CN108181812A - BP neural network-based valve positioner PI parameter setting method - Google Patents

Abstract

The invention discloses a method for setting PI parameters of a regulating valve positioner by using a BP neural network, which is characterized by searching for optimal PI parameters of the regulating valve under different specifications through an empirical rule and then obtaining the valve parameters of the regulating valve with different specifications through an offline test; the method can quickly and accurately set the PI parameters required by the regulating valve positioner, and is simple and convenient to operate.

Description

A PI parameter tuning method of valve positioner based on BP neural network

technical field

The invention belongs to the field of automatic control and relates to a control method, in particular to a method for adjusting PI parameters of a regulating valve positioner by using a BP neural network.

Background technique

The intelligent valve positioner plays an important role in the field of industrial control. It is the control core of the pneumatic control valve and can greatly improve the dynamic response and steady-state characteristics of the pneumatic valve. For the traditional PID controller, before it is put into operation, in order to obtain a more ideal control effect, three parameters must be adjusted: namely, the proportional coefficient (P), the integral (I), and the differential time (D). However, the actual industrial control process often has nonlinear and time-varying uncertainties, so it is difficult to establish an accurate mathematical model, and it is difficult to achieve the ideal control effect by using a conventional PID controller.

Contents of the invention

The purpose of the present invention is to aim at the deficiencies in the prior art, propose a kind of method based on BP neural network tuning positioner PI parameter, it synthesizes the advantage of traditional PID control theory and neural network theory.

For realizing the purpose of above-mentioned technology, reach above-mentioned technical effect, the present invention realizes through following technical scheme:

BP neural network is a multi-layer feed-forward neural network. The topological structure of 3-layer BP network includes input layer, output layer and a hidden layer. Each neuron is connected to all neurons in the next layer. Neurons in the same layer There is no connection between them. The basic principle of BP neural network is to use the gradient descent method to adjust the weight and threshold, so that the mean square error between the actual output value and the expected output value of the network is the smallest.

The standard BP algorithm does not consider the gradient direction of the previous moment when modifying the weight, so that the learning process often oscillates and converges slowly. Therefore, the present invention adopts the Levenberg-Marquardt algorithm to speed up the network training speed and improve the accuracy of the network training by optimizing the search direction of the BP neural network.

Levenberg-Marquardt algorithm, including the following:

1) Let w(i) represent the vector composed of the weight and threshold of the i-th iteration, and w(i+1) is the vector composed of the new weight and threshold, as shown in the formula: w(i+1 )=w(i)+Δw;

2) For Newton's method, the expression of Δw is: where E(w)

Let the evaluation function be the mean square error: Where e(w) is the error (i=1,2,3,...,N),

3) In the formula, J is the Jacobian matrix, and its form is:

4) For the Gauss-Newton law, there is ΔE=-[J ^T (w)*J(W)] ^-1 *J(w)e(w) LM algorithm is its improvement, then ΔE=-[J ^T (w) )*J(W)+μI] ^-1 *J(w)e(w) In the formula, the proportional coefficient μ>0 is a constant, and I is an identity matrix,

5) It can be seen from the above formula that if the proportional coefficient μ=0, the L-M algorithm becomes the Gauss-Newton method; if the value of μ is close to 1, the L-M algorithm is transformed into the gradient descent method, and as the number of successful iterations increases, The value of μ gradually decreases, and when the error is close to the minimum, the L-M algorithm gradually evolves into the Gauss-Newton method. Therefore, it has been shown through practice that the L-M algorithm can increase the speed by dozens or even hundreds of times compared with the original gradient descent method.

The determination of the input data of the neural network is the extraction of feature quantities (influence factors), mainly considering whether it has a clear causal relationship with the system output indicators. If the network input data has no connection with the output indicators, the connection between them cannot be established. Therefore, it is necessary to accurately analyze the factors that are strongly correlated with the output PI parameters and use them as input quantities.

