Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a hot continuous rolling strip steel head width prediction method integrating a rolling mechanism and deep learning, which comprises the following steps:
step 1: the method comprises the steps that production data of the same measuring position of M different strip steel heads in a hot continuous rolling site are obtained, wherein each strip steel head corresponds to a group of production process data, and the production data comprise each type of measurement data detected by each instrument arranged on a hot continuous rolling production line and each type of parameter data in rolling regulation data issued by a process automation level of hot continuous rolling production;
step 2: removing outlier data from production data by using a Pauta criterion to obtain m sample data;
and step 3: dividing all the same type of data representing the width of the head of the strip steel in the sample data into a reference sequence, and dividing each type of data remaining in the sample data into a comparison sequence;
and 4, step 4: screening the data in the comparison sequence according to the influence factors of rolling and spreading to obtain N groups of influence factor data influencing the head width of the strip steel;
and 5: constructing a rolling mechanism prediction model of each frame, and calculating a prediction reference value of the head width of the hot continuous rolling strip steel according to the influence factor data;
step 6: subtracting the prediction reference value of the head width of each strip steel from the data representing the head width of the strip steel in the reference sequence to obtain prediction deviation value data;
and 7: eliminating dimension difference of each type of influence factor data by adopting a min-max standardization method to obtain standardization data;
and 8: taking the standardized data corresponding to the influence factor data as input data of the deep belief neural network model, taking the predicted deviation value data as output data of the deep belief neural network model, and training the model to obtain the deep belief neural network model with optimal parameters;
and step 9: predicting production data of a measuring position of the head of the strip steel to be processed at an outlet by using a depth confidence neural network model with optimal parameters to obtain a prediction correction value of the width of the head of the strip steel;
step 10: and adding the prediction reference value and the prediction correction value of the strip steel head width to obtain the final prediction value of the strip steel head width at the outlet of the measurement position.
The step 2 comprises the following steps:
and (3) taking the Pauta criterion described in the formula (1) as a screening criterion, judging the data meeting the criterion as outlier data and removing the outlier data:
in the formula: y is_{i}The values in the production data which characterize the strip head exit width are indicated, i ═ 1,2,3, …, M,is y_{i}Average value of (1), S_{y}Is y_{i}Standard deviation of (2).
The step 5 comprises the following steps:
step 5.1: according to the thickness of the outlet of the rack and the flow equation of second, the thickness h of the inlet of the rack is inversely calculated by using a formula (2)_{0}：
In the formula, h_{1}Indicating the thickness of the exit of the rack, v_{0}Indicating the gantry entrance velocity, v_{1}Representing the gantry exit velocity;
step 5.2: calculating the contact length l of the deformation zone by using the formula (3)_{c}：
Wherein R represents a roll radius;
step 5.3: calculating the broadening coefficient S by using Hill formula shown in formula (4)_{B}：
In the formula, b_{0}Representing the rack entrance width, C being a constant;
step 5.4: calculating the width expansion DB of the flat roll rolling in the finish rolling area by using a formula (5):
step 5.5: calculating the gantry exit width b using equation (6)_{1}：
b_{1}＝b_{0}+DB (6)
Further, for a hot continuous rolling finishing mill group with multiple racks, calculating the outlet width output by the rolling mechanism prediction model of the previous rack as the inlet width of the next rack from rack to rack according to the running direction of the production line until the outlet width of the last rack is calculated as the prediction reference value of the head width of the hot continuous rolling strip steel;
the step 7 comprises the following steps:
calculating corresponding standardized data x 'after dimension difference of data in the influencing factor data set is eliminated by using formula (7)'_{jk}，
In the formula, x_{jk}Representing the kth data element, x, in the jth class of data_{jmin}Denotes the minimum value, x, in class j data_{jmax}Representing the maximum value in the jth data, and N representing the number of data types in the influence factor data set;
the bottom layer of the depth confidence network model adopts an unsupervised pre-trained restricted Boltzmann machine model, the top layer adopts an error inverse propagation regression model with supervision and fine adjustment, the activation function adopts a ReLU function, and the regularization method adopts a dropout method to prevent overfitting.
