A cold-rolled plate shape probability prediction method based on a deep belief network

Abstract

The application provides a cold-rolled plate shape probability prediction method based on a deep belief network, and relates to the technical field of metallurgical automation. The method first determines the number of Softmax classifiers required for plate shape prediction according to the number of plate shape roll detection units; and performs plate shape statistics and grid division to obtain the number of categories of the Softmax classifier and a category probability label vector; then establishes a deep belief network stacked by multiple RBM layers and fully connected with each Softmax classifier as a plate shape prediction network; and takes the strip steel rolling data and the plate shape control data as the network input to perform unsupervised learning on the stacked RBM. The overall output value of the stacked RBM network is taken as the input to perform supervised learning on each Softmax classifier respectively. When the plate shape prediction network is used for plate shape prediction, the middle value of the grid plate shape where the maximum probability value of each classifier is located is taken as the initial value of the plate shape of each unit, and finally the plate shape prediction value of each unit plate shape is determined according to the plate shape self-balancing principle.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metallurgical automation, and in particular to a cold-rolled plate shape probability prediction method based on a deep belief network. BACKGROUND

[0002] Plate shape prediction and control technology is a technology with high complexity and technical content in the current steel rolling technology field. The influence function method is widely used in cold-rolled plate shape presetting calculation and multivariable optimal control of cold-rolled plate shape actuator adjustment based on control efficacy, and can better predict common symmetric wave shapes such as middle wave, edge wave, and rib wave. In recent years, in order to reduce inter-roller contact stress and prolong the service life of the roller, an asymmetric intermediate roller profile has been developed; in order to reduce the transverse thickness difference and improve the silicon steel sheet coefficient, a large edge crown roller profile has been developed, but the online application of these special roller profiles also brings the so-called "double edge tight" plate shape with negative edge plate shape values and large absolute values. The existing plate shape prediction model based on the influence function method and the plate shape pattern recognition method based on Legendre or Chebyshev orthogonal polynomials are difficult to accurately predict or recognize this.

[0003] With the continuous application of artificial intelligence technologies such as big data analysis, deep learning, and machine learning in the steel industry, it is necessary to make full use of the massive rolling data and plate shape data accumulated by cold rolling mills over the years to solve the problem that the current plate shape mechanism model and plate shape pattern recognition model are difficult to accurately predict and recognize.

[0004] Patent CN201911387734.9 discloses a cold-rolled strip plate shape prediction method based on deep learning. This method uses a deep neural network (DNN) to analyze the data accumulated during the steel rolling process, determines the factors affecting the plate shape, and uses them as the input features of the neural network. The plate shape value at the mill exit is used as the output result of the output layer of the neural network, thereby constructing a plate shape prediction model, and the Adam algorithm is used for offline training. This patent mainly realizes the prediction of plate shape by increasing the number of hidden layers.

[0005] Patent CN202210444007.7 discloses a cold-rolled plate shape dynamic prediction method based on a deep recurrent neural network. This method mainly establishes a cold-rolled plate shape prediction network for each stand of the cold rolling mill by stacking a recurrent neural network, and uses historical rolling time series data, historical plate shape data, and intermediate stand plate shape calculation data obtained by the influence function method to train the plate shape prediction network offline. This patent mainly uses the characteristics of the recurrent neural network structure and the space-time similarity of the continuous rolling characteristics of the cold rolling mill to predict the plate shape.

[0006] The above patents all take the strip plate shape prediction as a regression problem, and do not specially consider and design the special plate shape distribution such as the "double tight", and there is no research on taking the plate shape prediction as a classification problem.

[0007] A restricted Boltzmann machine (RBM) is a two-layer neural network based on an energy model, the first layer is a visible layer, and the second layer is a hidden layer, and there is no connection in the layer, but full connection between layers. The RBM is a probability generation model, and can make the entire neural network generate training data according to the maximum probability. Hinton proposed an efficient unsupervised learning algorithm for training RBM, namely the contrastive divergence (CD) algorithm.

[0008] A deep belief network (DBN) is a commonly used neural network in machine learning, generally composed of two parts, the first part is a plurality of stacked RBMs, used for network pre-training; and the second part is a feedforward backpropagation network, which can adopt a BP network or a Softmax classifier, used for network supervised learning. The DBN network can be used for dimension reduction, classification, regression, collaborative filtering and feature learning.

