阅读说明：本技术 一种基于lstm的热轧弯辊力预测方法 (Hot rolling bending force prediction method based on LSTM ) 是由 李旭 栾峰 王琳 吴艳 韩月娇 张殿华 于 2020-12-04 设计创作，主要内容包括：本发明涉及一种基于LSTM的热轧弯辊力预测方法,采集不锈钢轧机热轧工艺的最终机架轧制数据并划分为训练集traindata和测试集testdata；对traindata进行归一化处理；构造矩阵P；将矩阵P的最后一行作为训练集的标签即真实值；对网络的输出值和真实值计算并更新；网络训练完成后,取出LSTM网络的最后m个输出数据作为下一时刻的输入,得到网络下一时刻的输出,即为下一时刻的弯辊力预测值；重复上述步骤直至获得足够数目的预测数据；将处理后的数据与testdata中的真实值进行比较,检验网络的有效性。本发明方法与传统LSTM网络相比,加入了更新机制后的网络模型准确度有所提升,且网络更加稳定。(The invention relates to a hot rolling roll bending force prediction method based on LSTM, which comprises the steps of collecting the final frame rolling data of a stainless steel rolling mill hot rolling process and dividing the final frame rolling data into a training set tranndadata and a test set testdata; carrying out normalization processing on the traindata; constructing a matrix P; taking the last row of the matrix P as a label of a training set, namely a true value; calculating and updating the output value and the real value of the network; after the network training is finished, taking out the last m output data of the LSTM network as the input of the next moment to obtain the output of the network at the next moment, namely the predicted value of the bending force at the next moment; repeating the above steps until a sufficient number of predicted data are obtained; and comparing the processed data with the true value in the testdata, and checking the validity of the network. Compared with the traditional LSTM network, the method of the invention improves the accuracy of the network model after adding the updating mechanism, and the network is more stable.)

1. A hot rolling bending force prediction method based on LSTM is characterized by comprising the following steps:

1) collecting final rack rolling data of a hot rolling process of a stainless steel rolling mill, and collecting roll bending force data for an experiment;

2) dividing the roll bending force data into a training set tranndata and a test set testdata in a time sequence according to a specified proportion;

3) carrying out normalization processing on the tranndata to obtain a vector after normalization;

4) constructing a matrix P by using the vector A in the step 3);

5) taking the first m rows of the matrix P in the step 4) as input and sending the input to the LSTM network;

6) taking the last row of the matrix P as a label of a training set, namely a true value; calculating the output value and the true value of the network by using a formula to obtain an error, and updating the weight and the offset of the network by using a gradient descent method;

7) after the network training is finished, taking out the last m output data of the LSTM network as the input of the next moment so as to obtain the output of the network at the next moment, wherein the output is the predicted value of the bending force at the next moment;

8) repeating step 7) until a sufficient number of prediction data are obtained;

9) and performing inverse normalization processing on the obtained roll bending force predicted value, comparing the processed data with the actual value in the testdata, and checking the effectiveness of the network.

2. The LSTM-based hot rolling bending force prediction method of claim 1, wherein: in step 4), the matrixWherein the parameter m is the number of input layer units of the neural network, and n is the number of samples included in the training set of the neural network.

3. The LSTM-based hot rolling bending force prediction method of claim 1, wherein: in step 5), the LSTM network may adopt a traditional LSTM network, or may adopt an ON-LSTM network or a double-layer ON-LSTM network, and the double-layer ON-LSTM network is adopted as follows:

taking the first m rows of the matrix P in the step 4) as the input of the first layer LSTM, and sending the obtained output data as the input to the second layer LSTM, wherein the output data of the second layer LSTM is the output of the whole network.