Correlation analysis is performed between the collected input performance parameter sets and the manually optimized PI parameter sets. Judging the closeness of the correlation between variables needs to be judged by an indicator. Here, the correlation coefficient is used, which is represented by r. The value range of the correlation coefficient is [-1, +1]. Less than zero means negative correlation, and greater than zero Indicates a positive correlation.

The formula for calculating the correlation coefficient is:

The standard of the correlation coefficient to judge the degree of linear correlation between two variables is: when |r|=0, it means that x and y are completely irrelevant; when 0<|r|<0.3, it means that x and y are not related; when 0.3 When <|r|<0.5, x and y are considered to be lowly correlated; when 0.5<|r|<0.8, x and y are considered to be significantly correlated; when 0.8<|r|1, x and y are considered to be highly correlated.

Due to the different units of input variables, the magnitude difference is also large, and the output of neurons is usually limited within a certain range. The model designed in this paper uses a single-ended S-type activation function, and the output is limited between 0 and 1, so the original data needs to be normalized to avoid neuron oversaturation.

The normalization formula is as follows:

In the formula: x is the system data outside the interval [0, 1]; X _max and X _min are the maximum and minimum values in the system data respectively; y is the normalized value of the system data x.

The improvement of BP network training accuracy can be obtained by using a hidden layer to increase the number of neurons, but it is not that the more hidden nodes, the better. If the number of hidden nodes is large, the accuracy of network output results will be higher. decline.

The network of this method uses the same input and output sample pair, the maximum number of network training is set to 1000 times, and the error is set to 0.0001. The network is trained by continuously changing the number of nodes in the hidden layer. After many comparison experiments, the number of hidden nodes is selected is 13, and the network at this time can achieve better accuracy. The transfer function of the hidden layer is selected as the tansig function, the transfer function of the output layer is selected as the tansig function, and the output result is between [0, 1].

For the output of the neural network, denormalization processing is required to convert the value of the network between [0, 1] into the actual output value of the system,

The denormalization formula corresponding to the normalization formula is as follows: x=y(x _max -x _min )+x _min

In the formula, y is a value between [0, 1]; x _max and x _min are the maximum and minimum values in the system data respectively; x is the actual data of the system after denormalization of the network output.

On the simulation model of the pneumatic regulating valve, the optimal PI parameters are manually found by using the closed-loop step test. This PI parameter can make the rising time of the regulating valve short and the overshoot small.

Use the basic parameters (stroke, diaphragm area, spring stiffness, valve trim quality) on the mechanism model of the pneumatic control valve to simulate various types of control valves, do open-loop experiments for each control valve, and obtain T1 (full off fall time), T2 (full open rise time), S1 (area enclosed by time T1 when the curve falls;), S2 (area enclosed by time T2 when the curve rises), K, and then adjust this type Valve does closed-loop experiments and manually finds the optimal PI parameters to obtain multiple sets of BP neural network training samples.

The present invention proposes a PI parameter tuning method based on BP neural network. Without establishing a mathematical model of the regulating valve, the BP network is used for self-adaptive learning, parallel distributed processing, and strong robustness and fault tolerance. The neural network can understand the structure, parameters, uncertainty and nonlinearity of the system through its own learning process, and give the control law required by the system. The controller composed of neural network has good adjustment ability and robustness, which greatly improves the control accuracy of the valve positioner.

Description of drawings

Fig. 1 is the closed-loop positioning control system diagram of the BP neural network regulating valve used in the present invention;

Fig. 2 is the used BP neural network structural diagram of the present invention;

Fig. 3 is the step response diagram of the control valve simulation model;

Figure 4 is a partial enlarged view of the rise of the regulating valve;

Figure 5 is a response diagram of the closed-loop control of the regulating valve.

Detailed ways

The present invention will be described below with reference to the accompanying drawings and examples.

With reference to shown in Fig. 1, the valve positioner PI parameter tuning method based on BP neural network of the present invention and the control valve closed-loop control system of its composition, obtain the input variable T1 of BP neural network by the control valve closed-loop system, T2, K, and PI control parameters Kp, Ti obtained by BP neural network training.