[0009] The Softmax classifier can solve the multi-classification problem, and can convert the plurality of probability values output by the stacked RBM into the probability of a plurality of classification results according to formula (1) and formula (2), and ensure that the classification probabilities of all categories add up to 1.

[0010] z i =∑x j *w ij +b i (1)

[0011]

[0012] In the formula, x j is the output value of the jth node in the hidden layer of the last RBM network; w ij is the weight value of the jth node in the hidden layer of the last RBM network and the ith class node of the Softmax classifier; b i is the threshold value of the ith class node of the Softmax classifier; z i is the weighted total input of the ith class node of the Softmax classifier; s i is the probability value of the ith class node calculated by the Softmax classifier. Formula (2) is also called the activation function of the Softmax classifier. SUMMARY

[0013] The technical problem solved by the present application is to provide a cold-rolled plate shape probability prediction method based on a deep belief network to solve the problems of the prior art.

[0014] To solve the above technical problems, the technical solution adopted by the present application is: a cold-rolled plate shape probability prediction method based on a deep belief network, first, according to the number of plate shape roll detection units configured by a cold rolling mill, the number of unit plate shape Softmax classifiers is determined; according to the plate shape measurement data of the historical rolled strip, the corresponding plate shape value range of each plate shape roll detection unit is counted and grid division is performed, the strip plate shape value at the corresponding grid position is calculated, and the category number and probability label value of each unit plate shape Softmax classifier are assigned; then an integrated DBN deep belief network formed by multiple stacked restricted Boltzmann machines (RBM) and fully connected with each unit plate shape Softmax classifier is established as a cold-rolled plate shape prediction network; then the historical strip rolling data and plate shape control data are used as the input of the cold-rolled plate shape prediction network, and the category probability label value is used as the network expectation, and the cold-rolled plate shape prediction network is trained offline; finally, the verified cold-rolled plate shape prediction network is used for cold-rolled plate shape prediction, and the grid plate shape intermediate value where the maximum probability value predicted by each unit plate shape Softmax classifier is located is used as the plate shape value of each unit plate shape, and the plate shape prediction value of each unit plate shape is determined according to the plate shape self-balancing principle; specifically comprising the following steps:

[0015] Step 1: determining the number of unit plate shape Softmax classifiers according to the number of plate shape roll detection units configured by the cold rolling mill;

[0016] Step 2: establishing a training data set required by the cold-rolled plate shape prediction network, each record being composed of strip rolling data, plate shape control data and plate shape measurement data;

[0017] The strip rolling data includes the initial thickness of the strip before rolling, the inlet thickness, the outlet thickness, the rolling force, the inlet tension, the outlet tension, the rolling speed and the motor torque of the last rack; the strip plate shape control data includes the roll inclination value, the work roll bending force, the work roll roll shifting amount, the intermediate roll bending force and the intermediate roll roll shifting amount; the plate shape measurement data includes the strip plate shape value measured by each plate shape roll detection unit;

[0018] Step 3: counting the strip plate shape measurement data in the training set, obtaining the corresponding plate shape value range of each plate shape roll detection unit and performing grid division, determining the category number of each unit plate shape Softmax classifier; at the same time, calculating the strip plate shape value at the corresponding grid position, and assigning the category probability label value of the unit plate shape Softmax classifier;

[0019] Step 3.1: According to the number of plate shape roller detection units, the maximum and minimum values of the plate shape measurement data in all records in the training set are respectively calculated to obtain the plate shape value change range of each plate shape roller detection unit;

[0020] Step 3.2: According to the same plate shape accuracy, the plate shape value change interval of each plate shape roller detection unit is divided to form a plate shape grid of each plate shape roller detection unit, and the number of categories of each unit plate shape Softmax classifier is determined accordingly;

[0021] Step 3.3: The position of the plate shape measurement data in each record in the training set in the divided grid of each plate shape roller detection unit is calculated, the category in each unit plate shape Softmax classifier is determined, and the category probability value is 1 and the probability values of other categories are 0;

