd> </mtr> </mtable> </mfenced> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>1.16</mn> <mo>)</mo> </mrow> </mrow> </math>

f(g)＝R*(1-g/G_max) (1.17)

In the formula, a_maxIs gene a_ijThe upper bound of (c); a is_minIs gene a_ijThe lower bound of (c); r is a random number; g is the current iteration number; g_maxThe maximum number of evolutions; r is [0, 1 ]]Random number of intervals.

7) And (3) setting other parameters: the population size sizepop was 10 and the maximum number of generations of inheritance was 10.

The invention has the beneficial effects that:

(1) the OPC technology is applied to field process data acquisition of an intermittent reaction kettle, data of a connected Wincc OPC server can be conveniently read and written through the OPCToolbox, and data exchange between an Matlab platform based on the OPC technology and WincC configuration software is realized. And after a series of data preprocessing such as abnormal data elimination, normalization processing, digital filtering and the like, interference components and singular values in the measured data are eliminated, and the precision of data measurement is improved. The principal component analysis method is used for carrying out dimensionality reduction on the field data, auxiliary variables of the soft measurement model are screened out, and a better data preprocessing method is provided for better soft measurement modeling of subsequent sections.

(2) The method has the advantages that the defects of BP neural network soft measurement modeling are overcome, a genetic algorithm is introduced to optimize the BP neural network soft measurement modeling, an improved self-adaptive genetic algorithm which is added with a nonlinear design idea and is based on a self-adaptive intersection and variation algorithm is provided and designed, the advantages of the BP neural network are combined to form a genetic algorithm, the BP neural network hybrid algorithm is combined, and a reactor concentration soft measurement model is established. The soft measurement model established by the BP neural network hybrid algorithm is optimized through the improved adaptive genetic algorithm, so that the accuracy is high, the generalization error is small, and the requirement of process control can be met.

Drawings

FIG. 1 is a block diagram of OPC-based communication structure between WinCC and MATLAB

FIG. 2 is a diagram of a multi-scale wavelet analysis

FIG. 3BP neural network algorithm flow chart

FIG. 4GA optimization BP neural network algorithm flow chart

FIG. 5WinCC reaction kettle temperature trend curve

FIG. 6 temperature collection curve of Matlab reaction kettle

FIG. 7 is a diagram of the effect of abnormal data culling

FIG. 8 is a diagram of exception signals being processed

FIG. 9 raw data before processing

FIG. 10 Effect after normalization

FIG. 11 is a graph of the filtering effect of different threshold selection rules

FIG. 12 original signal and wavelet filtered signal

FIG. 13 wavelet filtered noise signal

FIG. 14IAGA-BPNN learning prediction graph

FIG. 15AGA-BPNN learning error curve

FIG. 16IAGA-BPNN regression Effect

Detailed Description

The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

The invention relates to a soft measurement method for the concentration of reactants in an unsaturated polyester resin reaction kettle, which comprises the following steps:

step 1: a data acquisition system based on OPC technology is established, MATLAB is used as an OPC client, WinCC is used as an OPC server, and the field process data of the batch reactor are acquired in real time, wherein the specific acquired data are shown in a table 1-1.

TABLE 1-1 measurement parameters of batch reactor

The realization process of the data exchange of MATLAB and WinCC based on the OPC technology is as follows:

firstly, selecting monitoring configuration software as an OPC server, activating WinCC configuration software, and establishing the OPC server for an acquisition system.

Second, the names of all servers, group objects, and items are set.

Thirdly, in MATLAB, finding and positioning OPC data read-write objects established by using an OPC server, and establishing effective connection with the OPC server.

And fourthly, creating an OPC data read-write group object and newly creating a corresponding item.

Fifthly, the acquisition time, the sampling period and the data acquisition times of the data are set.

The dynamic exchange of data between Matlab and WinCC is realized by utilizing a Matlab tool box OPC Toolbox, and the specific setting steps are as follows:

(1) register Windows core component with command OPCregist: OPC communication interface component OPCcom ps.dll, OPC agent placeholder component OPCProxy.dll and OPC automation interface component OPCdaauto.dll etc.

