阅读说明：本技术 基于卷积简单循环单元网络的水泥成品比表面积预测方法 (Cement product specific surface area prediction method based on convolution simple cycle unit network ) 是由 孙超 张园 赵海超 郭浩然 刘鹏飞 郝晓辰 于 2021-07-14 设计创作，主要内容包括：本发明公开了基于卷积简单循环单元网络的水泥成品比表面积预测方法,属于预测领域,包括以下流程：分析整个水泥磨生产工艺过程,选取与水泥成品比表面积相关的8个输入变量,然后将排列后的数据进行归一化,将归一化后的数据经过卷积网络,将经过卷积网络处理后的训练数据输入到简单循环单元网络模型中进行训练,采用基于时间的反向传播算法,然后根据相应误差项,计算权重的梯度和使用自适应矩阵估计法更新权重参数和偏置参数,直到满足要求或者达到迭代次数为止,完成模型将卷积网络处理后的水泥研磨过程中的过程变量数据给到训练好的简单循环单元网络模型中,实现水泥成品比表面积的在线预测。(The invention discloses a method for predicting the specific surface area of a cement product based on a convolution simple cycle unit network, which belongs to the field of prediction and comprises the following processes: analyzing the whole cement grinding mill production process, selecting 8 input variables related to the specific surface area of a cement finished product, then normalizing the arranged data, passing the normalized data through a convolution network, inputting training data processed by the convolution network into a simple cycle unit network model for training, adopting a time-based back propagation algorithm, then calculating the gradient of weight according to corresponding error items, updating weight parameters and bias parameters by using an adaptive matrix estimation method until the requirements are met or the iteration times are reached, and sending the process variable data processed by the convolution network in the cement grinding process to the trained simple cycle unit network model by using a completion model to realize the on-line prediction of the specific surface area of the cement finished product.)

1. The cement finished product specific surface area prediction method based on the convolution simple cycle unit network is characterized by comprising the following steps of: the method comprises the following steps:

step S1: analyzing the whole cement mill production process, selecting 8 input variables related to the specific surface area of a cement finished product, wherein the input variables are respectively that firstly, selected time variable data are sorted according to a time sequence, and then, the arranged data are normalized;

step S2: the normalized data sequentially passes through an input layer, a convolution layer, a pooling layer, a full-link layer and an output layer of a convolution network to obtain data;

step S3: inputting the data processed by the convolution network into a simple cycle unit network model for training, firstly, simply linearizing the input data, calculating through a forgetting gate, a resetting gate and a state gate, and finally, calculating the unit state to obtain a final state, thereby completing the forward propagation of the simple cycle unit network;

step S4: adopting a time-based back propagation algorithm, reversely carrying out error items of each neuron in the simple cyclic unit network neural network, wherein the error items can carry out back propagation in a time-delay manner, calculating the error items of the previous time from the current time, the error items are propagated to the upper layer, the error items are transmitted to an output layer from input data and weights, finally, a predicted value is solved and forms a loss function with a target value, in the back propagation process, the partial derivative of each node is solved from the loss function, then, the gradient of the weights is calculated according to the corresponding error items, the weight parameters and the bias parameters are updated by using an adaptive matrix estimation method, the processes are repeated until the requirements are met or the iteration times are reached, if the error items are solved and are smaller than a threshold value, the update of the weight matrix and the bias items is carried out until the set error threshold value is reached or the iteration times are maximized, completing model training and exiting the cycle;

and 5: and (3) the process variable data in the cement grinding process after the convolution network processing is sent to a trained simple cycle unit network model, so that the on-line prediction of the specific surface area of the cement finished product is realized.

2. The cement product specific surface area prediction method based on the convolution simple cycle unit network as claimed in claim 1, characterized in that: and 8 input variables in the step S1 comprise feeding feedback P1, A mill main machine current P2, mill tail dust collection fan baffle opening degree feedback P3, A mill circulating fan frequency conversion feedback P4, A mill bucket lifting current feedback P5, powder concentrator current feedback P6, powder concentrator rotating speed feedback P7 and ball mill semi-closed circuit material distribution baffle feedback P8.