4. The LSTM-based hot rolling bending force prediction method of claim 3, wherein: the first layer and the second layer of the double-layer ON-LSTM network introduce ordered neurons LSTM of an updating mechanism, namely ON-LSTM, and when the LSTM is subjected to parameter updating, the neurons are ordered and the importance degree, namely the hierarchical structure, is introduced; information with higher hierarchy represents important information and needs to be reserved in the network; in contrast, the low-level information represents unimportant information, which needs to be updated with new input data, and the detailed process and calculation formula are as follows:

assuming that the primary hierarchical position corresponding to the important information is represented by S1 and the secondary hierarchical position corresponding to the unimportant information is represented by S2, S1 and S2 are calculated as:

the function of the indexmax function is to calculate a position sequence number corresponding to the maximum element in the vector; x is the number of_tTo input data, h_t-1In order to be able to recur the data,is a master forgetting gate weight matrix,is the main forgetting gate offset and is,is a matrix of the weight of the primary input gate,biasing the main input gate;

when S2 is more than or equal to S1, the positions corresponding to the important information and the unimportant information are partially overlapped, and the current cell state c_tThe calculation formula of (2):

when S2< S1, the positions corresponding to the important information and the secondary information are independent of each other, and the current cell state ct is calculated by the following formula:

wherein k is c_tDimension of (d), f_tAnd i_tRespectively a forgetting gate output and an input gate output,is an intermediate cell state, c_t-1The cell state at the previous time.

5. The LSTM-based hot rolling bending force prediction method of claim 4, wherein: respectively to be provided withAnddefining as a main forgetting gate and a main input gate; w is a_t1、w_t2And w_t3Representing the high, medium and low levels in the hierarchy, respectively, there are:

wherein the content of the first and second substances,is mainly output by the forgetting gate,is the output of the main input gate.

Technical Field

The invention relates to a rolling bending force prediction technology, in particular to a hot rolling bending force prediction method based on LSTM.

Background

The high-precision rolling requires a rolled material plate shape control system to have high precision, quick response and disturbance resistance. The principle is that the roll bending force is applied to a working roll and a roll neck of a supporting roll to enable the roll to generate bending deformation to change the convexity of a roll gap, so that the shape of the loaded roll gap in the rolling process is influenced, the shape of the strip steel is changed along the width direction, the bad plate shape caused by the change of other rolling process factors is compensated, and the precision of the plate shape of the strip steel outlet is ensured. Therefore, the prediction precision of the bending roll force has important influence on the shape and the shape control precision, especially the head of the strip steel. The higher roll bending force prediction accuracy is beneficial to the closed-loop feedback control of the roll bending force prediction. In actual production, the roll bending force is generally calculated comprehensively according to the strip steel temperature, thickness, width, rolling force, material, roll thermal expansion, roll wear, flatness and convexity and other influence factors. However, the existing bending force prediction model based on the traditional mathematical method is very complex in structure, and each influence factor has the characteristics of nonlinearity, strong coupling and time-varying property, so that the precision improvement of the bending force prediction model is seriously influenced. Therefore, it is necessary to search a new roll bending force prediction model.

The neural network is widely applied to many fields by the unique property thereof, and is characterized by being capable of completing system modeling with complex internal relations and solving the traditional nonlinear problem. In addition, the method has good robustness and fault tolerance. In recent years, many scholars gradually apply the neural network algorithm to the field of roll bending force prediction and achieve better prediction effect. Presetting and controlling the roll bending force of the hot continuous rolling mill according to a genetic algorithm; presetting and optimizing the roll bending force of the BP neural network; and optimizing a hot-rolled strip steel roll bending force forecasting model of the neural network by a genetic algorithm. However, most of the existing neural network roll bending force prediction models need to acquire a plurality of parameters affecting roll bending force as inputs, and the parameters are dispersed in all components of the whole rolling process, so that the data acquisition work difficulty is large, and the task is heavy. Meanwhile, more input parameters not only increase the task of data processing, but also easily cause the complexity of the model, thereby influencing the training speed and the prediction precision of the network. The analysis of the rolling process can find that in the actual rolling process, the rolling bending forces are continuous, each rolling bending force is related to the previous stage, and therefore the rolling bending forces have the time series characteristic, and the common neural network does not consider time context information, so that the common neural network is difficult to learn the trend of the time series, the accuracy of a prediction model is low, and the accurate prediction of the bending forces at the subsequent moment is difficult to make.