(2) DCOM configuration is respectively carried out in the OPC server and the client, the network access node name of the OPC client is found, and the object successful connection between the OPC client and the OPC server is ensured.

(3) And (3) creating variables in the WinCC, activating an OPC (optical proximity correction) driver, and then setting running parameters of the WinCC to ensure that the OPC server is in a normal communication state.

(4) Creating an OPC client object and establishing connection with an OPC server, adding the object, the Group (Group) and the item (Items), setting corresponding attributes respectively, and finally reading the process data of the Items and displaying on line, wherein the main codes are as follows:

searching available OPC server on local computer

The hostnfo ═ opcserverinfo ('localhost')% obtains the server id of the server, the specification used and the manufacturer of the OPC server.

② verifying returned OPC detailed information

Serverid'; % assigns OPC Server ID information to AllServers

Creating OPC data access client end object

After acquiring the host name and the OPC server ID, the OPC server is connected with the host name and the OPC server ID, but an opcda (OPC data access) object corresponding to the server needs to be created, and the codes are as follows:

da ═ opcda ('localhost', opcserver

Fourthly, successfully establishing the connection between the Matlab and the OPC server

connect (da)% connection server

The execution results are as follows:

Summary of OPC Data Client Object：local/opcserver.wincc

ServerID：opcserver.wincc

Status：connected

Timeout：10senconds

from the execution result Status: connected can see that the OPC client and the server successfully establish connection.

Creating an OPC data access group object

grp＝addgroup(da，′Group1′)

Sixthly, adding the items into the group object

itmCollection＝additem(grp)

Seventhly, configuring attributes of OPC tool box

The OPC toolkit attributes are configured with the SET command and retrieved with the GET command.

logDuration ═ 5 × 60; % Collection time 5 min

logRate ═ 1; % sampling period 1s

numRecords ceil (logduration./logRate); % number of data collected

set(grp，′UpdateRate′，logRate，′RecordsToAcquire′，numRecords)；

Data from OPC server

start(grp)；

[logIDs，logVal，logQual，logTime，logEvtTime]＝getdata(grp，′double′)；

Ninthly drawing data figure

The data pattern is converted to a formatted data string according to a format tag (datatack) function.

plot(logTime，logVal)；

axis tight；

datetick(′x′，′keeplimits′)；

Fig. 5 is a temperature trend curve of a WinCC reaction kettle acquired on site, and the temperature acquisition curve of the Matlab reaction kettle obtained by data exchange between Matlab based on OPC technology and monitoring configuration software WinCC is shown in fig. 6.

Step 2: eliminating abnormal data, adopting 3 sigma criterion (three-sigma code criterion) as criterion for judging gross error, i.e. adopting 3 sigma as criterion

<math> <mrow> <mo>|</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <mover> <mi>X</mi> <mo>&OverBar;</mo> </mover> <mo>|</mo> <mo>></mo> <mn>3</mn> <mi>&sigma;</mi> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2.1</mn> <mo>)</mo> </mrow> </mrow> </math>

Determine data x_i(i ═ 1, 2, …, n) is anomalous data that must be culled, where

Is the arithmetic mean of n data and σ is the mean square error or standard deviation. In order to ensure the continuity of data, after the data is removed, the removed vacancy needs to be filled, and a first-order difference equation is used for calculating and filling, wherein the calculation equation is as follows:

x_k′＝x_k-1+(x_k-1-x_k-2) (2.2)

in the formula, x_k-1Is the sampled value at time k-1, x_k-2Is the sampled value at time k-2, x_k' is an estimated value at time k. The effect of the reaction kettle temperature T after 3 sigma criterion abnormal data elimination is shown in FIG. 7, and the corresponding processed abnormal signal is shown in FIG. 8.