3. The cement product specific surface area prediction method based on the convolution simple cycle unit network as claimed in claim 1, characterized in that: in the step S2, performing data feature extraction by using convolution, taking the obtained normalized data as input data of a convolution network, and sequentially passing through an input layer, a convolution layer, a pooling layer, a full-link layer and an output layer of the convolution network to obtain an output sequence of the convolution network;

the method aims to avoid overfitting and accelerated training, and layer normalization is added into the convolutional layer, wherein a layer normalization formula is as follows:

wherein H is a convolutional layer input neuron, x_iAs a variable, μ_tFor translation parameter at time t, σ_tScaling for time tA parameter;

secondly, extracting the characteristics of the variable data related to the specific surface area of the cement by one-dimensional convolution, wherein the calculation formula of the convolution layer after adding layer normalization is as follows:

in formula, an is the matrix product symbol,is the ith characteristic of the output value of the ith layer,a weight matrix that is the ith convolution kernel of the l-th layer; g_iThe dimension is the same as H; operators represent convolution operations; x is the number of^(l-1)Is the output of layer l-1;is a bias term; the function f is an output activation function and is a linear rectifying unit (ReLU);

compressing data through a pooling layer, simplifying network computation complexity, extracting main data characteristics, and obtaining a formula of the pooling layer as follows:

in the formula (I), the compound is shown in the specification,element in ith feature map representing pooled layer l +1, D_jFor the jth pooling zone, the number of pooling zones,elements of the ith feature map of the ith layer in the range of the pooling core;

and fourthly, the full connection layer connects all the extracted data signs for output, and data distortion is prevented.

4. The cement product specific surface area prediction method based on the convolution simple cycle unit network as claimed in claim 1, characterized in that: in the step S3, a simple cyclic unit network neural network is used for prediction, each neural unit of the simple cyclic unit network is a processing unit, each processing unit includes a plurality of thresholds, and the thresholds are used for controlling information flow;

firstly, a forgetting gate is used for determining how much information needs to be transmitted continuously at the current time, the output of the forgetting gate is determined by the input of the current time, and the forgetting gate has the calculation formula as follows:

f_t＝σ(W_fx_t+b_f)

in the formula, W_fIs the weight matrix of the forgetting gate, x_tFor the input at the current moment, σ is the forgetting gate activation function Sigmoid, b_fBiasing the item for the forgetting gate;

resetting the gate to determine how much information needs to be forgotten, wherein the output of the resetting gate is determined by the input of the current moment, and the formula of the resetting gate is as follows:

r_t＝σ(W_rx_t+b_r)

in the formula, W_rIs a weight matrix of reset gates, x_tFor the input of the reset gate, σ is the reset gate activation function Sigmoid, b_rA reset gate bias term;

the current input unit state calculation formula:

c_t＝f_t⊙c_t-1+(1-f_t)⊙Wx_t

in the formula (f)_tIs forgetting to output at the present moment, c_t-1Is the output of the state gate of the cell at the last moment, W is the weight matrix linearized by the input cell, x_tInput for the current time;

the final output of the hidden layer is determined by the current time output of the reset gate, the current time output of the unit state gate and the current time input together, and the state calculation formula of the hidden layer is as follows:

h_t＝r_t⊙Tanh(c_t)+(1-r_t)⊙x_t

in the formula, h_tFor the final output of the hidden layer, c_tIs the cell state at time t, r_tFor resetting the output of the gate at the current moment, Tanh is that the hidden layer activation function is a hyperbolic tangent activation function, x_tInput for the current time;

the final output sequence is as follows:

y_t＝σ(W_yh_t+b)

in the formula, y_tFor the predicted output of the current time t, h_tFor final output of the hidden layer, W_yB is a weight matrix of the prediction output layer, and b is a bias vector of the prediction output layer.

