CN101539781A - Electrogalvanizing zinc coating thickness BP neural network control method and application in PLC thereof - Google Patents

Abstract

The invention relates to an electrogalvanizing zinc coating thickness BP neural network control method and an application in PLC thereof. The method includes the following steps of: (1) collecting electrogalvanizing sample data; (2) establishing BP neural network; (3) BP neural network learning and training; and (4) recording the trained BP neural network in a PLC controller. The adoption of an electrogalvanizing zinc coating thickness BP neural network controller designed by the invention can control galvanizing thickness with high accuracy, effectively inhibit industrial interference, and has intelligent self-adaptability, and good accuracy and fault-tolerance.

Description

Electrogalvanizing zinc coating thickness BP neural network control method and the application on PLC thereof

Technical field

The present invention relates to a kind of BP neural network control method of electrogalvanizing zinc coating thickness, also relate to the application of BP neural network control method on PLC of electrogalvanizing zinc coating thickness.

Background technology

On the electrogalvanizing production line, zinc coating thickness control is the core control technology.Aspect the zinc coating thickness design of Controller, mentality of designing in the past will be made precise math model to controlling object often; Carry out design of Controller according to precise math model afterwards.

Electro-galvanized layer THICKNESS CONTROL object has between multivariate, strong coupling and the variable characteristics such as amplitude of variation difference is big, is difficult to set up point-device controlling object model, and each variable is carried out de control; Simultaneously, accurately fixing mathematical model has shortcomings such as anti-interference difference and fault-tolerant ability difference.

Summary of the invention

In order to improve the product quality of electrogalvanizing production line, make product thickness even, accurately reach the setting thickness requirement, improve the top grade rate of product, the PLC control method of the electrogalvanizing zinc coating thickness based on the BP neural network algorithm of the present invention comprises the following steps:

A, training sample data are gathered

Gather the data sample of the actual production of some at the electrogalvanizing production line, sample is approximately wanted 10,000 groups of samples, and guarantees the ergodicity of data.Sample is gathered typing with the form of input vector and output vector.1 input vector and 1 corresponding output vector are formed one group of sample data.Input vector comprises thickness of coating, following thickness of coating, elements such as width of steel band and coating bath total current number; Output vector comprises that thickness of coating is calculated speed and following thickness of coating is calculated two elements of speed.

In conjunction with the experience and the disturbing factor of actual production, consider that simultaneously the defective of some data is taked to revise.According to senior zincincation slip-stick artist's empirical data and the ideal effect that expectation reaches, data are analyzed and handled, obtain outstanding training data.

B, BP neural network are set up

According to electroplating thickness control requirement, network approximation accuracy and computer computation ability, and input and output vector situation is set up amphineura network structure.They contain 1 input layer and two hidden layers and 1 BP neural network structure that output layer is formed to set up two neural networks with same structure.

The input layer of first network structure is the column vector that contains 3 elements, is respectively thickness of coating, coil of strip width and coating bath total current number; It is linear function that hidden layer 1 has 20 its excitation functions of neuron; It is the S function that hidden layer 2 has 20 its excitation functions of neuron, as formula (1):

$S\left(n\right)=\log \mathrm{sig}\left(n\right)=\frac{1}{1+e^{-n}}---\left(1\right)$

Output layer adopts the linear incentive function, is output as thickness of coating and calculates speed.

Second network structure is identical with first.3 elements of input column vector are respectively: following thickness of coating, coil of strip width and coating bath total current number; Be output as down thickness of coating and calculate speed.Two networks are with parallel way computing data.

C, e-learning and training

Network training and study employing change gradient algorithm (Conjugate Gradient Backpropagation, CGBP).Its for the first time iteration begin to search for along the steepest gradient descent direction, suc as formula (2):

p(0)＝-g(0) (2)

Then, the linear search of decision optimum distance carries out along the direction of current search:

x(k+1)＝x(k)+αp(k) (3)

P (k)=-g (k)+β (k) p (k-1) (4) becomes the gradient modified value and adopts the Fletcher-Reeves correction algorithm, suc as formula (5):

$\mathrm{\β}\left(k\right)=\frac{g^T\left(k\right)g\left(k\right)}{g^T(k-1)g(k-1)}---\left(5\right)$

Adopt Fletcher-Reeves correction algorithm speed fast, and the BP algorithm more variable than common learning rate (Variable Learning RateBackpropagation, VLBP) training algorithm is fast, saves the Computer Storage space simultaneously.

The error performance function is square error MES (Mean Square Error).

Programming of the BP network architecture and learning algorithm programming realize on computers, adopt mathematical simulation emulational language MATLAB or other higher level lanquages.To import the MATLAB program of computing machines through 10000 groups of sample datas of Screening Treatment, write the corresponding program training network, obtain corresponding weights and deviate.

The Industry Control of D, BP network realizes

With network structure with the corresponding descriptive language coding of PLC.Because the storage space of PLC is smaller, thus the method for off-line training adopted, what promptly only typing had trained in computing machine in program in PLC, reach the network of accuracy requirement.

Adopt the electrogalvanizing zinc coating thickness BP neural network controller of the present invention's design, can control zinc-plated thickness accurately, suppress industrial noise, intelligent adaptive effectively, have good accuracy and fault-tolerance.

Description of drawings

Fig. 1 is the process flow diagram of electrogalvanizing zinc coating thickness BP neural network control method;

Fig. 2 is error performance function and the graph of relation of training step number;

Fig. 3 is the comparison diagram of nerve network controller simulation result and actual result.