And step 3: the data is normalized and preprocessed, the standard deviation is normalized to obtain the mean value and the standard deviation of each variable which are respectively 0 and 1, and the corresponding formula is as follows:

<math> <mrow> <msub> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>k</mi> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>-</mo> <mover> <mi>x</mi> <mo>&OverBar;</mo> </mover> </mrow> <mi>s</mi> </mfrac> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>1,2</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2</mn> <mo>.</mo> <mn>3</mn> <mo>)</mo> </mrow> </mrow> </math>

wherein, <math> <mrow> <mover> <mi>x</mi> <mo>&OverBar;</mo> </mover> <mo>=</mo> <mfrac> <mn>1</mn> <mi>n</mi> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> </mrow> </math> <math> <mrow> <mi>s</mi> <mo>=</mo> <msqrt> <mfrac> <mn>1</mn> <mrow> <mi>n</mi> <mo>-</mo> <mn>1</mn> </mrow> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msup> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <mi>x</mi> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> </mrow> </math>

the raw data before processing is shown in fig. 9, and the effect after the standard deviation normalization processing is shown in fig. 10.

And 4, step 4: the digital filtering wavelet processing comprises the following main steps:

(1) wavelet decomposition is performed.

The orthogonal wavelet transform is performed on the signal containing noise by using Mallat fast algorithm by selecting proper wavelet and decomposition layer number, namely, the decomposition is performed by using equation (2.3.7) and equation (2.3.8). Fig. 3 shows a wavelet decomposition process with a layer number of three, and noise is partially contained in cD1, cD2, and cD 3.

<math> <mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>n</mi> </munder> <mi>h</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>0,1</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2.4</mn> <mo>)</mo> </mrow> </mrow> </math>

<math> <mrow> <msub> <mi>d</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>n</mi> </munder> <mi>g</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>0,1</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2.5</mn> <mo>)</mo> </mrow> </mrow> </math>

In the formula, c_j，kIs a scale coefficient, j is the number of decomposition layers, n is the number of discrete sampling points, d_j，kH (-) g (-) is a pairwise quadrature mirror filter bank.

(2) The wavelet coefficients are processed.

For the decomposed wavelet coefficient d_j，kPerforming nonlinear processing, only keeping wavelet coefficient with absolute value greater than set threshold, eliminating others to set as 0, and finally obtaining processed wavelet coefficient d'_j，k。

(3) And (5) signal reconstruction.

And (4) reconstructing the wavelet coefficient after threshold processing according to the formula (2.3.9) to obtain an estimated value of the original signal, namely the optimal estimated value corresponding to the original signal.

<math> <mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>k</mi> </munder> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mi>h</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <mo>+</mo> <munder> <mi>&Sigma;</mi> <mi>k</mi> </munder> <msub> <msup> <mi>d</mi> <mo>&prime;</mo> </msup> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mi>g</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <mo>-</mo> <mo>-</mo> <mo>-</mo> <mrow> <mo>(</mo> <mn>2.6</mn> <mo>)</mo> </mrow> </mrow> </math>

Now, for the detected data signal of the batch process reaction kettle control system, db3 wavelet analysis is used to perform 5-layer decomposition, and then the noisy signal is filtered by the selection rules of rigrsure, sqtwolog and heursure threshold values, respectively, and the obtained filtering comparison effect graph is shown in fig. 11. Fig. 12 shows the original signal and the wavelet-filtered signal, and fig. 13 shows the noise signal (error curve) after wavelet filtering, which shows that the wavelet denoising effect is better, and the noise signal can be effectively filtered by using the wavelet analysis method, thereby improving the quality of the detected data signal in the batch process reaction kettle control system.