Training the model in the step 8 to obtain a depth confidence neural network model with optimal parameters, which is specifically expressed as follows:
step 8.1: setting an initial learning rate as alpha, the number of initial hidden layers as A, the number of initial nodes of the hidden layers as B and the maximum iteration number as x;
step 8.2: setting the updating step length of the node number as b, updating the node number in each iteration by using the step length b, calculating the mean square error after each iteration by using a formula (8), and taking the node number corresponding to the iteration with the minimum mean square error value as the optimal node number of the hidden layer when the maximum iteration times x is reached
Where MSE represents the mean square error value, Y_{k}A predicted deviation value of the input is represented,a predictive modification value representing a head width output by the deep belief neural network model;
step 8.3: the number of nodes of each hidden layer is set asSetting the updating step length of the layer number as a, updating the hidden layer number in each iteration by using the step length a, calculating the mean square error after each iteration by using a formula (8), and taking the layer number corresponding to the iteration with the minimum mean square error value as the optimal layer number of the hidden layer when the maximum iteration number x is reached
Step 8.4: the number of nodes of each hidden layer is set asThe number of the hidden layers is set asSetting learning rateAnd (3) updating the learning rate of each iteration by using the step length d, calculating the mean square error after each iteration by using a formula (8), and taking the learning rate corresponding to the iteration with the minimum mean square error value as the optimal learning rate of the model when the maximum iteration times x is reached.
The invention has the beneficial effects that:
the invention provides a hot continuous rolling strip steel head width prediction method integrating a rolling mechanism and deep learning, which is characterized in that outlier data are removed by applying a Pauta criterion according to actual production data of a hot continuous rolling field, so that a model is not interfered by an abnormal value in a modeling process; the modeling method can overcome the defects that the width error predicted by a rolling mechanism prediction model is large or the width interpretability predicted by a neural network prediction model is poor and the reliability is low. In addition, the selected deep belief neural network has higher prediction precision due to the structural characteristics of multiple hidden layers, and combines the characteristics of the training modes of unsupervised pre-training and supervised fine-tuning, so that the model has higher convergence speed and is not easy to fall into a local extremum. The method has the advantages of high prediction precision, strong generalization capability, high reliability and easy maintenance of the model, solves the problem of weak capability of the traditional width prediction model in adapting to the actual production process, saves the production investment cost and provides a good foundation for adjusting the parameters of the process automation level setting model.
Detailed Description
The invention is further described with reference to the following figures and specific examples.
As shown in fig. 1, a method for predicting the head width of a hot continuous rolling strip steel by combining a rolling mechanism and deep learning, comprises the following steps:
step 1: the method comprises the steps that production data of the same measuring position of M different strip steel heads in a hot continuous rolling site are obtained, wherein each strip steel head corresponds to a group of production process data, and the production data comprise each type of measurement data detected by each instrument arranged on a hot continuous rolling production line and each type of parameter data in rolling regulation data issued by a process automation level of hot continuous rolling production;
in this embodiment, a typical finishing mill group of a hot continuous rolling line is adopted, and the arrangement of main equipment and detection instruments of the rolling line is shown in fig. 2, wherein RE represents a rough rolling edger, R represents a rough rolling plain-barreled mill, FE represents a finish rolling edger, and F represents a finish rolling plain-barreled mill. In the hot continuous rolling production process, the production data mainly comes from the actual data measured by the detecting instrument and the rolling regulation data calculated in the process automation level, wherein the measured data used in the embodiment includes: the device comprises a thickness gauge, a width gauge, a temperature gauge, a speed sensor, a rolling mill pressure sensor, a position sensor, a rolling mill power sensor, a rolling mill rotating speed sensor and an angle sensor, wherein the thickness gauge measures the thickness of a rolled piece, the width gauge measures the width of the rolled piece, the surface temperature of the rolled piece measured by the temperature gauge, the linear speed of a rolling mill measured by the speed sensor, the rolling force of the rolled piece in a deformation process measured by the rolling mill pressure sensor, the roll gap and roll gap deviation measured by the position sensor, the rolling mill power measured by the power sensor, the rolling mill rotating speed measured by the rotating speed sensor and the loop angle deviation measured by the angle sensor. The measuring signal generated by the detecting instrument is transmitted to the process automation level from the basic automation level; the parameter data in the rolling schedule data used in the present embodiment includes: the method comprises the following steps of setting the thickness of an intermediate billet, setting the tension between finishing mill frames, setting the rolling force of a finishing vertical roll, setting the roll gap of the finishing vertical roll, setting the linear speed of the finishing vertical roll, setting the diameter of a roll and compensating coefficient of roll abrasion. All the data constitute the production data of the hot continuous rolling field acquired by the embodiment.