Embodiment

Below in conjunction with drawings and the specific embodiments the present invention is further described.

As shown in Figure 1, the PLC control method of the electrogalvanizing zinc coating thickness based on the BP neural network algorithm of the present invention comprises the following steps:

A, training sample data are gathered

Gather the data sample of the actual production of some at the electrogalvanizing production line, sample is approximately wanted 10,000 groups of samples, and guarantees the ergodicity of data.Sample is gathered typing with the form of input vector and output vector.1 input vector and 1 corresponding output vector are formed one group of sample data.Input vector comprises thickness of coating, following thickness of coating, elements such as width of steel band and coating bath total current number; Output vector comprises that thickness of coating is calculated speed and following thickness of coating is calculated two elements of speed.

In conjunction with the experience and the disturbing factor of actual production, consider that simultaneously the defective of some data is taked to revise.According to senior zincincation slip-stick artist's empirical data and the ideal effect that expectation reaches, data are analyzed and handled, obtain outstanding training data.

B, BP neural network are set up

According to electroplating thickness control requirement, network approximation accuracy and computer computation ability, and input and output vector situation is set up amphineura network structure.They contain 1 input layer and two hidden layers and 1 BP neural network structure that output layer is formed to set up two neural networks with same structure.

The input layer of first network structure is the column vector that contains 3 elements, is respectively thickness of coating, coil of strip width and coating bath total current number; It is linear function that hidden layer 1 has 20 its excitation functions of neuron; It is the S function that hidden layer 2 has 20 its excitation functions of neuron, as formula (1):

$S\left(n\right)=\log \mathrm{sig}\left(n\right)=\frac{1}{1+e^{-n}}---\left(1\right)$

Output layer adopts the linear incentive function, is output as thickness of coating and calculates speed.

Second network structure is identical with first.3 elements of input column vector are respectively: following thickness of coating, coil of strip width and coating bath total current number; Be output as down thickness of coating and calculate speed.Two networks are with parallel way computing data.

C, e-learning and training

Network training and study employing change gradient algorithm (Conjugate Gradient Backpropagation, CGBP).Its for the first time iteration begin to search for along the steepest gradient descent direction, suc as formula (2):

p(0)＝-g(0) (2)

Then, the linear search of decision optimum distance carries out along the direction of current search:

x(k+1)＝x(k)+αp(k) (3)

p(k)＝-g(k)+β(k)p(k-1) (4)

Become the gradient modified value and adopt the Fletcher-Reeves correction algorithm, suc as formula (5):

$\mathrm{\β}\left(k\right)=\frac{g^T\left(k\right)g\left(k\right)}{g^T(k-1)g(k-1)}---\left(5\right)$

Adopt Fletcher-Reeves correction algorithm speed fast, and save the Computer Storage space than common training algorithm.

The error performance function is square error MES (Mean Square Error).

Programming of the BP network architecture and learning algorithm programming realize on computers, adopt mathematical simulation emulational language MATLAB or other higher level lanquages.To import the MATLAB program of computing machines through 10000 groups of sample datas of Screening Treatment, write the corresponding program training network, obtain corresponding weights and deviate.The error performance function is distinguished as shown in Figures 2 and 3 with the graph of relation of training step number and the comparison diagram of nerve network controller simulation result and actual result, and controller is through emulation testing, and output valve and actual result error are within 1%.

The Industry Control of D, BP network realizes

With network structure with the corresponding descriptive language coding of PLC.Because the storage space of PLC is smaller, thus the method for off-line training adopted, what promptly only typing had trained in computing machine in program in PLC, reach the network of accuracy requirement.

Claims (5)

1, electrogalvanizing zinc coating thickness BP neural network control method is characterized in that: comprise the following steps:

(1) gathers the electrogalvanizing sample data;

(2) set up the BP neural network;

(3) BP neural network learning and training.

2, electrogalvanizing zinc coating thickness BP neural network control method as claimed in claim 1, it is characterized in that: described electrogalvanizing sample is gathered typing with the form of input vector and output vector, 1 input vector and 1 corresponding output vector are formed one group of sample data, input vector comprises thickness of coating, following thickness of coating, width of steel band and coating bath total current number, output vector comprise that thickness of coating is calculated speed and following thickness of coating is calculated speed.

3, electrogalvanizing zinc coating thickness BP neural network control method as claimed in claim 1, it is characterized in that: set up two neural networks with same structure, each neural network is and contains 1 input layer and two hidden layers and 1 BP neural network structure that output layer is formed; The input layer of first network structure is the column vector that contains thickness of coating, coil of strip width and several 3 elements of coating bath total current, hidden layer 1 has 20 neurons, its excitation function is a linear function, hidden layer 2 has 20 neurons, its excitation function is the S function, output layer adopts the linear incentive function, is output as thickness of coating and calculates speed; Second network structure is identical with first, and 3 elements of input column vector are respectively: following thickness of coating, coil of strip width and coating bath total current number; Be output as thickness of coating and calculate speed, two networks are with parallel way computing data.

4, electrogalvanizing zinc coating thickness BP neural network control method as claimed in claim 1, it is characterized in that: described BP neural network learning adopts with training and becomes gradient algorithm, its for the first time iteration begin to search for along the steepest gradient descent direction, then, the linear search of decision optimum distance carries out along the direction of current search, becomes the gradient modified value and adopts the Fletcher-Reeves correction algorithm.

5, electrogalvanizing zinc coating thickness BP neural network control method as claimed in claim 1 is characterized in that: with the BP neural network typing PLC controller that trains.

Images