The BP neural network is a multi-layer feed-forward neural network. The topology of the 3-layer BP network is shown in Figure 2, including an input layer, an output layer, and a hidden layer. Each neuron is connected to all neurons in the next layer. , there is no connection between neurons in the same layer. The basic principle of BP neural network is to use the gradient descent method to adjust the weight and threshold, so that the mean square error between the actual output value and the expected output value of the network is the smallest.

The standard BP algorithm does not consider the gradient direction of the previous moment when modifying the weight, so that the learning process often oscillates and converges slowly. Therefore, the present invention adopts the Levenberg-Marquardt algorithm to speed up the network training speed and improve the accuracy of the network training by optimizing the search direction of the BP neural network.

Levenberg-Marquardt algorithm, including the following:

1) Let w(i) represent the vector composed of the weight and threshold of the i-th iteration, and w(i+1) is the vector composed of the new weight and threshold, as shown in the formula: w(i+1 )=w(i)+Δw;

2) For Newton's method, the expression of Δw is: where E(w)

Let the evaluation function be the mean square error: Where e(w) is the error (i=1,2,3,...,N),

3) In the formula, J is the Jacobian matrix, and its form is:

4) For the Gauss-Newton law, there is ΔE=-[J ^T (w)*J(W)] ^-1 *J(w)e(w) LM algorithm is its improvement, then ΔE=-[J ^T (w) )*J(W)+μI] ^-1 *J(w)e(w) In the formula, the proportional coefficient μ>0 is a constant, and I is an identity matrix,

5) It can be seen from the above formula that if the proportional coefficient μ=0, the L-M algorithm becomes the Gauss-Newton method; if the value of μ is close to 1, the L-M algorithm is transformed into the gradient descent method, and as the number of successful iterations increases, The value of μ gradually decreases, and when the error is close to the minimum, the L-M algorithm gradually evolves into the Gauss-Newton method. Therefore, it has been shown through practice that the L-M algorithm can increase the speed by dozens or even hundreds of times compared with the original gradient descent method.

The present invention indirectly obtains the parameters that can reflect the control performance through the step test. The step experiment of the control valve simulation model is shown in Figure 3.

Make a partial enlarged diagram for the rise time, and mark the performance parameters collected during the ascent on the diagram. The partial enlarged diagram is shown in Figure 4: the input variables of the BP neural network are obtained from it.

Among them, T2 is the time required for the valve position to go from 0% to 100% when the regulating valve is fully open and fully closed; S2 is the area surrounded by the time T2 when the curve rises. During the inflation and deflation process of the regulating valve, due to the switching between sonic inflation and subsonic inflation, between constant volume inflation and isobaric inflation, the rising curve will have an inflection point, so S2 is selected to describe this characteristic. t2 is the pure delay time of rising, that is, the time interval from the moment when the input quantity changes to the moment when the output quantity starts to change.

In addition, what is not clearly marked in the figure is the data when the corresponding curve drops. T1 is the time required for the valve position to change from 100% to 0% when the regulating valve is fully open and fully closed; S1 is the time and time when the curve drops The area enclosed by T1. t1 is the pure delay time for falling. K is the magnification of the regulating valve (the ratio of the output valve position difference to the input current difference).

The determination of the input data of the neural network is actually the extraction of the feature quantity (influencing factor), mainly considering whether it has a clear causal relationship with the system output index. If the network input data has no connection with the output index, the relationship between them cannot be established. Therefore, it is necessary to accurately analyze the factors that are strongly correlated with the output PI parameters and use them as input quantities.

Correlation analysis is performed between the collected input performance parameter sets and the manually optimized PI parameter sets. Judging the closeness of the correlation between variables needs to be judged by an indicator. Here, the correlation coefficient is used, which is represented by r. The value range of the correlation coefficient is [-1, +1]. Less than zero means negative correlation, and greater than zero Indicates a positive correlation. The formula for calculating the correlation coefficient is:

The standard for judging the degree of linear correlation between two variables through the correlation coefficient is: when |r|=0, it means that x and y are completely irrelevant; when 0<|r|<0.3, it means that x and y are not related; when When 0.3<|r|<0.5, x and y are considered to be lowly correlated; when 0.5<|r|<0.8, x and y are considered to be significantly correlated; when 0.8<|r|1, x and y are considered to be highly correlated.

