CN106862284B - A kind of cold rolled sheet signal mode knowledge method for distinguishing - Google Patents

Abstract

A method for signal pattern recognition of cold-rolled strips, the method comprising the following steps: collecting the shape measurement values of each measurement section in the width direction of the cold-rolled strip steel measured online by a shapemeter, and obtaining the shape value of each measurement section; The original data output by the shape meter is input to an n-layer neural network as the feature extraction layer, mainly through training to let the network automatically extract features to eliminate artificial traces; use the improved quantum neural network based on genetic algorithm for plate shape recognition. The invention applies the improved quantum neural network of the multi-layer excitation function optimized by the genetic algorithm to the shape pattern recognition technology, which significantly improves the training efficiency of the network and effectively solves the problems of accuracy and real-time performance encountered in the traditional shape recognition method It is not ideal enough, the network structure is complex, the training time is long, and the stability and robustness are poor.

Description

A method for signal pattern recognition of cold-rolled strip

technical field

The invention belongs to the field of cold-rolled strips and relates to a method for identifying signal patterns of cold-rolled strips.

Background technique

In the production process of cold-rolled strips, the pattern recognition of cold-rolled strip shape is an important part of the cold-rolled strip shape control system. In this link, the recognition accuracy of the flatness defect state is crucial to the subsequent flatness control effect. role. The stress signal data in the plate on the production site is collected by the shape meter for identification, and the type of the current shape defect state is judged, and fed back to the shape control mechanism, and the defect degree of the shape is reduced by controlling the adjustment of the actuator, and finally To make the output of the shape meet the production standard of the process, it is necessary to seek and study the identification method of the high-precision shape defect mode.

The traditional plate shape signal pattern recognition method is based on the polynomial decomposition method of the least square method and the improved orthogonal polynomial regression decomposition method. These methods have poor anti-interference ability and have defects in theory. Meet the high-precision shape control requirements. The plate shape recognition method based on the fuzzy principle is simple and effective, but the accuracy and real-time performance are not ideal, so it is seldom used in practice. In recent years, experts and scholars in this field have studied the technology of plate shape signal recognition and application based on neural network, and achieved very good results. However, due to the complexity of the neural network and the insufficiency of the existing optimization technology, the recognition model structure based on the neural network pattern recognition system is complex, and there are problems such as long network training time and poor stability.

Plate shape refers to the distribution of internal residual stress along the plate width direction. The task of plate shape recognition is to map a set of discrete values of tension distribution detected online into a few characteristic parameters after certain processing, and can be compared A good reflection of the classification of shape defects. The quantum neural network of the multi-layer excitation function adopts the superposition of multiple Sigmoid functions to divide the feature space into more energy levels to represent more states. The superimposed hidden layer excitation function has multiple different quantum intervals, and Each quantum interval corresponds to a different quantum energy level or step, that is, it can correspond to a different spatial structure. During the training process of the model, when the sample data enters the feature space, it is not necessary to search for the corresponding spatial structure in the entire feature space, that is, [0,1], but to search in the layered spatial structure. , so that the sample information can be quickly mapped to the corresponding magnitude or ladder; at the same time, if the actual output of the model and the expected output have a large error, the quantum interval can be adjusted through an appropriate learning algorithm to change the divided quantum energy level or Step width, to adapt to the structural characteristics of existing data samples, so as to map to the corresponding spatial structure better and faster, so that the uncertainty in the sample data is acquired and quantified by the quantum neural network, thus greatly shortening the training of the model Time, improve the accuracy and convergence speed of model prediction results.

On the other hand, intelligent optimization algorithms have developed rapidly in recent decades, and the optimal solutions of many nonlinear optimization problems can be obtained through intelligent optimization methods. Applying intelligent optimization algorithm technology to quantum neural network training improves modeling accuracy and efficiency, which is a promising and valuable research direction in the field of nonlinear modeling. This also provides technical support and theoretical basis for high-precision and high-efficiency cold-rolled strip signal online pattern recognition technology.