Disclosure of Invention

Aiming at the defects in the prior art, the invention aims to provide a hot rolling bending force prediction method based on LSTM, which can accurately and efficiently predict the hot rolling strip steel bending force.

In order to solve the technical problems, the invention adopts the technical scheme that:

the invention provides an LSTM-based hot rolling bending force prediction method, which comprises the following steps of:

1) collecting final rack rolling data of a hot rolling process of a stainless steel rolling mill, and collecting roll bending force data for an experiment;

2) dividing the roll bending force data into a training set tranndata and a test set testdata in a time sequence according to a specified proportion;

3) carrying out normalization processing on the tranndata to obtain a vector after normalization;

4) constructing a matrix P by using the vector A in the step 3);

5) taking the first m rows of the matrix P in the step 4) as input and sending the input to the LSTM network;

6) taking the last row of the matrix P as a label of a training set, namely a true value; calculating the output value and the true value of the network by using a formula to obtain an error, and updating the weight and the offset of the network by using a gradient descent method;

7) after the network training is finished, taking out the last m output data of the LSTM network as the input of the next moment so as to obtain the output of the network at the next moment, wherein the output is the predicted value of the bending force at the next moment;

8) repeating step 7) until a sufficient number of prediction data are obtained;

9) and performing inverse normalization processing on the obtained roll bending force predicted value, comparing the processed data with the actual value in the testdata, and checking the effectiveness of the network.

In step 4), the matrixWherein the parameter m is the number of input layer units of the neural network, and n is the number of samples included in the training set of the neural network.

In step 5), the LSTM network may adopt a traditional LSTM network, or may adopt an ON-LSTM network or a double-layer ON-LSTM network, and the double-layer ON-LSTM network is adopted as follows:

taking the first m rows of the matrix P in the step 4) as the input of the first layer LSTM, and sending the obtained output data as the input to the second layer LSTM, wherein the output data of the second layer LSTM is the output of the whole network.

The first layer and the second layer of the double-layer ON-LSTM network introduce ordered neurons LSTM of an updating mechanism, namely ON-LSTM, and when the LSTM is subjected to parameter updating, the neurons are ordered and the importance degree, namely the hierarchical structure, is introduced; information with higher hierarchy represents important information and needs to be reserved in the network; in contrast, the low-level information represents unimportant information, which needs to be updated with new input data, and the detailed process and calculation formula are as follows:

assuming that the primary hierarchical position corresponding to the important information is represented by S1 and the secondary hierarchical position corresponding to the unimportant information is represented by S2, S1 and S2 are calculated as:

S1＝F₁(x_t,h_t-1)＝index max(soft max(W_f～x_t+U_f～h_t-1+b_f～))

S2＝F₂(x_t,h_t-1)＝index max(soft max(W_i～x_t+U_i～h_t-1+b_i～))

wherein the function of indexmax is to find the position order corresponding to the largest element in the vectorNumber; x is the number of_tTo input data, h_t-1For recursive data, W_f～、U_f～As a master forgetting gate weight matrix, b_f～Is a main forgetting gate bias, W_i～、U_i～Is a master input gate weight matrix, b_i～Biasing the main input gate;

when S2 is more than or equal to S1, the positions corresponding to the important information and the unimportant information are partially overlapped, and the current cell state c_tThe calculation formula of (2):

when S2< S1, the positions corresponding to the important information and the secondary information are independent of each other, and the current cell state ct is calculated by the following formula:

wherein k is c_tDimension of (d), f_tAnd i_tRespectively a forgetting gate output and an input gate output,is an intermediate cell state, c_t-1The cell state at the previous time.

Respectively to be provided withAnddefining as a main forgetting gate and a main input gate; w is a_t1、w_t2And w_t3Representing the high, medium and low levels in the hierarchy, respectively, there are:

wherein the content of the first and second substances,is mainly output by the forgetting gate,is the output of the main input gate.