Now take the analysis of the relationship between the main parameters of the input performance and the output parameters Kp and Ti as an example, and the correlation coefficients between them are shown in Table 1 below.

Table 1 is a table of correlation coefficients between input parameters and output parameters.

It can be seen from the table that T1, T2, S1, S2, and k have the highest correlation degree according to the correlation degree standard, and finally determine that the input variables of the BP neural network are 3, and the output variables are Kp and Ti parameters.

Due to the different units of input variables, the magnitude difference is also large, and the output of neurons is usually limited within a certain range. The model designed in this paper uses a single-ended S-type activation function, and the output is limited between 0 and 1, so the original data needs to be normalized to avoid neuron oversaturation. The normalization formula is shown in formula (11):

In the formula: x is the system data outside the interval [0, 1]; X _max and X _min are the maximum and minimum values in the system data respectively; y is the normalized value of the system data x.

The improvement of BP network training accuracy can be obtained by using a hidden layer to increase the number of neurons, but it is not that the more hidden nodes, the better. If the number of hidden nodes is large, the accuracy of network output results will be higher. decline. The network of this method uses the same input and output sample pair, the maximum number of network training is set to 1000 times, and the error is set to 0.0001. The network is trained by continuously changing the number of nodes in the hidden layer. After many comparison experiments, the number of hidden nodes is selected is 13, and the network at this time can achieve better accuracy. The transfer function of the hidden layer is selected as the tansig function, the transfer function of the output layer is selected as the tansig function, and the output result is between [0, 1].

For the output of the neural network, denormalization processing is required to convert the value of the network between [0, 1] into the actual output value of the system. The denormalization formula is as follows: x=y(x _max -x _min )+ _xmin ;

In the formula, y is a value between [0, 1]; X _max and X _min are the maximum and minimum values in the system data respectively; x is the actual data of the system after denormalization of the network output.

Use the trained BP neural network to verify the simulation model of the pneumatic control valve, obtain the corresponding PI parameters, and use this parameter to do closed-loop multi-step test. The closed-loop response results are shown in Figure 5.

It can be seen from Figure 5 that the test results are good, and the response curve of the simulation model can achieve "fast, accurate and stable" control effects at each step. It can be shown that, on the control valve simulation model, the PI parameter setting design method of the intelligent positioner based on the BP neural network is accurate.

Claims (6)

1. A kind of method that utilizes BP neural network to tune intelligent locator PI parameter, it is characterized in that, obtain the input and output variable of BP neural network by carrying out autocorrelation analysis to control valve system, determine the hidden node number of model simultaneously and Transfer function, and then determine the system parameters, after training to obtain the required PI parameters. 2. The method according to claim 1, wherein the BP neural network is used to adjust the PI parameters of the intelligent positioner, wherein the regulating valve system is a first-order inertia plus pure hysteresis model system. 3. utilize BP neural network according to claim 1 to come the method for tuning intelligent locator PI parameter, it is characterized in that, described autocorrelation analysis method first carries out preprocessing to data by normalization method, utilizes autocorrelation analysis simultaneously method to determine the degree of linear correlation between variables. 4. utilize BP neural network according to claim 1 to come the method for setting intelligent locator PI parameter, it is characterized in that, described neural network input variable is full closing time T1, full opening time T2, magnification factor k, output variable For the ratio Kp, integral Ti. 5. utilize BP neural network according to claim 1 to come the method for setting intelligent locator PI parameter, it is characterized in that, described BP neural network is a three-layer neural network, is respectively input layer, hidden layer and output layer, The input layer contains three neurons, the hidden layer contains nine neurons, and the output layer contains two neurons. 6. utilize BP neural network according to claim 5 to come the method for tuning intelligent locator PI parameter, it is characterized in that, described transfer function is selected as tansig function, output result is between [0,1].