In summary, it is a further improvement to develop a high-precision and high-efficiency cold-rolled flat shape signal online pattern recognition method to provide a reliable control basis for the control system, thereby producing high-quality cold-rolled strip products. A key technical problem that needs to be solved urgently at the current cold-rolled strip shape control level.

Contents of the invention

The purpose of the present invention is to provide a method for signal pattern recognition of cold-rolled strips, which can effectively solve the problems of insufficient accuracy and real-time performance encountered in traditional flatness pattern recognition methods, complex identification model structure and too long network training time , Poor stability and other technical problems, can provide a reliable control basis for the control system, and provide a strong guarantee for improving the quality of cold-rolled strip shape control.

In order to solve the above problems, the technical solution provided by the invention is:

A method for signal pattern recognition of cold-rolled strips, the method comprising the following steps:

Step 1 Collect the flatness measurement values of each measurement section in the width direction of the cold-rolled strip steel measured online by the flatness meter, and obtain the flatness value of each measurement section; let the number of measurement sections be m, and the flatness measurement value of the _i -th measurement section is F _i ;

Step 2 Input the original data output by the shape meter into an n-layer neural network as the feature extraction layer, mainly through training to let the network automatically extract features to eliminate artificial traces; because the plate shape mainly has left waves, right waves, and middle waves , two-sided wave, right three-point wave, left three-point wave, four-point wave and side-center wave eight defect modes, so the network output is a _i , i=1,2,3...m, where m is the output The number, here m=8, the calculation formula of each layer of the network is as follows:

a ^l _i ＝f(W _l x _l +b _l ),l＝1,2,...n

Where l is the lth layer of the network; W _l is the network weight of the lth layer; x _l is the input of the lth layer; b _l is the network neuron bias;

Step 3 uses the improved quantum neural network based on genetic algorithm to recognize the plate shape, takes the output a _i of the data processing module as the input of the quantum neural network, and the output u ₁ , u ₂ , u ₃ , u ₄ of the quantum neural network are the plates The degree of membership of the shape;

If u ₁ >0, it means that the board shape has a left wave; if u ₁ <0, it means that the board shape has a right wave;

If u ₂ >0, it means that the board shape has middle waves; if u ₂ <0, it means that the board shape has double-sided waves;

If u ₃ >0, it means that the board shape has a left three-point wave; if u ₃ <0, it means that the board shape has a right three-point wave;

If u ₄ >0, it means that the board shape has a quarter wave, and if u ₄ <0, it means that the board shape has a side wave.

In step 3, the improved quantum neural network based on genetic algorithm, its establishment process includes the following steps:

1) Determine the control parameters for the quantum neural network training learning of plate shape pattern recognition:

The specific control parameters include: the initial population number NP used for network training genetic algorithm, the maximum learning algebra N of genetic algorithm, the target value e of network training effect, and the number of training samples M;

2) Processing of training sample data:

Input the data output by the shape meter directly into the network, let the multi-layer neural network at the front end of the network perform automatic feature extraction, and then input it into the quantum neural network of the multi-layer activation function;

3) Setting of multi-layer activation function improved quantum neural network:

The multi-layer excitation function improved quantum neural network model adopts a mesh topology, which is composed of a feature extraction layer, a hidden layer and an output layer. The feature extraction layer is also called the input layer; Level transformation function, each multi-level transformation function is a linear superposition of a series of n _s Sigmoid functions with quantum interval offset, that is, the multi-layer activation function; therefore, the output y _h of neurons in the hidden layer can be expressed as:

Where V is the hidden layer weight of the multi-layer excitation function quantum neural network, a is the output of the feature extraction layer, θ ^r is the quantum interval, and n _s is the number of quantum intervals;

4) Define the weights and biases of the feature extraction layer network, the hidden layer weights, quantum intervals, and output layer weights of the multi-layer excitation function quantum neural network as individual vectors of the genetic algorithm, and determine NP random distributions according to equal probability The initial value of the population of the individual vector;

5) According to the mutation operation, crossover operation and selection operation of the genetic algorithm, the optimization learning of the individual vector is carried out; the network parameters obtained by the optimization learning are used to configure the parameters of the improved quantum neural network of the multi-layer activation function, and the M training sample data is brought to The improved quantum neural network with multi-layer activation function is used to calculate the output value, and then the sum of the squares of the arithmetic difference between the 4M network output values corresponding to the above M training sample data and their actual values is calculated; the sum of the squares of the arithmetic difference Defined as the error index function value; if the error index function value is less than the target value g of the network training effect, it means that the network training is successful, otherwise, the network should continue to be trained, and g should be selected according to the number of training samples and the scale of the network;

6) Record the network parameters of training and learning, and obtain an improved quantum neural network with a multi-layer activation function of 15 inputs and 4 outputs;

After the network is established according to the above steps, the plate shape recognition is carried out.