The invention has the following beneficial effects and advantages:

1. according to the hot-rolling bending force prediction method based on the LSTM, an LSTM (Long Short-Term Memory) neural network model is selected, and the bending force data at the future moment can be effectively predicted only by inputting a certain number of bending force historical data into a network; in addition to traditional LSTM, predictions are made using ON-LSTM networks with added update mechanisms. Compared with the traditional LSTM network, the accuracy of the network model added with the updating mechanism is improved, and the network is more stable.

2. The invention integrates an updating mechanism with a double-layer structure, improves the network performance, further improves the prediction precision of the network, and proves the effectiveness of the LSTM (traditional LSTM, ON-LSTM, double-layer ON-LSTM) network model provided by the invention through carrying out experiments ON the three data sets of the traditional LSTM, ON-LSTM and double-layer ON-LSTM networks and carrying out comparison experiments ON the Artificial Neural Network (ANN) by using the same data set.

Drawings

FIG. 1 is a flow chart of the hot rolling bending force prediction method based on LSTM according to the present invention;

FIG. 2 is a diagram of a conventional LSTM hidden layer structure;

FIG. 3 is a diagram of hidden layer structure of ON-LSTM;

FIG. 4 is a diagram of a dual layer LSTM structure;

FIG. 5A is a data set of error contrast maps for four network models (one);

FIG. 5B is a data set of error contrast plots for the four network models (two);

FIG. 5C is a data set of error contrast plots for the four network models (three);

FIG. 6A is a roll force prediction result dataset (one) for a conventional LSTM network;

FIG. 6B is a data set of roll bending force prediction results for a conventional LSTM network (two);

FIG. 6C is a data set of roll bending force prediction results for a conventional LSTM network (III);

FIG. 7A is a set of roll force prediction results data (one) for an ON-LSTM network;

FIG. 7B is a set of roll force prediction results data for the ON-LSTM network (two);

FIG. 7C is a set of roll force prediction results data for the ON-LSTM network (III);

FIG. 8A is a roll force prediction result dataset (one) for a two-layer ON-LSTM network;

FIG. 8B is a data set of roll bending force prediction results for a two-layer ON-LSTM network (II);

FIG. 8C is a set of roll force prediction results data for a two-layer ON-LSTM network (III);

FIG. 9A is a roll force prediction result dataset (one) for an ANN network;

FIG. 9B is a bending force prediction result dataset for an ANN network (two);

fig. 9C is the bending force prediction result dataset (three) for the ANN network.

Detailed Description

The invention is further elucidated with reference to the accompanying drawings.

As shown in fig. 1, the present invention provides an LSTM-based hot rolling bending force prediction method, comprising the steps of:

1) collecting final frame rolling data of a 1580mm hot rolling process of a stainless steel rolling mill, and collecting roll bending force data for an experiment;

2) dividing the roll bending force data into a training set tranndata and a test set testdata in a time sequence according to a specified proportion;

3) normalizing the input data to obtain a normalized vector, specifically:

performing linear normalization processing on the tranndata to obtain a normalized vector A ═ a₁,a₂,......a_n)^TWherein the parameter n is the number of samples contained in the neural network training set, and a 1-an are the samples contained in the neural network training set;

4) constructing a matrix P by using the vector A in the step 3);wherein the parameter m is the number of input layer units of the neural network, and n is the number of samples;

5) taking the first m rows of the matrix P in the step 4) as input and sending the input to the LSTM network;

6) taking the last row of the matrix P as a label of a training set, namely a true value; calculating the output value and the true value of the network by using a formula to obtain an error, and updating the weight and the offset of the network by using a gradient descent method;

7) after the network training is finished, taking out the last m output data of the LSTM network as the input of the next moment so as to obtain the output of the network at the next moment, wherein the output is the predicted value of the bending force at the next moment;

8) repeating step 7) until a sufficient number of prediction data are obtained;

9) and performing inverse normalization processing on the obtained roll bending force predicted value, comparing the processed data with the actual value in the testdata, and checking the effectiveness of the network.

In step 5), the LSTM network may adopt a conventional LSTM network (hidden layer structure is shown in fig. 2), or may adopt an ON-LSTM network (hidden layer structure is shown in fig. 3) or a dual-layer ON-LSTM network (hidden layer structure is shown in fig. 4).