And 5: the main factor analysis, the reaction volume V, the reactant concentration C and the coolant flow rate Q are carried out on the original data_cAnd the outlet flow rate Q of the reaction kettle_oTemperature T of reaction vessel, temperature T of material entering reaction vessel_ATemperature T of coolant_cTemperature T at the inlet of the reactor_inForm a column vector X, where X ═ VQ_cQ_oTT_AT_inT_c]^TAnd performing dimension reduction treatment on the composite material, which comprises the following specific steps:

(1) for the input data matrix X ═ X₁，x₂，…x_n)^TAnd standardizing it

<math> <mrow> <mover> <mi>X</mi> <mo>~</mo> </mover> <mo>=</mo> <mrow> <mo>(</mo> <mi>X</mi> <mo>-</mo> <mover> <mi>X</mi> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> <mo>/</mo> <msqrt> <msub> <mi>D</mi> <mi>&sigma;</mi> </msub> </msqrt> <mo>,</mo> </mrow> </math> <math> <mrow> <msub> <mi>D</mi> <mi>&sigma;</mi> </msub> <mo>=</mo> <mi>diag</mi> <mrow> <mo>(</mo> <msubsup> <mi>&sigma;</mi> <mn>1</mn> <mn>2</mn> </msubsup> <mo>,</mo> <msubsup> <mi>&sigma;</mi> <mn>2</mn> <mn>2</mn> </msubsup> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <msubsup> <mi>&sigma;</mi> <mi>n</mi> <mn>2</mn> </msubsup> <mo>)</mo> </mrow> <mo>,</mo> </mrow> </math> Is the mean value of X;

is x_iThe variance of (c).

(2) Computing

Covariance matrix of

(3) For matrix

Performing feature decomposition

Wherein L ═ L₁，l₂，…，l_n)，l₁，l₂，…，l_nIs the corresponding regularized feature vector;

Λ＝diag(λ₁，λ₂，…，λ_n)，λ₁≥λ₂≥…≥λ_nnot less than 0 isThe characteristic value of (2).

(4) Calculating the cumulative variance of the first p principal elementsContribution ratio n_pWherein

p≤n

Since the main objective of principal component analysis is dimensionality reduction, the original variable x needs to be replaced by a few principal components without much information loss₁，x₂，…x_nFor subsequent analysis and processing. The general approach is to remove p smaller so that the cumulative variance contribution of p principal components does not fall below a certain level (n is generally required_p＞0.85)。

The principal component analysis results obtained according to the principal component analysis algorithm steps are shown in tables 2-3. Tables 2-3 are variance contribution ratio analysis tables, and it can be seen from the tables that the first principal component explains 32.9873%, the second principal component explains 29.9238%, the first two principal components explain 62.9111%, the third principal component explains 17.3398%, the first three principal components explain 80.2509%, the fourth principal component explains 8.1702%, the first four principal components explain 88.4211%, and n is a general requirement_pAnd the first four main components reach the requirements when the temperature is higher than 0.85.

TABLE 2-3 ANALYSIS TABLE OF VARIATION COVERY

Step 6: optimizing BP neural network soft measurement by self-adaptive genetic algorithm, and p in self-adaptive genetic algorithm_cAnd p_mThe selection of the adaptive genetic algorithm is not simply changed linearly according to the fitness value f, but a nonlinear design is added, and an improved adaptive genetic algorithm based on an adaptive crossover and mutation algorithm is proposed and designed, and the method comprises the following specific steps:

$p_c=\left\{\begin{array}{c}{p}_{c1}-\frac{\mathrm{<figure-callout\; id="1"\; label="P"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">P</figure-callout>}}{1+\exp (-P)}(p_{c1}-p_{c2})\\ p_{c2}-\frac{\mathrm{<figure-callout\; id="1"\; label="Q"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">Q</figure-callout>}}{1+\exp (-Q)}(p_{c2}-p_{c3})\end{array}\right.---\left(6.1\right)$