The beneficial effects of the present invention are:

1) Through a multi-layer neural network, the data input by the shape meter is automatically extracted, processed automatically, and artificial interference factors are eliminated, which can significantly improve the accuracy of network recognition of shape.

2) The improved quantum neural network with multi-layer excitation function optimized by genetic algorithm is applied to the shape pattern recognition technology, which significantly improves the training efficiency of the network and effectively solves the problems of accuracy and real-time performance encountered in traditional shape recognition methods Problems such as unsatisfactory, complex network structure, long training time, poor stability and robustness provide a reliable basis for shape control.

Description of drawings

Fig. 1 is the implementation flowchart of the present invention;

Fig. 2 is the improved quantum neural network structure topological diagram of type multi-layer activation function;

Fig. 3 is the improved quantum neural network training and learning flowchart of the multi-layer activation function based on genetic algorithm in one embodiment of the present invention;

Fig. 4 is a recognition effect diagram obtained after adopting the present invention for plate shape recognition.

Detailed ways

The present invention will be further described below in conjunction with specific examples and accompanying drawings.

A kind of method for cold-rolled strip signal pattern recognition of the present invention, Fig. 1 is the flowchart of an embodiment of the present invention, and it comprises the following steps:

Step 1 collects the flatness measurement values of each measurement section in the width direction of the cold-rolled strip measured by the shapemeter on-line, and obtains the flatness value of each measurement section; let the number of measurement sections be m, m=15, and the _i -th measurement section strip The shape measurement value is F _i ;

Step 2 Input the original data output by the shape meter into an n-layer neural network as the feature extraction layer, mainly through training to let the network automatically extract features to eliminate artificial traces; since the plate shape mainly has left waves, right waves, and middle waves , two-sided wave, right three-point wave, left three-point wave, four-point wave and side-center wave eight defect modes, so the network output is a _i , i=1,2,3...m, where m is the output The number, here m=8; the calculation formula of each layer of the network is as follows:

a ^l _i ＝f(W _l x _l +b _l ),l＝1,2,...n

Where l is the lth layer of the network; W _l is the network weight of the lth layer; x _l is the input of the lth layer; b _l is the network neuron bias;

Step 3 Use the improved quantum neural network based on genetic algorithm to recognize the shape of the plate, and use the output a _i of the data processing module as the input of the quantum neural network; the output u ₁ , u ₂ , u ₃ , u ₄ of the quantum neural network are the plate The degree of membership of the shape;

If u ₁ >0, it means that the board shape has a left wave; if u ₁ <0, it means that the board shape has a right wave;

If u ₂ >0, it means that the board shape has middle waves; if u ₂ <0, it means that the board shape has double-sided waves;

If u ₃ >0, it means that the board shape has a left three-point wave; if u ₃ <0, it means that the board shape has a right three-point wave;

If u ₄ >0, it means that the board shape has a quarter wave, and if u ₄ <0, it means that the board shape has a side wave.

According to the above scheme, the establishment process of the improved quantum neural network based on genetic algorithm is:

1) Determine the control parameters for the quantum neural network training learning of plate shape pattern recognition:

The specific control parameters include: the initial population number NP used for network training genetic algorithm, the maximum learning algebra N of genetic algorithm, the target value e of network training effect, and the number of training samples M;

2) Processing of training sample data:

The data output by the shape meter is directly input into the network, and the multi-layer neural network at the front end of the network is used for automatic feature extraction, and then input into the quantum neural network of the multi-layer activation function.