Step 2: removing outlier data from production data by using a Pauta criterion to obtain m sample data, wherein the method comprises the following steps:
and (3) taking the Pauta criterion described in the formula (1) as a screening criterion, judging the data meeting the criterion as outlier data and removing the outlier data:
in the formula: y is_{i}Data representing the width of the strip head outlet in the production data, i ═ 1,2,3, …, M,is y_{i}Average value of (1), S_{y}Is y_{i}Standard deviation of (2).
In this embodiment, the production data of the M-2744 group is divided into 685mm, 710mm, 735mm and 737mm according to the target width of the finished product, and the production data is removed by using the Pauta criterion. The rejection results are shown in table 1. Finally, 14 groups of outlier data are removed altogether, and sample data m is selected to be 2730 groups.
TABLE 1 outlier rejection results
And step 3: dividing all the same type of data representing the width of the head of the strip steel in the sample data into a reference sequence, and dividing each type of data remaining in the sample data into a comparison sequence;
and 4, step 4: screening the data in the comparison sequence according to the influence factors of rolling and spreading to obtain N groups of influence factor data influencing the head width of the strip steel;
because of the influence of factors of main fertilization vertical roll parameters, the reduction rate of a finish rolling flat roll and the tension between finish rolling stands in the width expansion of a finish rolling area, 48 groups of influence factor data are finally selected in the embodiment, wherein the influence factor data comprise finish rolling inlet temperature, finish rolling outlet temperature, intermediate billet thickness, finish rolling outlet thickness, finish rolling finished product target width, rough rolling outlet width, finish rolling vertical roll rolling force, finish rolling vertical roll gap, finish rolling vertical roll linear velocity, roll diameter, roll linear velocity, roll wear compensation, tension between stands and thickness back value in a rolling mill unit F1-F8;
and 5: constructing a rolling mechanism prediction model of each frame, and calculating a prediction reference value of the head width of the hot continuous rolling strip steel according to the influence factor data, wherein the method comprises the following steps:
step 5.1: according to the thickness of the outlet of the rack and the flow equation of second, the thickness h of the inlet of the rack is inversely calculated by using a formula (2)_{0}：
In the formula, h_{1}Indicating the thickness of the exit of the rack, v_{0}Indicating the gantry entrance velocity, v_{1}Representing the gantry exit velocity;
step 5.2: calculating the contact length l of the deformation zone by using the formula (3)_{c}：
Wherein R represents a roll radius;
step 5.3: calculating the broadening coefficient S by using Hill formula shown in formula (4)_{B}：
In the formula, b_{0}Representing the width of the entrance of the rack, wherein C is a constant and is generally equal to 0.5;
step 5.4: calculating the width expansion DB of the flat roll rolling in the finish rolling area by using a formula (5):
step 5.5: calculating the gantry exit width b using equation (6)_{1}：
b_{1}＝b_{0}+DB (6)
For a single stand rolling mill, the stand exit width b calculated by step 5.5_{1}The prediction reference value of the head width of the strip steel is obtained, but for a hot continuous rolling finishing mill group with a plurality of frames, the outlet width output by a rolling mechanism prediction model of the previous frame is taken as the inlet width of the next frame to be calculated frame by frame according to the running direction of a production line until the outlet width of the last frame is calculated to be the prediction reference value of the head width of the hot continuous rolling strip steel;
the prediction flow chart of the rolling mechanism model in the present invention is shown in fig. 3. In this embodiment, the algorithm for calculating the thickness between the stands according to the second flow equation in the formula (2) has been written into the process automation level program of the production field, and the field can automatically calculate the thickness between the finishing mill stands according to the instrument parameters and store the thickness. The method is specifically realized in a way that the formula (2) is programmed and packaged into a function by C + + language, the hot continuous rolling finishing mill group with multiple racks calls the packaged function of the formula (2) to calculate the outlet thickness of the previous rack according to the final rack roller linear speed and the finishing rolling outlet thickness measured by the instrument, and the process is repeated until the outlet thickness of the first rack is calculated.
In this example, since 1 vertical rolling mill and 8 flat rolling mills are provided in the finish rolling zone, the calculation work of step 5.2 to step 5.5 is repeated with the finish rolling vertical roll gap as a starting point. And calculating the outlet width of the previous frame as the inlet width of the frame one by one until the outlet width of the last frame is calculated to be the prediction reference value of the head width of the hot continuous rolling strip steel, wherein the head width of the strip steel refers to the outlet width of the head of the strip steel at the outlet position of the last frame, and the outlet width of the frame refers to the outlet width of the head of the strip steel at the outlet position of the middle frame.