And (3) adopting an ON-LSTM network, wherein the steps 1) to 9) are the same as those of the traditional LSTM network.

Adopting a double-layer ON-LSTM network, wherein the steps 1) to 4) and 6) to 9) are the same as those of the traditional LSTM network, and the step 5) is; the first m rows of the matrix P in step 4) are used as inputs for the first layer LSTM and the resulting output data is sent as input to the second layer LSTM. The output data of the second layer LSTM is the output of the network as a whole.

In step 5), the first layer and the second layer of LSTM are ordered neuron LSTM introduced with update mechanism, namely ON-LSTM. The updating mechanism has the specific effect that when the LSTM is subjected to parameter updating, all the neurons are ordered according to a certain sequence, and the importance degree, namely the hierarchical structure, is introduced. Higher level information represents important information and needs to be retained in the network. Conversely, information at a lower level indicates less important information. For example, the roll force data at a certain time is greatly different from the previous and subsequent times, a large jump is generated, the trend learning of the roll force data does not greatly contribute to the whole trend learning, and the roll force data is unimportant information, and other data is important information relatively.

It needs to be updated with new input data. The detailed process and the calculation formula are as follows:

it is assumed that the primary hierarchical position corresponding to the important information is represented by S1 and the secondary hierarchical position corresponding to the unimportant information is represented by S2. By x_tAnd h_t-1S1 and S2 are calculated as:

S1＝F₁(x_t,h_t-1)＝index max(softmax(W_f～x_t+U_f～h_t-1+b_f～))

S2＝F₂(x_t,h_t-1)＝index max(softmax(W_i～x_t+U_i～h_t-1+b_i～))

the function of the indexmax function is to find the position sequence number (ordered from 1) corresponding to the maximum element in the vector; xt is the input data, ht-1 is the recursive data, W and U are the weight matrices, and b is the bias.

The network updates ct based on these hierarchical locations. Considering the relative magnitude between S1 and S2, there are two types of update models: (1) when S2 is not less than S1, the positions corresponding to the important information and the secondary information are partially overlapped, and the current cell state c_tThe calculation formula of (2):

when S2<At S1, the positions corresponding to the important information and the minor information are independent of each other, and the current cell state c_tCalculated from the following formula:

wherein k is c_tDimension of (d), f_tAnd i_tRespectively a forgetting gate output and an input gate output,is the intermediate cell state.

For the convenience of describing the above updating process, the following will be describedAnddefining as a main forgetting gate and a main input gate; w is a_t1、w_t2And w_t3Representing the high, medium and low levels in the hierarchy, respectively.

The complete update formula of the ON-LSTM network is as follows:

where the cumax function is shorthand for cumsum (softmax ()).

The updating calculation process of the LSTM network weight and bias mentioned in step 6) is specifically as follows:

the calculation formula of the weighted input error term of the traditional LSTM and ON-LSTM networks is as follows:

with the double-layer ON-LSTM of the double-layer structure, the calculation of the two layers of networks is slightly different when the weight and the bias are updated. The weighted input error term calculation formulas of the first layer and the second layer LSTM are respectively as follows:

(1) first layer LSTM:

(2) second layer LSTM:

wherein f is_tFor forgetting to inputGo out, i_tOutput o for input gate_tFor output of the output gate, c_tAs the current cell state, c_t-1The state of the cells at the previous moment,is an intermediate cell state, x_tTo input data, h_tFor recursive data, h_t-1Is the recursive data of the previous time instant,is mainly output by the forgetting gate,is the output of the main input gate, w_t1Is of high grade, w_t2Is of medium grade, w_t3At a low grade, W_fx、W_ix、W_cx、W_ox、 W_fh、W_ih、W_ch、W_oh、As a weight matrix, b_f、b_i、b_c、b_o、Is a bias vector; delta_tError term for time t, δ_f,t、δ_i,t、And delta_o,tAre respectively f_t、i_t、c_tAnd o_tThe four weights input the corresponding error terms.