3) Setting of multi-layer activation function improved quantum neural network:

The multi-layer excitation function improved quantum neural network model adopts the network topology shown in Figure 2, which is composed of a feature extraction layer, a hidden layer and an output layer. The feature extraction layer is also called the input layer; the neurons in the hidden layer of the model The excitation function adopts multi-quantum energy level transformation function, and each multi-energy level transformation function is a series of linear superposition of n _s Sigmoid functions with quantum interval offset, that is, multi-layer excitation function; therefore, the output of hidden layer neurons y _h can be expressed as:

Where V is the hidden layer weight of the multi-layer activation function improved quantum neural network, a is the output of the feature extraction layer, θ ^r is the quantum interval, n _s is the number of quantum intervals;

4) Define the weight and bias of the feature extraction layer network, the hidden layer weight of the multi-layer excitation function quantum neural network, the quantum interval and the output layer weight as the individual vector of the genetic algorithm, and determine NP according to the equal probability random distribution The initial value of the population of the individual vector;

5) According to the mutation operation, crossover operation and selection operation of the genetic algorithm, the optimization learning of the individual vector is carried out; the network parameters obtained by the optimization learning are used to configure the parameters of the improved quantum neural network of the multi-layer activation function, and the M training sample data is brought to The improved quantum neural network with multi-layer activation function is used to calculate the output value, and then the sum of the squares of the arithmetic difference between the 4M network output values corresponding to the above M training sample data and their actual values is calculated; the sum of the squares of the arithmetic difference Defined as the error index function value; if the error index function value is less than the target value g of the network training effect, it means that the network training is successful, otherwise, the network should continue to be trained, and g should be selected according to the number of training samples and the scale of the network;

6) Record the network parameters for training and learning, and obtain an improved quantum neural network with a multi-layer activation function of 15 inputs and 4 outputs; after the network is established according to the above steps, the plate shape recognition is performed.

The method of the invention can be used in four-roller, six-roller single-stand or multi-stand cold tandem rolling units. Taking a single-stand six-high rolling mill as an example, the products rolled by the six-high rolling mill include ordinary plates, high-strength steel, some stainless steel and silicon steel, etc. The model used in this example is 900HC rolling mill, the main technical performance indicators and equipment parameters of this unit are:

Rolling speed: Max 300m/min, rolling pressure: Max 8000KN, maximum rolling torque: 60KN×m, crimping tension: 4～80KN; incoming material thickness range: 2.0～4.0mm, plate width range: 460～780mm, Finished product thickness: 0.24~1.5mm;

Working roll size: φ270/φ245×900mm, intermediate roll size: φ340/φ320×920mm, backup roll size: φ850/φ790～×850;

The maximum lateral movement of the intermediate roll: 200mm, the positive/negative bending force of the working roll (one side): 400/254.5KN.

The flatness measuring device adopts a self-made flatness meter with 15 channels.

In this example, the specific calculation process of the shape signal of the failure channel is as follows:

(1) Collect and receive the flatness measurement values of each measurement section in the direction of the cold-rolled strip measured by the on-line measurement of the receiving flatness meter. The shape meter is equipped with several measurement areas with definite widths in the width direction of the cold-rolled strip, and each measurement area provides the shape value of the corresponding area. There are 15 effective measurement areas from the operation side of the shape meter to the eating side of the transmission, and the number of measurement segments is m-15, that is, F _i , i=1, 2,...15 is the international unit of shape I. The data measured by the shape meter are directly imported into the network without manual processing.

(2) Input the original data output by the shape meter into an n-layer neural network as the feature extraction layer, which mainly allows the network to automatically extract features and eliminate artificial traces through training. Since the board shape mainly has eight defect modes: left wave, right wave, middle wave, double-sided wave, right three-point wave, left three-point wave, quarter-point wave and side-center wave, the network output is a _i , i=1 ,2,3...m, where m is the number of outputs, here m=8. The calculation formula of each layer of the network is as follows:

a ^l _i ＝f(W _l x _l +b _l ),l＝1,2,...n

Where l is the first layer of the network; W _l is the network weight of the first layer; x _l is the input of the first layer; b _l is the bias of the network neuron.