$p_m=\left\{\begin{array}{c}{p}_{m1}-\frac{\mathrm{<figure-callout\; id="1"\; label="R"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">R</figure-callout>}}{1+\exp (-R)}(p_{m1}-p_{m2})\\ p_{m2}-\frac{\mathrm{<figure-callout\; id="1"\; label="S"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">S</figure-callout>}}{1+\exp (-S)}(p_{m2}-p_{m3})\end{array}\right.---\left(6.2\right)$

in the formula, <math> <mrow> <mi>P</mi> <mo>=</mo> <mfrac> <mrow> <msup> <mi>f</mi> <mo>&prime;</mo> </msup> <mo>-</mo> <msub> <mi>f</mi> <mi>avg</mi> </msub> </mrow> <mrow> <msub> <mi>f</mi> <mi>max</mi> </msub> <mo>-</mo> <msub> <mi>f</mi> <mi>avg</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math> <math> <mrow> <mi>Q</mi> <mo>=</mo> <mfrac> <mrow> <msup> <mi>f</mi> <mo>&prime;</mo> </msup> <mo>-</mo> <msub> <mi>f</mi> <mi>min</mi> </msub> </mrow> <mrow> <msub> <mi>f</mi> <mi>avg</mi> </msub> <mo>-</mo> <msub> <mi>f</mi> <mi>min</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math> $R=\frac{f-f_{\mathrm{avg}}}{f_{\max }-f_{\mathrm{avg}}};$ $S=\frac{f-f_{\mathrm{avg}}}{f_{\max }-f_{\mathrm{avg}}};$

f_maxthe maximum individual fitness in the population; f. of_avgThe average individual fitness of the population; f' is the greater fitness of the two individuals to be crossed; p is a radical of_c1＝0.9，p_c2＝0.6，p_c3＝0.5，p_m1＝0.1，p_m2＝0.01，p_m2＝0.008。

The standard genetic algorithm has the following basic steps:

1) selecting a coding method, and correspondingly converting the parameter set X and the domain into a space S with a bit string structure;

2) determining a suitable fitness function f (x);

3) selection and determination of genetic strategies, including selection of population size n, selection, crossover, mutation operators and methods and corresponding determination of crossover and mutationProbability p_c、p_mAnd other genetic control parameters;

4) continuously initializing the group P at random;

5) solving adaptive function values f (X) of individuals in the population;

6) evolving a new generation of population through operations such as selection, crossover and mutation operators;

7) and judging whether the group performance can meet the preset target requirement or not, or judging whether the preset iteration times are finished or not, if not, returning to the step 6), and if so, ending the algorithm process.

The neural network learning prediction curve is shown in fig. 14, the corresponding learning error curve graph 15 is shown, and table 3-2 is the comparison of the results of the actual reactor concentration value and the IAGA-BP predicted value. In the IAGA-BPNN, the optimal weight threshold is obtained by optimizing the BPNN by using the IAGA, the optimal weight threshold is learned by using the same BP algorithm, and after 31 iterations, the total error of the network reaches the MSE 6.1748 e-004. Therefore, the learning effect of the IAGA algorithm on the optimization of the neural network parameters is good, and the prediction error is small.

TABLE 3-2 comparison of the actual values of the reactor concentrations and the results of the predicted values of IAGA-BP

The hidden layer in the IAGA-BPNN is connected with the weight matrixes IW and LW1 and the threshold matrixes IB and LB1 as follows:

$\mathrm{IW}=\left[\begin{array}{cccc}0.5247& 1.2315& 0.3987& -1.3728\\ 0.0035& -0.0030& -0.2389& 0.0145\\ -0.1356& -0.4772& -1.9862& -1.3457\\ 1.2165& -0.5478& 0.3279& 1.3984\end{array}\right]$

$\mathrm{<figure-callout\; id="1"\; label="LW"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">LW</figure-callout>}1=\left[\begin{array}{cccc}-1.1753& -0.4825& -0.8447& 1.0279\\ 0.0621& 1.5045& 0.0071& 0.0002\\ -0.6443& 1.1623& -0.7849& 0.5482\end{array}\right]$