[0022] Step 4: An integrated DBN deep belief network composed of N stacked restricted Boltzmann machines (RBM) and fully connected with M unit plate shape Softmax classifiers is established as a cold-rolled plate shape prediction network;

[0023] Step 4.1: A neural network composed of N stacked RBMs is established. In the neural network, the number of hidden layer units of the previous RBM is the same as the number of visible layer units of the next RBM, and the hidden layer output of the previous RBM is the input of the visible layer of the next RBM. The number of visible layer units of the first RBM is consistent with the number of strip rolling data items and strip shape control data items in the records of the training set;

[0024] Step 4.2: A unit plate shape prediction network is established. The hidden layer units of the last RBM of the neural network composed of N stacked RBMs are connected with the unit plate shape Softmax classifier, that is, fully connected with all category nodes under the classifier;

[0025] Step 4.3: Repeat step 4.2 to sequentially establish M unit plate shape Softmax classifiers, and fully connect them with the RBM stacked neural network established in step 4.1, and finally establish an integrated cold-rolled plate shape prediction network composed of M unit plate shape prediction networks;

[0026] Step 5: The strip rolling data and plate shape control data in each record of the training set are used as the input of the cold-rolled plate shape prediction network, and the contrastive divergence (CD) unsupervised learning algorithm is used to train the stacked RBMs in the cold-rolled plate shape prediction network layer by layer;

[0027] Step 5.1: The strip rolling data and plate shape control data in each record of the training set are normalized or standardized, and used as the input of the first RBM;

[0028] Step 5.2: Train each RBM of the stacked RBMs in turn using the contrastive divergence (CD) unsupervised learning algorithm;

[0029] Step 6: Train the unit strip shape Softmax classifier using the output value of the last RBM as input and the class probability label value corresponding to the strip shape measurement value in each record of the training set as expectation;

[0030] Step 6.1: Calculate the total input of each class node of the unit strip shape Softmax classifier through the fully connected layer according to the output value of the last RBM, and calculate the probability value of each class of the unit strip shape Softmax classifier through the Softmax activation function;

[0031] Step 6.2: Calculate the deviation between the calculated probability value of each class of the unit strip shape Softmax classifier and the class label value, and use the error back propagation algorithm to learn and train the connection weights of the fully connected layer and the threshold values of the class nodes of the unit strip shape Softmax classifier;

[0032] Step 7: Repeat step 6 to train each unit strip shape Softmax classifier in turn;

[0033] Step 8: Use the trained integrated cold rolling strip shape prediction network for actual rolling shape prediction, take the middle value of the strip shape grid where the maximum probability value predicted by each unit strip shape Softmax classifier is located as the strip shape value of each unit strip shape, and determine the strip shape prediction value of each unit strip shape according to the strip shape self-balancing principle;

[0034] Step 8.1: After standardizing the actual rolling data and strip shape control data of the strip, input them into the integrated cold rolling strip shape prediction network, and input them into each unit strip shape Softmax classifier after forward calculation of the stacked RBMs;

[0035] Step 8.2: Determine the class with the maximum probability according to the class probability value predicted by the unit strip shape Softmax classifier, i.e. determine the position of the strip shape grid of the unit strip shape and the corresponding strip shape change interval, and take the middle value of the interval as the initial value of the unit strip shape;

[0036] Step 8.3: Calculate the average value of each unit strip shape initial value output by step 8.2, and subtract the average value from each unit strip shape initial value to obtain the unit strip shape prediction value that satisfies the strip shape stress self-balancing principle.

[0037] The beneficial effects produced by the technical scheme are as follows: the cold-rolled plate shape probability prediction method based on the deep belief network can perform feature learning on the input rolling data and plate shape control data through the stacked RBM network unsupervised learning, can perform supervised learning on the plate shape values at each plate shape roll detection unit through the M unit plate shape Softmax classifier, and can thus realize prediction on special plate shape distribution. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 A flowchart of the cold-rolled plate shape probability prediction method based on the deep belief network provided for the embodiment of the present application is shown in FIG. 1.

[0039] Figure 2 A measured plate shape change range statistical scatter diagram of each plate shape detection unit of the 1340mm cold-rolled mill provided for the embodiment of the present application is shown in FIG. 2.