Step 6: subtracting the prediction reference value of the head width of each strip steel from the data representing the head width of the strip steel in the reference sequence to obtain prediction deviation value data;
and 7: eliminating dimension difference of each type of influence factor data by adopting a min-max standardization method to obtain standardization data, wherein the standardization data comprises the following steps:
calculating corresponding standardized data x 'after dimension difference of data in the influencing factor data set is eliminated by using formula (7)'_{jk}，
In the formula, x_{jk}Representing the kth data element, x, in the jth class of data_{jmin}Denotes the minimum value, x, in class j data_{jmax}Representing the maximum value in the jth data, and N representing the number of data types in the influence factor data set;
and 8: taking the standardized data corresponding to the influence factor data as input data of the deep belief neural network model, taking the predicted deviation value data as output data of the deep belief neural network model, and training the model to obtain the deep belief neural network model with optimal parameters;
the bottom layer of the deep belief neural network model adopts an unsupervised pre-trained restricted Boltzmann machine model, the top layer adopts an error inverse propagation regression model with supervision and fine adjustment, the activation function adopts a ReLU function, the regularization method adopts a dropout method to prevent overfitting, and the discarding probability is 0.3.
Training the model in the step 8 to obtain a depth confidence neural network model with optimal parameters, which is specifically expressed as follows:
step 8.1: setting an initial learning rate alpha to be 0.0001, the number of initial hidden layer layers to be A to be 2, the number of initial nodes of the hidden layers to be B to be 50 and the maximum iteration number to be x;
step 8.2: setting the updating step length of the node number as b being 50, updating the node number in each iteration by using the step length b being 50, calculating the mean square error after each iteration by using a formula (8), and taking the node number corresponding to the iteration with the minimum mean square error value as the optimal node number of the hidden layer when the maximum iteration number x is reached
Where MSE represents the mean square error value, Y_{k}A predicted deviation value of the input is represented,a predictive modification value representing a head width output by the deep belief neural network model;
step 8.3: the number of nodes of each hidden layer is set asSetting the updating step length of the layer number as a being 1, updating the hidden layer number in each iteration by the step length a being 1, calculating the mean square error after each iteration by using a formula (8), and taking the layer number corresponding to the iteration with the minimum mean square error value as the optimal layer number of the hidden layer when the maximum iteration number x is reachedThe selection results of the hidden layer structure are shown in table 2;
TABLE 2 Effect of hidden layer Structure on deep belief network model
Step 8.4: the number of nodes of each hidden layer is set asThe number of the hidden layers is set asSetting the updating step length of the learning rate to be 0.0003, updating the learning rate in each iteration by the step length d, calculating the mean square error after each iteration by using a formula (8), and taking the learning rate corresponding to the iteration with the minimum mean square error value as the optimal learning rate of the model when the maximum iteration times x is reachedThe results of selecting the learning rate are shown in table 3.
TABLE 3 Effect of learning Rate on deep belief network model
And step 9: predicting production data of a measuring position of the head of the strip steel to be processed at an outlet by using a depth confidence neural network model with optimal parameters to obtain a prediction correction value of the width of the head of the strip steel;
step 10: and adding the prediction reference value and the prediction correction value of the strip steel head width to obtain the final prediction value of the strip steel head width at the outlet of the measurement position.
Specifically, due to the deep structure characteristics of the deep confidence network and the training mode combining unsupervised pre-training and supervised fine tuning, the width correction value predicted based on the deep confidence network model has the characteristics of high prediction precision, strong generalization capability and difficulty in falling into a local extreme value. The detailed structure and training process of the model is as follows:
the Deep Belief Network (DBN) is a probabilistic generation model, which is formed by stacking a plurality of constrained Boltzmann machines (RBMs). The DBN bottom-most layer receives the input data vector and transforms the input data to the hidden layer via the RBM, i.e., the input to a higher layer of RBMs is from the output of a lower layer of RBMs. An RBM is composed of a visible layer and a hidden layer, and the neurons of the visible layer and the neurons of the hidden layer are in full bidirectional connection. Assuming that a certain RBM visible layer has V neurons and the hidden layer has H neurons, for a given state (V, H), the energy function is defined as follows:
where θ ═ { W ', a ', b ' } is a parameter of the RBM, where W ' denotes a connection weight between the visible layer and the hidden layer, W '_{f,g}Represents the connection weight between the visible unit f and the invisible unit g, a 'represents the offset of the visible layer, a'_{f}Denotes the bias of the visible unit f, b 'denotes the bias of the hidden layer, b' denotes the bias of the hidden unit g, H denotes the number of hidden layer neurons, V denotes the number of visible layer neurons, E_{θ}(v, h) represents the energy function when the hidden layer neuron is h and the visible layer neuron is v.