$\mathrm{IB}=\left[\begin{array}{c}-1.9939\\ 0.0318\\ -1.9430\\ 2.5644\end{array}\right]$ $\mathrm{<figure-callout\; id="1"\; label="LB"\; filenames="HSA00000703616400042.png"\; state="\{\{state\}\}">LB</figure-callout>}1=\left[\begin{array}{c}1.8393\\ -0.1314\\ -2.4288\end{array}\right]$

the output layer connection weight matrix LW2 and the threshold matrix LB2 are:

LW2＝[0.0194 2.8516-0.2441]LB2＝[0.1664]

the regression effect of the prediction result of the IAGA-BPNN algorithm is further obtained as shown in fig. 16, where R is 0.98649, and Output is 0.98 Target + -0.64, which illustrates the better fitting effect and better generalization performance of the fitting prediction of the IAGA-BPNN algorithm provided herein.

Claims (10)

1. The soft measurement method for the concentration of reactants in the unsaturated polyester resin reaction kettle comprises an on-site intelligent instrument, a data storage device and an upper computer, wherein the on-site intelligent instrument is connected with an object in the production process of the intermittent reaction kettle, the data storage device is used for storing historical data, the on-site intelligent instrument, the data storage device and the upper computer are sequentially connected, and the upper computer is a soft measurement intelligent processor and is characterized in that: the intelligent soft measurement processor comprises a data acquisition module based on an OPC technology, an abnormal data elimination preprocessing module, a data normalization preprocessing module, a digital filtering wavelet processing module, a main factor analysis module and a self-adaptive genetic algorithm optimization BP neural network soft measurement module.

2. The method for soft measurement of reactant concentration in an unsaturated polyester resin reaction kettle according to claim 1, wherein: the data acquisition module of the OPC technology takes MATLAB as an OPC client and WinCC as an OPC server to acquire data in real time from the field process data of the batch reactor.

3. The method for soft measurement of reactant concentration in an unsaturated polyester resin reaction kettle according to claim 1, wherein: the abnormal data elimination preprocessing module adopts a 3 sigma criterion (three sigma code criterion) as a criterion for judging gross errors, namely adopts 3 sigma as a criterion and meets the requirements

<math> <mrow> <mo>|</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <mover> <mi>X</mi> <mo>&OverBar;</mo> </mover> <mo>|</mo> <mo>></mo> <mn>3</mn> <mi>&sigma;</mi> </mrow> </math>

Determine data x_i(i ═ 1, 2, …, n) is anomalous data that must be culled, where

Is the arithmetic mean of n data and σ is the mean square error or standard deviation. And the first-order difference equation is used for calculating and filling, and the calculation equation is as follows:

x_k′＝x_k-1+(x_k-1-x_k-2)

in the formula, x_k-1Is the sampled value at time k-1, x_k-2Is the sampled value at time k-2, x_k' is an estimated value at time k.

4. The unsaturated polyester resin of claim 1The soft measurement method for the concentration of reactants in the reactor is characterized by comprising the following steps: the data normalization preprocessing module is used for normalizing (standardizing) the data, namely converting all the data into the same range, and obtaining the mean value and the standard deviation of each variable which are respectively 0 and 1. The concrete corresponding formula is as follows: <math> <mrow> <msub> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> <mi>k</mi> </msub> <mo>=</mo> <mfrac> <mrow> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>-</mo> <mover> <mi>x</mi> <mo>&OverBar;</mo> </mover> </mrow> <mi>s</mi> </mfrac> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>1,2</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>)</mo> </mrow> </mrow> </math>

wherein, <math> <mrow> <mover> <mi>x</mi> <mo>&OverBar;</mo> </mover> <mo>=</mo> <mfrac> <mn>1</mn> <mi>n</mi> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>,</mo> </mrow> </math> <math> <mrow> <mi>s</mi> <mo>=</mo> <msqrt> <mfrac> <mn>1</mn> <mrow> <mi>n</mi> <mo>-</mo> <mn>1</mn> </mrow> </mfrac> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <msup> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <mi>x</mi> <mo>)</mo> </mrow> <mn>2</mn> </msup> </msqrt> </mrow> </math>

5. the method for soft measurement of reactant concentration in an unsaturated polyester resin reaction kettle according to claim 1, wherein: the digital filtering wavelet processing module mainly comprises the following steps:

step 1: wavelet decomposition is performed.