[0040] Figure 3 A distribution diagram of plate shape measurement data contained in the third record of the training set provided for the embodiment of the present application is shown in FIG. 3.

[0041] Figure 4 A structure schematic diagram of the cold-rolled plate shape prediction network based on the stacked RBM and Softmax classifier provided for the embodiment of the present application is shown in FIG. 4.

[0042] Figure 5 A schematic diagram of full connection of the last layer RBM and all category nodes of the jth unit plate shape Softmax classifier provided for the embodiment of the present application is shown in FIG. 5.

[0043] Figure 6 A schematic diagram of the corresponding relationship between the plate shape grid division of the plate shape roll detection unit and the categories of the unit plate shape Softmax classifier provided for the embodiment of the present application is shown in FIG. 6.

[0044] Figure 7 A comparison result diagram of the plate shape prediction value and the actual value provided for the embodiment of the present application is shown in FIG. 7. DETAILED DESCRIPTION

[0045] The specific embodiments of the present application are described in further detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application.

[0046] The embodiment takes a certain 1340mm four-stand six-roll HC cold-rolled mill as an example, and uses the cold-rolled plate shape probability prediction method based on the deep belief network to predict the cold-rolled plate shape of the cold-rolled mill.

[0047] In the embodiment, a cold-rolled strip shape probability prediction method based on a deep belief network is provided. First, the number of unit strip shape Softmax classifiers is determined according to the number of strip shape roll detection units of a cold rolling mill; the corresponding strip shape value range of each strip shape roll detection unit is calculated and grid partitioned according to the historical strip shape measurement data, and the strip shape value of the strip steel in each corresponding grid position is calculated and the number of categories and probability label values of each unit strip shape Softmax classifier are assigned; then, an integrated DBN deep belief network stacked by multiple restricted Boltzmann machine (RBM) layers and fully connected with each unit strip shape Softmax classifier is established as a cold-rolled strip shape prediction network; then, the historical strip rolling data and strip shape control data are taken as the input of the cold-rolled strip shape prediction network, and the category probability label value is taken as the network expectation, so as to perform offline training on the cold-rolled strip shape prediction network; finally, the verified cold-rolled strip shape prediction network is used for cold-rolled strip shape prediction, the grid strip shape intermediate value where the maximum probability value predicted by each unit strip shape Softmax classifier is located is taken as the strip shape value of each unit strip shape, and the strip shape prediction value of each unit strip shape is determined according to the strip shape self-balancing principle; as shown in FIG. 8, the method specifically comprises the following steps: Figure 1

[0048] Step 1: determining the number of unit strip shape Softmax classifiers M according to the number of strip shape roll detection units of a cold rolling mill;

[0049] In the embodiment, the cold rolling mill is provided with strip shape rolls after the last stand, and has 23 strip shape roll detection units. Therefore, the number of unit strip shape Softmax classifiers is determined to be 23 according to the number of strip shape roll detection units;

[0050] Step 2: establishing a training data set required by the cold-rolled strip shape prediction network, each record being composed of strip rolling data, strip shape control data and strip shape measurement data;

[0051] The strip rolling data includes the initial thickness of the strip steel before rolling, the inlet thickness, the outlet thickness, the rolling force, the inlet tension, the outlet tension, the rolling speed and the motor torque of the last stand; the strip shape control data includes the roll inclination value, the work roll bending force, the work roll roll shifting amount, the intermediate roll bending force and the intermediate roll roll shifting amount; and the strip shape measurement data includes the strip shape values measured by each strip shape roll detection unit;

[0052] In the embodiment, the strip rolling data, the strip shape control data and the strip shape measurement data during continuous rolling are exported from the iba data storage device of the cold rolling mill production site, and a total of 20000 sampling time data, i.e. 20000 records in the training data set, are included in the training data set;

[0053] ​Step 3: The strip shape measurement data in the training set is counted to obtain the corresponding strip shape value range of each strip shape roll detection unit and to perform grid division, so as to determine the number of categories of the unit strip shape Softmax classifier; meanwhile, the strip shape value of the strip steel is calculated at the corresponding grid position, and the category probability label value of the unit strip shape Softmax classifier is given;