Based on the above energy functions, the joint probability distribution for a given state (v, h) is given by:
in the formula Z_{θ}The allocation function is represented.
Due to the special structure that RBM layers are connected with each other and not connected in the layers, when the state of each neuron of the visible layer is given, the activation states of each neuron of the hidden layer are mutually independent, and similarly, when the state of each neuron of the hidden layer is given, the activation states of each neuron of the visible layer are also mutually independent, so the activation probabilities of the g hidden layer neuron and the f visible layer neuron are respectively as follows:
where σ represents the activation function.
Due to the distribution function Z_{θ}Are difficult to calculate, resulting in a joint probability distribution p_{θ}(v, h) cannot be calculated. And a contrast divergence algorithm can be adopted to accelerate RBM training learning. Training the RBM through a contrast divergence algorithm, wherein each parameter updating rule is as follows:
W′＝W′+ρ(hv^{T}-h′(v′)^{T} (14)
b′＝b′+ρ(h-h′) (15)
a′＝a′+ρ(v-v′) (16)
in the formula: v ' represents the reconstruction of the visible layer v, h ' represents the hidden layer obtained from the reconstruction v ', and ρ represents the learning rate.
However, stacking the RBMs can only obtain some high-level features from complex raw data, and cannot perform direct regression prediction on the data, and in order to obtain a complete DBN model, a conventional supervised regressor needs to be added at the topmost layer of the stacked RBMs. The basic structure of a DBN is shown in fig. 4. As can be seen from the figure, the training process of the DBN consists of two processes, unsupervised layer-by-layer pre-training and supervised fine-tuning. And (3) forming an RBM by two adjacent layers of neurons, performing unsupervised pre-training on the RBM layer by layer from bottom to top, inputting a final result into a supervised regression device at the top layer, and performing fine adjustment on the network weight and the bias by adopting a back propagation algorithm. The overall training process is shown in fig. 5.
In the embodiment, 2730 groups of screened actual data are used as experimental data of the model, the rolling mechanism prediction model established based on the rolling mechanism predicts the reference value of the head width of the strip steel in the finish rolling process, and the prediction model established based on the depth confidence neural network predicts the correction value of the head width of the strip steel in the finish rolling process. The data are randomly divided into 2180 groups of training sets and 550 groups of testing sets, and the training sets and the testing sets are applied to the prediction model based on the deep belief neural network. And selecting influence factor data of 48 groups of variables including a finish rolling inlet temperature, a finish rolling outlet temperature, an intermediate billet thickness, a finish rolling outlet thickness, a finish rolling finished product target width, a rough rolling outlet width, a finish rolling vertical roll rolling force, a finish rolling vertical roll gap, a finish rolling vertical roll linear velocity, and roll diameters, roll linear velocities, roll wear compensation, tension between frames and thickness back calculation values in the mill train F1-F8 as the input of a depth confidence network prediction model, and summing a head width reference value predicted by the rolling mechanism prediction model and a head width correction value predicted by the depth confidence neural network model to obtain a final predicted value of the head width of the strip steel. The embodiment is implemented by using python language programming, and the obtained predicted result and actual measurement result are compared as shown in fig. 6.
In conclusion, compared with the traditional method for predicting the width of the head of the finish rolling, the method for predicting the width of the head of the hot continuous rolling strip steel, which integrates the rolling mechanism and the depth confidence neural network, is accurate and efficient in predicting the width of the head of the strip steel in the finish rolling process. The method has better generalization performance for the strip steels with different finished product target specifications; meanwhile, the defects that the width is predicted only by a rolling mechanism, the precision is low, the capability of adapting to actual production is weak, the reliability of the width is predicted only by a neural network is low, and the interpretability is poor are overcome. In addition, the prediction model is constructed based on the rolling mechanism and deep learning without improving the equipment of the existing hot continuous rolling production line, so that the production investment cost is saved, and a good foundation is provided for the adjustment of the process automation level setting model parameters.