Selecting proper wavelets and decomposition layer number, and carrying out orthogonal wavelet transform on the signal containing noise by using a Mallat fast algorithm, wherein the decomposition formula is as follows:

<math> <mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>n</mi> </munder> <mi>h</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>0,1</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow> </math>

<math> <mrow> <msub> <mi>d</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>n</mi> </munder> <mi>g</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mrow> <mo>(</mo> <mi>k</mi> <mo>=</mo> <mn>0,1</mn> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <mi>n</mi> <mo>-</mo> <mn>1</mn> <mo>)</mo> </mrow> </mrow> </math>

in the formula, c_j，kIs a scale coefficient, j is the number of decomposition layers, n is the number of discrete sampling points, d_j，kH (-) g (-) is a pairwise quadrature mirror filter bank.

Step 2: the wavelet coefficients are processed.

For the decomposed wavelet coefficient d_j，kPerforming nonlinear processing, only keeping wavelet coefficient with absolute value greater than set threshold, eliminating others to set as 0, and finally obtaining processed wavelet coefficient d'_i，k。

And 3, step 3: and (5) signal reconstruction.

And reconstructing the wavelet coefficient after threshold processing to obtain an estimated value of the original signal, namely the optimal estimated value corresponding to the original signal.

<math> <mrow> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>-</mo> <mn>1</mn> <mo>,</mo> <mi>n</mi> </mrow> </msub> <mo>=</mo> <munder> <mi>&Sigma;</mi> <mi>k</mi> </munder> <msub> <mi>c</mi> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mi>h</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> <mo>+</mo> <munder> <mi>&Sigma;</mi> <mi>k</mi> </munder> <msub> <msup> <mi>d</mi> <mo>&prime;</mo> </msup> <mrow> <mi>j</mi> <mo>,</mo> <mi>k</mi> </mrow> </msub> <mi>g</mi> <mrow> <mo>(</mo> <mi>n</mi> <mo>-</mo> <mn>2</mn> <mi>k</mi> <mo>)</mo> </mrow> </mrow> </math>

The main factor analysis module is a method for converting a plurality of variables with correlation into a few independent variables, and the specific process is as follows:

(1) for the input data matrix X ═ X₁，x₂，…x_n)^TAnd standardizing it

<math> <mrow> <mover> <mi>X</mi> <mo>~</mo> </mover> <mo>=</mo> <mrow> <mo>(</mo> <mi>X</mi> <mo>-</mo> <mover> <mi>X</mi> <mo>&OverBar;</mo> </mover> <mo>)</mo> </mrow> <mo>/</mo> <msqrt> <msub> <mi>D</mi> <mi>&sigma;</mi> </msub> </msqrt> <mo>,</mo> </mrow> </math> <math> <mrow> <msub> <mi>D</mi> <mi>&sigma;</mi> </msub> <mo>=</mo> <mi>diag</mi> <mrow> <mo>(</mo> <msubsup> <mi>&sigma;</mi> <mn>1</mn> <mn>2</mn> </msubsup> <mo>,</mo> <msubsup> <mi>&sigma;</mi> <mn>2</mn> <mn>2</mn> </msubsup> <mo>,</mo> <mo>.</mo> <mo>.</mo> <mo>.</mo> <mo>,</mo> <msubsup> <mi>&sigma;</mi> <mi>n</mi> <mn>2</mn> </msubsup> <mo>)</mo> </mrow> <mo>,</mo> </mrow> </math>

Is the mean value of X;

is x_iThe variance of (c).

(2) Computing

Covariance matrix of

(3) For matrix

Performing feature decomposition

Wherein L ═ L₁，l₂，…，l_n)，l₁，l₂，…，l_nIs the corresponding regularized feature vector;

Λ＝diag(λ₁，λ₂，…，λ_n)，λ₁≥λ₂≥…≥λ_nnot less than 0 is

The characteristic value of (2).