[0054] Step 3.1: The maximum value and the minimum value of the strip shape measurement data in all records in the training set are counted respectively according to the number of strip shape roll detection units, so as to obtain the strip shape value change range of each strip shape roll detection unit;

[0055] Step 3.2: The strip shape value change interval of each strip shape roll detection unit is divided according to the same strip shape accuracy, so as to form the strip shape grid of each strip shape roll detection unit, and the number of categories of each unit strip shape Softmax classifier is determined respectively;

[0056] In this embodiment, the number of categories of the unit strip shape Softmax classifier is determined according to the strip shape grid division of the unit strip shape, and is calculated and determined according to the following formula:

[0057]

[0058] wherein, is the integer obtained by rounding up the maximum value of the strip shape of the i-th unit strip shape, is the integer obtained by rounding down the minimum value of the strip shape of the i-th unit strip shape, and Tolernance is the strip shape division accuracy, which can be set to 0.1-1.0 IU, and the specific value can be determined in combination with the strip shape prediction accuracy and the training time requirement.

[0059] Step 3.3: The position of the strip shape measurement data in each record in the training set in the divided grid of each strip shape roll detection unit is calculated, the category in each unit strip shape Softmax classifier is determined, and the probability value of the category is given as 1, and the probability values of other categories are given as 0;

[0060] In this embodiment, the strip shape measurement data in the training data set is counted according to the strip shape roll detection unit, the strip shape value change range is counted, as shown in FIG. 2, and the strip shape grid is divided according to the same strip shape accuracy, as shown in FIG. 3. Figure 2The grid division is performed, and the number of categories of the plate shape Softmax classifier of each unit is determined. According to statistics, the plate shape value range measured by the first plate shape roller detection unit is [-46.7052 IU, 0], 1.0 IU is taken as the grid division precision, the plate shape value distribution range [-47 IU, 0] can be divided into 47 grid intervals, and the number of categories of the first unit plate shape Softmax classifier is equal to 47; the plate shape value range measured by the tenth plate shape roller detection unit is [-6.88536 IU, 2.39936 IU], 1.0 IU is still taken as the grid division precision, the plate shape value distribution range [-7, 3 IU] can be divided into 10 grid intervals, and the number of categories of the tenth unit plate shape Softmax classifier is 10, and the same is true for the other 23 unit plate shape Softmax classifiers.

[0061] According to the grid division of each unit plate shape, the plate shape measurement data in each record in the training data set is calculated for the category position and determined for the category probability label. For example, the plate shape measurement data in the third data record is -20.882, -13.0536, -8.26112, -3.46859, -2.40045, 17.4205, 10.7398, 4.05903, 1.39539, -1.26826, 2.31933, 3.37251, 4.72319, 6.07387, 8.22179, 7.33824, 4.89307, 2.44791, 2.75785, -0.718385, -2.55917, -8.18577, -14.965 IU, and the plate shape distribution is shown in FIG. 6. Figure 3The first plate shape value measured by the first plate shape roller detection unit in the third data record is-20.882 IU, and in the corresponding plate shape grid of [-47 IU, 0], the position is the 27th plate shape interval in the order from small to large. Therefore, the probability value of the 27th class of the first unit plate shape Softmax classifier in the third data record is 1.0, and the probability values of other classes are all 0. The class label vector is (0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0). Similarly, the plate shape value measured by the tenth plate shape roller detection unit in the third data record is-1.26826 IU, and in the corresponding plate shape grid of [-7, 3 IU], the position is the 6th. Therefore, the probability value of the 6th class of the tenth unit plate shape Softmax classifier in the third data record is 1.0, and the probability values of other classes are all 0. The class label vector is (0, 0, 0, 0, 0, 1, 0, 0, 0, 0).

[0062] Step 4: Establish an integrated DBN deep belief network stacked by N restricted Boltzmann machines (RBM) and fully connected with M unit plate shape Softmax classifiers as the cold-rolled plate shape prediction network, as shown in Figure 4 ;

[0063] Step 4.1: Establish a neural network stacked by N restricted Boltzmann machines (RBM). In the neural network, the number of hidden layer units of the previous RBM is the same as the number of visible layer units of the next RBM, and the output of the hidden layer of the previous RBM is the input of the visible layer of the next RBM. The number of visible layer units of the first RBM is consistent with the number of strip rolling data items and strip plate shape control data items in the training set records.