Calculation of weight gradient:

calculation of bias gradient:

wherein E is the error of the first image,is h_t-1The transpose of (a) is performed,is x_tTranspose of (W)_oh,tW at time t_ohIn the same way as W_fh,t、W_ih,t、W_ch,t、W_ox,t、W_fx,t、W_ix,t、W_cx,t、b_o,t、b_f,t、b_i,t、b_c,tW at time t_fh、W_ih、W_ch、W_ox、W_fx、W_ix、W_cx、b_o、b_f、b_i、b_c。

According to the continuity and time sequence characteristics of the rolling process, an LSTM (Long Short-Term Memory) neural network model is selected. The LSTM is a long-short term memory network, is a time Recurrent Neural Network (RNN), and mainly aims to solve the problems of gradient extinction and gradient explosion in the long sequence training process. In short, LSTM can perform better in longer sequences than normal RNNs.

The invention selects the LSTM network to predict the roll bending force. Different from the traditional neural network or other machine learning methods, the invention does not need to collect a large number of input parameters influencing the roll bending force in advance, simplifies the preliminary data preparation work, and only needs to form a time sequence by the roll bending force data collected up to the current moment at each moment as input to the network, so that the LSTM network can be trained and the roll bending force at the subsequent moment can be predicted.

The invention adopts three different LSTM network models to predict the roller bending force, and provides two other improved LSTM networks for experiments besides the traditional LSTM network. The two improved LSTM networks are:

(1) ON-LSTM network: compared with the traditional LSTM, the ON-LSTM network can effectively enhance the robustness of the network and improve the accuracy of the network.

(2) Double-layer ON-LSTM: the dual-layer ON-LSTM network organically combines the update mechanism with the dual-layer structure. Because the double-layer structure has strong function fitting capability, the double-layer ON-LSTM can further improve the prediction accuracy of the network ON the basis of the ON-LSTM.

In order to prove the effectiveness of the LSTM (traditional LSTM, ON-LSTM and double-layer ON-LSTM) network model provided by the invention, roll bending force data of a 1580mm hot rolling process of a stainless steel rolling mill are collected and divided into three different data sets, and each data set comprises 500 sample points. Experiments are respectively carried out ON the three data sets of the traditional LSTM, the ON-LSTM and the double-layer ON-LSTM networks, comparison experiments are carried out ON the Artificial Neural Network (ANN) by using the same data set, and the experimental results are compared. The maximum and average errors for the three different network models over the three data sets are shown in fig. 5A-5C below. According to fig. 5A to 5C), it can be seen that the prediction errors of the three LSTM networks are much smaller than those of the ANN network, the three LSTM networks all have good performance ON the data sets 1 to 3 (the data sets 1 to 3 are obtained by cutting and dividing the collected roll bending force data, and three representative different data are selected as the three data sets for experiments), and it can be found that the ON-LSTM network added with the update mechanism is superior to the conventional LSTM network. In addition, after the ON-LSTM is constructed into a double-layer structure, the network performance is further improved, and the double-layer ON-LSTM network has the smallest error among the three LSTM networks and performs best.

The first 450 sample points of the three different data sets are respectively subjected to network training, and the trained LSTM network is used for predicting the bending force data of the future 50 sample points, namely the future 50 moments, and the results are shown in the following fig. 6A-6C, fig. 7A-7C and fig. 8A-8C. The three LSTM networks excel to a high degree and a low degree respectively: double-layer ON-LSTM, traditional LSTM.

The results of comparative experiments were performed by replacing the LSTM network used in the present invention with the ANN network while keeping the conditions of the data set used, the number of training samples, and the number of prediction samples constant, as shown in fig. 9A to 9C below. Experimental results show that the result error obtained by predicting the roll bending force by adopting the ANN network is larger, and the effect is far less than that of the three LSTM networks provided by the patent. And through comparing with the ANN network, the LSTM network has strong advantages in the aspect of processing continuous time sequences, can effectively solve the problem of roll bending force prediction provided by the patent, and has great practical significance and application value.