(4) Calculating the cumulative variance contribution rate n of the first p principal elements_pWherein

p≤n

The general approach is to remove p smaller so that the cumulative variance contribution of p principal components does not fall below a certain level (n is generally required_p＞0.85)。

6. The method for soft measurement of reactant concentration in an unsaturated polyester resin reaction kettle according to claim 1, wherein: the adaptive genetic algorithm optimizes the BP neural network soft measurement module, and p is subjected to the adaptive genetic algorithm_cAnd p_mInstead of simply changing linearly according to the fitness value f, a non-linear design is added, and the design is proposed and based on self-adaptationThe improved self-adaptive genetic algorithm of the cross and mutation algorithm is specifically as follows:

$p_c=\left\{\begin{array}{c}{p}_{c1}-\frac{P}{1+\exp (-P)}(p_{c1}-p_{c2})\\ p_{c2}-\frac{Q}{1+\exp (-Q)}(p_{c2}-p_{c3})\end{array}\right.$

in the formula, <math> <mrow> <mi>P</mi> <mo>=</mo> <mfrac> <mrow> <msup> <mi>f</mi> <mo>&prime;</mo> </msup> <mo>-</mo> <msub> <mi>f</mi> <mi>avg</mi> </msub> </mrow> <mrow> <msub> <mi>f</mi> <mi>max</mi> </msub> <mo>-</mo> <msub> <mi>f</mi> <mi>avg</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math> <math> <mrow> <mi>Q</mi> <mo>=</mo> <mfrac> <mrow> <msup> <mi>f</mi> <mo>&prime;</mo> </msup> <mo>-</mo> <msub> <mi>f</mi> <mi>min</mi> </msub> </mrow> <mrow> <msub> <mi>f</mi> <mi>avg</mi> </msub> <mo>-</mo> <msub> <mi>f</mi> <mi>min</mi> </msub> </mrow> </mfrac> <mo>;</mo> </mrow> </math> $R=\frac{f-f_{\mathrm{avg}}}{f_{\max }-f_{\mathrm{avg}}};$ $S=\frac{f-f_{\mathrm{avg}}}{f_{\max }-f_{\mathrm{avg}}};$

f_maxthe maximum individual fitness in the population; f. of_avgThe average individual fitness of the population; f' is the greater fitness of the two individuals to be crossed; p is a radical of_c1＝0.9，p_c2＝0.6，p_c3＝0.5，p_m1＝0.1，p_m2＝0.01，p_m2＝0.008。

7. The genetic algorithm optimized BP neural network soft measurement module according to claim 6, wherein: the genetic algorithm optimization BP neural network process is divided into three parts, namely determination of a BP neural network structure, optimization of an improved adaptive genetic algorithm IAGA (inertial navigation algorithm) and prediction of soft measurement modeling.

8. The genetic algorithm optimized BP neural network soft measurement module according to claim 6 or 7, wherein: the BP neural network structure is determined by determining a network topological structure according to the number of input and output parameters of soft measurement modeling data, and further determining the individual length of a soft color body of an IAGA algorithm.

9. The genetic algorithm optimized BP neural network soft measurement module according to claim 6 or 7, wherein: the genetic algorithm IAGA optimizing module optimizes the connection weight and threshold of the BP neural network by using GA, all the weights and thresholds of the network are contained in individuals in a population, and an individual fitness value is calculated through a fitness function, and then the individuals with the optimal fitness value are found through selection, crossing and mutation operations, so that the optimal weight and threshold are provided for a third module, and the second module is an intermediate transition module.

10. The genetic algorithm optimized BP neural network soft measurement module according to claim 6 or 7, wherein: the soft measurement modeling prediction is to assign optimal individuals obtained by a genetic algorithm to initial neural network weights and threshold values, and finally predict target output after the network is trained.

Images