[0064] Step 4.2: Establish a unit plate shape prediction network. Connect the hidden layer units of the last RBM of the neural network stacked by N restricted Boltzmann machines (RBM) with the unit plate shape Softmax classifier, as shown in Figure 5 , that is, fully connected with all class nodes under the classifier, as shown in Figure 6 ;

[0065] Step 4.3: Repeat step 4.2 to sequentially establish M unit plate shape Softmax classifiers and fully connect them with the RBM stacked neural network established in step 4.1, and finally establish an integrated cold-rolled plate shape prediction network composed of M unit plate shape prediction networks.

[0066] In this embodiment, an integrated DBN deep belief neural network is established, which is stacked by 3 layers of RBMs and fully connected with 23 unit plate shape Softmax classifiers, wherein the number of visible layer nodes of the first layer of RBMs is 8, and the number of hidden layer nodes is 7; the number of visible layer nodes of the second layer of RBMs is 7, and the number of hidden layer nodes is 5; the number of visible layer nodes of the third layer of RBMs is 5, and the number of hidden layer nodes is 4. The number of visible layer units of the first layer of RBMs is consistent with the number of rolling data items (entry thickness of the last rack, exit thickness, rolling force, entry tension, exit tension, rolling speed) and plate shape control data items (roll inclination value, work roll bending force) in the training set data records.

[0067] Step 5: The strip rolling data and plate shape control data in each record of the training set are taken as the input of the cold-rolled plate shape prediction network, and the RBMs stacked in the cold-rolled plate shape prediction network are trained layer by layer by using the contrast divergence (CD) unsupervised learning algorithm.

[0068] Step 5.1: The strip rolling data and plate shape control data in each record of the training set are normalized or standardized, and taken as the input of the first RBM;

[0069] Step 5.2: Each RBM stacked in sequence is trained by using the contrast divergence (CD) unsupervised learning algorithm.

[0070] In this embodiment, after the rolling data and plate shape control data of each record in the training data set are standardized by Z-Score, they are input into the first RBM, and the first RBM network is trained by unsupervised learning by using the contrast divergence algorithm. After the training is completed, the output of the first RBM network is taken as the input of the next RBM, and the second RBM and the third RBM network stacked in sequence are also trained.

[0071] Step 6: The output value of the last layer of RBMs is taken as the input, and the class probability label value of the unit plate shape Softmax classifier corresponding to the strip plate shape measurement value in each record of the training set is taken as the expectation, so as to train the unit plate shape Softmax classifier;

[0072] Step 6.1: According to the output value of the last RBMs, the total input of each class node of the unit plate shape Softmax classifier is calculated through the full connection layer, and the full class probability value of the unit plate shape Softmax classifier is calculated through the Softmax activation function;

[0073] Step 6.2: The deviation between the calculation probability value of all classes of the unit plate shape Softmax classifier and the class label value is calculated, and the connection weights of the full connection layer and the threshold values of the class nodes of the unit plate shape Softmax classifier are trained by using the error back propagation algorithm;

[0074] In this embodiment, the output value of the third RBM is taken as input into the first unit plate shape Softmax classifier, and the class probability label vector of the Softmax classifier corresponding to the measurement value of the first detection unit of the plate shape roller in each training data record is taken as the expectation, and the first unit plate shape Softmax classifier is trained by using the error back propagation algorithm, the full connection weight value between the hidden layer nodes of the third RBM network and the class nodes of the first unit plate shape Softmax classifier is updated, and the threshold value of the class nodes of the first unit plate shape Softmax classifier is also updated.

[0075] Step 7: Repeat step 6 to sequentially train each unit plate shape Softmax classifier.

[0076] In this embodiment, step 6 is repeated to sequentially train the second to twenty-third unit plate shape Softmax classifiers, and the corresponding weight values and threshold values are updated.

[0077] Step 8: The trained integrated cold rolling plate shape prediction network is used for actual rolling plate shape prediction, the middle value of the plate shape grid where the maximum probability value predicted by each unit plate shape Softmax classifier is located is taken as the plate shape value of each unit plate shape, and the plate shape prediction value of each unit plate shape is determined according to the plate shape self-balancing principle.