(3) Use the improved quantum neural network based on genetic algorithm to recognize the plate shape, and use the output a _i of the data processing module as the input of the quantum neural network; the output u ₁ , u ₂ , u ₃ , u ₄ of the quantum neural network are The degree of membership of the plate shape;

If u ₁ >0, it means that the board shape has a left wave; if u ₁ <0, it means that the board shape has a right wave;

If u ₂ >0, it means that the board shape has middle waves; if u ₂ <0, it means that the board shape has double-sided waves;

If u ₃ >0, it means that the board shape has a left three-point wave; if u ₃ <0, it means that the board shape has a right three-point wave;

If u ₄ >0, it means that the board shape has a quarter wave, and if u ₄ <0, it means that the board shape has a side wave.

According to the above scheme, as shown in Figure 3, the establishment process of the new quantum neural network based on the genetic algorithm is as follows:

1. Determine the control parameters for the quantum neural network training study of plate shape pattern recognition:

The specific parameters include: the initial population number NP of the genetic algorithm used for network training, the maximum learning algebra N of the genetic algorithm, the target value of the network training effect e, and the number of training samples M;

2. Processing of training sample data:

The data output by the shape meter is directly input to the network, and the multi-layer neural network at the front end of the network is used for automatic feature extraction, and then input into the quantum neural network of the multi-layer activation function.

3. The settings of the improved quantum neural network with multi-layer activation function:

The improved quantum neural network model of multi-layer activation function adopts the network topology shown in Figure 2, which is composed of feature extraction layer (input layer), hidden layer and output layer. The neuron activation function of the hidden layer of the model adopts multi-quantum energy level transformation functions, and each multi-level energy level transformation function is a series of linear superposition of n _s Sigmoid functions with quantum interval offsets, that is, multi-layer activation functions; therefore, hidden The output _yh of neurons in a layer can be expressed as:

Among them, V is the hidden layer weight of the multi-layer excitation function improved quantum neural network, a is the output of the feature extraction layer, θ ^r is the quantum interval, and n _s is the number of quantum intervals.

4. Define the weight and bias of the feature extraction layer network, the hidden layer weight of the multi-layer excitation function improved quantum neural network, the quantum interval and the output layer weight as individual vectors of the genetic algorithm, which are determined according to an equal probability random distribution The initial value of the population of NP individual vectors;

5. According to the mutation operation, crossover operation and selection operation of the genetic algorithm, the optimization learning of the individual vector is carried out; the network parameters obtained by the optimization learning are used to configure the improved quantum neural network parameters of the multi-layer activation function, and M training sample data are brought into The improved quantum neural network of the multi-layer activation function calculates the output value, and then calculates the sum of the squares of the arithmetic differences between the 4M network output values corresponding to the above M training sample data and their actual values. The sum of squares of the arithmetic difference is defined as the error index function value; if the error index function value is less than the target value g of the network training effect, it means that the network training is successful, otherwise, the network should continue to be trained, and g should be based on the number of training samples and the network performance. size to choose;

6. Record the network parameters for training and learning, and obtain an improved quantum neural network with a multi-layer activation function of 15 inputs and 4 outputs;

After the network is established according to the above steps, the plate shape recognition is carried out.

Figure 4 shows the recognition effect diagram obtained after online pattern recognition in this example. Figure 4-a is the distribution diagram of the shape defect curve identified by the network; Figure 4-b is the left wave component after identification; Figure 4-c is the bilateral wave component after identification; Figure 4-d is the left third after identification wave; Figure 4-e is the wave in the edge after identification. It can be seen from the figure that the technical solution proposed by the present invention can well complete the identification task of plate shape defects. Solve the technical problems encountered by conventional methods such as unsatisfactory accuracy and real-time performance, complex identification model structure, long network training time, too many human factors in data feature processing, poor stability and robustness, and provide reliable control for the control system. The basis provides a strong guarantee for improving the quality of cold-rolled strips.