[0078] Step 8.1: After the actual rolling data and the plate shape control data of the strip steel are standardized, they are input into the integrated cold rolling plate shape prediction network, and after the forward calculation of the stacked RBMs, they are input into each unit plate shape Softmax classifier.

[0079] Step 8.2: According to the class probability value predicted by the unit plate shape Softmax classifier, the class with the maximum probability is determined, i.e. the position of the plate shape grid of the unit plate shape and the corresponding plate shape change interval are determined, and the middle value of the interval is taken as the initial value of the unit plate shape.

[0080] Step 8.3: The average value of each unit plate shape initial value output by step 8.2 is calculated, and each unit plate shape initial value is subtracted from the average value, so as to obtain the unit plate shape prediction value satisfying the plate shape stress self-balancing principle.

[0081] In the embodiment, the strip rolling data and the shape control data to be predicted are standardized by Z-Score, and then input into the three stacked RBMs, sequentially calculated in the forward direction, and then input into the 23 unit shape Softmax classifiers. In the total class probability output by each unit shape Softmax classifier, the class where the maximum probability value is located is determined, the corresponding shape grid interval is found, and the shape intermediate value of the interval is taken as the initial value of the shape predicted by the unit shape Softmax classifier. Finally, the initial values of the 23 unit shapes are subtracted by the average value, and the shape prediction value satisfying the shape stress self-balancing principle is obtained. The distribution comparison of the shape prediction value and the actual value obtained by the embodiment is shown in the following table. Figure 7 .

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features. Such modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present application.

Claims

1. A deep belief network-based cold-rolled strip shape probability prediction method, characterized in that: the number of unit shape Softmax classifiers is determined according to the number of shape roll detection units configured for a cold rolling mill train; the range of shape values corresponding to each shape roll detection unit is calculated and grid partitioned according to historical strip shape measurement data, the strip shape value in each corresponding grid position is calculated, and the number of categories and probability label values of each unit shape Softmax classifier are assigned; an integrated DBN deep belief network stacked by multiple restricted Boltzmann machines (RBMs) and fully connected with each unit shape Softmax classifier is established as a cold-rolled strip shape prediction network; the historical strip rolling data and shape control data are used as the input of the cold-rolled strip shape prediction network, and the category probability label values are used as the network expectation for offline training of the cold-rolled strip shape prediction network; the verified cold-rolled strip shape prediction network is used for cold-rolled strip shape prediction, the middle value of the grid where the maximum category probability value in each unit shape Softmax classifier is located is taken as the shape value of each unit shape, and the shape prediction value of each unit shape is determined according to the shape self-balancing principle; the method comprises the following steps: Step 1: determining the number of unit shape Softmax classifiers M according to the number of shape roll detection units configured for a cold rolling mill train; Step 2: establishing a training data set required by the cold-rolled strip shape prediction network, each record being composed of strip rolling data, shape control data and shape measurement data; the strip rolling data including the initial thickness of the strip before rolling, the inlet thickness, the outlet thickness, the rolling force, the inlet tension, the outlet tension, the rolling speed and the motor torque of the last stand; the strip shape control data including the roll inclination value, the work roll bending force, the work roll roll shifting amount, the intermediate roll bending force and the intermediate roll roll shifting amount; and the strip shape measurement data including the strip shape values measured by each shape roll detection unit; Step 3: calculating the range of shape values corresponding to each shape roll detection unit and performing grid partitioning based on the strip shape measurement data in the training set, determining the number of categories of each unit shape Softmax classifier; and calculating the strip shape value in each corresponding grid position and assigning the category probability label value of the unit shape Softmax classifier; Step 4: establishing an integrated DBN deep belief network stacked by N restricted Boltzmann machines (RBMs) and fully connected with M unit shape Softmax classifiers as a cold-rolled strip shape prediction network; Step 5: using the strip rolling data and shape control data in each record of the training set as the input of the cold-rolled strip shape prediction network, and training the RBMs stacked in the cold-rolled strip shape prediction network layer by layer using the contrastive divergence (CD) unsupervised learning algorithm; Step 6: using the output value of the last layer of RBMs as the input, and using the category probability label value of the unit shape Softmax classifier corresponding to the strip shape measurement value in each record of the training set as the expectation to train the unit shape Softmax classifier; Step 7: repeating Step 6 to train each unit shape Softmax classifier in turn. Step 8: The trained integrated cold-rolled plate shape prediction network is used for actual rolling plate shape prediction, the middle value of the plate shape grid where the maximum probability value predicted by each unit plate shape Softmax classifier is located is taken as the plate shape value of each unit plate shape, and the plate shape prediction value of each unit plate shape is determined according to the plate shape self-balancing principle.