The above embodiments are only used to illustrate the calculation idea and characteristics of the present invention, and its purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above examples. Therefore, all equivalent changes or modifications based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims (2)

1. A method for cold-rolled strip signal pattern recognition, characterized in that: the method content may further comprise the steps: Step 1 Collect the flatness measurement values of each measurement section in the width direction of the cold-rolled strip measured online by the flatness meter, and obtain the flatness value of each measurement section; let the number of measurement sections be m, and the flatness measurement value of the i-th measurement section is F _i ; Step 2 Input the original data output by the shape meter into an n-layer neural network as the feature extraction layer, mainly through training to let the network automatically extract features to eliminate artificial traces; because the plate shape mainly has left waves, right waves, and middle waves , two-sided wave, right three-point wave, left three-point wave, four-point wave and side-center wave eight defect modes, so the network output is a _i , i=1,2,3...m, where m is the output The number, here m=8, the calculation formula of each layer of the network is as follows: a ^l _i ＝f(W _l x _l +b _l ),l＝1,2,...n Where l is the lth layer of the network; W _l is the network weight of the lth layer; x _l is the input of the lth layer; b _l is the network neuron bias; Step 3 Use the improved quantum neural network based on genetic algorithm to recognize the plate shape, take the output a _i of the data processing module as the input of the quantum neural network, and the output u ₁ , u ₂ , u ₃ , and u ₄ of the quantum neural network are the plates The degree of membership of the shape; If u ₁ >0, it means that the board shape has a left wave; if u ₁ <0, it means that the board shape has a right wave; If u ₂ >0, it means that the board shape has middle waves; if u ₂ <0, it means that the board shape has double-sided waves; If u ₃ >0, it means that the board shape has a left three-point wave; if u ₃ <0, it means that the board shape has a right three-point wave; If u ₄ >0, it means that the board shape has a quarter wave, and if u ₄ <0, it means that the board shape has a side wave. 2. the method for a kind of cold-rolled strip signal pattern recognition according to claim 1, is characterized in that: described improved quantum neural network based on genetic algorithm, its establishment process comprises the steps: 1) Determine the control parameters for the quantum neural network training learning of plate shape pattern recognition: The specific control parameters include: the initial population number NP used for network training genetic algorithm, the maximum learning algebra N of genetic algorithm, the target value e of network training effect, and the number of training samples M; 2) Processing of training sample data: Input the data output by the shape meter directly into the network, let the multi-layer neural network at the front end of the network perform automatic feature extraction, and then input it into the quantum neural network of the multi-layer activation function; 3) Setting of multi-layer activation function improved quantum neural network: The multi-layer excitation function improved quantum neural network model adopts a mesh topology, which is composed of a feature extraction layer, a hidden layer and an output layer. The feature extraction layer is also called the input layer; Level transformation function, each multi-level transformation function is a linear superposition of a series of n _s Sigmoid functions with quantum interval offset, that is, the multi-layer activation function; therefore, the output y _h of neurons in the hidden layer can be expressed as: Where V is the hidden layer weight of the multi-layer excitation function quantum neural network, a is the output of the feature extraction layer, θ ^r is the quantum interval, and n _s is the number of quantum intervals; 4) Define the weights and biases of the feature extraction layer network, the hidden layer weights, quantum intervals, and output layer weights of the multi-layer excitation function quantum neural network as individual vectors of the genetic algorithm, and determine NP random distributions according to equal probability The initial value of the population of the individual vector; 5) According to the mutation operation, crossover operation and selection operation of the genetic algorithm, the optimization learning of the individual vector is carried out; the network parameters obtained by the optimization learning are used to configure the parameters of the improved quantum neural network of the multi-layer activation function, and the M training sample data is brought to The improved quantum neural network with multi-layer activation function is used to calculate the output value, and then the sum of the squares of the arithmetic difference between the 4M network output values corresponding to the above M training sample data and their actual values is calculated; the sum of the squares of the arithmetic difference Defined as the error index function value; if the error index function value is less than the target value g of the network training effect, it means that the network training is successful, otherwise, the network should continue to be trained, and g should be selected according to the number of training samples and the scale of the network; 6) Record the network parameters of training and learning, and obtain an improved quantum neural network with a multi-layer activation function of 15 inputs and 4 outputs; After the network is established according to the above steps, the plate shape recognition is carried out.