2. The cold strip shape probabilistic prediction method based on deep belief network according to claim 1, characterized in that: The specific method of step 3 is: Step 3.1: According to the number of plate shape roller detection units, the maximum and minimum values of the plate shape measurement data in all records in the training set are counted to obtain the plate shape value change range of each plate shape roller detection unit; Step 3.2: According to the same plate shape accuracy, the plate shape value change interval of each plate shape roller detection unit is divided to form a plate shape grid of each plate shape roller detection unit, and the number of categories of each unit plate shape Softmax classifier is determined accordingly; Step 3.3: The position of the plate shape measurement data in each record in the training set in the divided grid of each plate shape roller detection unit is calculated to determine the category in each unit plate shape Softmax classifier, and the category probability value is given as 1 and the probability values of other categories are given as 0.

3. The cold strip shape probabilistic prediction method based on deep belief network according to claim 1, characterized in that: The specific method of step 4 is: Step 4.1: A neural network stacked by N restricted Boltzmann machines (RBM) is established; in the neural network, the number of hidden layer units of the previous RBM is the same as the number of visible layer units of the next RBM, and the hidden layer output of the previous RBM is the input of the visible layer of the next RBM; The number of visible layer units of the first RBM is consistent with the number of strip rolling data items and strip plate shape control data items in the training set records; Step 4.2: A unit plate shape prediction network is established; the hidden layer units of the last RBM of the neural network stacked by N RBMs are connected with the unit plate shape Softmax classifier, that is, all category nodes under the classifier are fully connected; Step 4.3: Repeat step 4.2 to sequentially establish M unit plate shape Softmax classifiers, and fully connect them with the RBM-stacked neural network established in step 4.1, and finally establish an integrated cold-rolled plate shape prediction network composed of M unit plate shape prediction networks.

4. The cold strip shape probability prediction method based on deep belief network according to claim 1, characterized in that: The specific method of step 5 is: Step 5.1: The strip rolling data and plate shape control data in each record of the training set are normalized or standardized, and are taken as the input of the first RBM; Step 5.2: Each RBM in the stacked network is trained in turn using the contrastive divergence (CD) unsupervised learning algorithm.

5. The cold strip shape probabilistic prediction method based on deep belief network according to claim 1, characterized in that: The specific method of step 6 is: Step 6.1: According to the output value of the last RBM, the total input of each category node of the unit plate shape Softmax classifier is calculated through the fully connected layer, and the probability values of all categories of the unit plate shape Softmax classifier are calculated through the Softmax activation function; Step 6.2: The deviation between the calculated probability values of all categories of the unit plate shape Softmax classifier and the category label values is calculated, and the connection weights of the fully connected layer and the threshold values of the category nodes of the unit plate shape Softmax classifier are learned and trained using the error back propagation algorithm.

6. The cold strip shape probabilistic prediction method based on deep belief network according to claim 1, characterized in that: The specific method of step 8 is: Step 8.1: After the actual rolling data and the shape control data of the strip steel are standardized, the data are input into the integrated cold-rolled shape prediction network, and after the forward calculation of the stacked RBMs, the data are respectively input into each unit shape Softmax classifier; Step 8.2: According to the category probability value predicted by the unit shape Softmax classifier, the category with the maximum probability is determined, that is, the position of the shape grid of the unit shape and the corresponding shape change interval are determined, and the middle value of the interval is taken as the initial value of the unit shape; Step 8.3: The average value of each unit shape initial value output in step 8.2 is calculated, and each unit shape initial value is subtracted from the average value, so that each unit shape prediction value satisfying the self-balancing principle of the shape stress is obtained.

Images