5. The cement product specific surface area prediction method based on the convolution simple cycle unit network as claimed in claim 1, characterized in that: in said step S4;

firstly, an error term is propagated in two directions, one is propagated along the time direction, and the error term of each time is calculated from the time t; the other direction is to propagate the error spatially one layer up, defining a loss function:

in the formula (I), the compound is shown in the specification,and y_tRespectively a predicted output value and an expected output at the time t;

at time t, the hidden layer output of the simple cyclic unit network is h_tDefining the error term delta at time t_tComprises the following steps:

the state unit information c can be known from the calculation chart of the simple cycle unit network and the chain derivation method_tThe gradient of (d) is:

wherein L is a loss function, r_tIs the reset gate output;

the forgetting gate gradient is:

the reset gate gradient is:

the formula for back propagation along the time sequence is:

the error term propagates to the upper layer: assuming that the error term defining l-1 for the current layer as l is the derivative of the weighted input of the error function to l-1, then there is

In the formula, delta_t ^l-1Is an error term of layer L-1, L is an error function, net_t ^l-1Is a weighted input for level l-1;

calculating weight gradient:

the bias gradients corresponding to the weights are:

updating the weight:

in the formula, eta is the learning rate of the model, in order to increase the applicability of the formula, the formula is generalized, W represents the network node weight, and b represents the node corresponding bias term.

Technical Field

The invention relates to the technical field of cement mill cement finished product specific surface area prediction, in particular to a method for predicting the specific surface area of a cement finished product in a cement mill grinding process based on a convolution simple cycle unit network.

Background

As a traditional industry in China, the cement industry is one of process industries of basic raw materials, plays an important role in economic construction in China, the performance of cement can directly influence the quality of concrete and indirectly influence the development of the building industry, and the performance of cement is not as dense as the specific surface area of cement. The specific surface area of cement is the total surface area of the cement powder per unit mass, and the specific surface area of cement has a great relationship with the process of grinding the cement, so that the specific surface area of cement can be used as one of the indexes for evaluating the finished cement product, and the thinner the cement is, the larger the specific surface area is, and conversely, the smaller the specific surface area of cement is. Generally, the cement product has too large specific surface area, which causes too fast hydration speed of the cement, too fast heat release and concentration, which causes quality problems such as early cracking of concrete, but if the specific surface area of the cement is too small, the cement particles are too coarse, which also affects the quality of poured concrete. Therefore, the specific surface area of the cement as an important index for evaluating the quality of the cement should be kept within a reasonable range, and the realization of the on-line prediction of the specific surface area of the cement is of great significance for improving the quality of the concrete. However, in the aspect of on-line prediction of the specific surface area of a cement product, the traditional linear prediction model is difficult to accurately predict due to the fact that the cement grinding process flow has the characteristics of time lag and randomness.

At present, the specific surface area of a cement product is detected in an on-line mode and an off-line mode, the off-line detection of the granularity is sampling from hour to site, and the specific surface area is obtained by detecting the sample by an analyzer in a laboratory. However, the sampling amount of off-line detection is small, the representativeness is insufficient, the sampling time interval is 1 hour, and the interval is long, so that the final measurement result is inconsistent with the actual production result. Therefore, the guiding action of operators in actual production is delayed, so that the specific surface area and the fineness of the cement cannot be monitored in real time, the produced product is easy to fail, and the product is possibly wasted because the national standard cannot be met and is difficult to adapt to the actual production requirement. The on-line detection is to directly detect the specific surface area and the fineness of the cement in the cement production process and transmit the information such as the specific surface area and the fineness of the detected cement to a DCS (distributed control system). At present, the equipment for measuring the granularity of the cement on line is a line granularity monitor, which can detect the granularity of the cement in real time in time, quickly, continuously and truly, and provides a more advanced means for stabilizing the quality of the cement.

In order to better adapt to the complex cement grinding process flow, the data driving technology is applied to the optimization of cement production. The data driving technology adopts observation data for modeling, avoids the defect that direct modeling cannot be carried out due to a complex process, can mine the coupling rule between process parameters and control variables from cement production data, and reduces the modeling complexity. A cement specific surface area prediction model is realized by establishing a unitary linear regression equation about the cement fineness of 45 mu m and the cement specific surface area. However, the method is not suitable for data characteristics in cement process production, and has large error and more unavailable data in the actual production process. The existing cement raw material fineness soft measurement model based on mutual information and a least square support vector machine (MI-LSSVM) solves the time delay problem existing in data, is high in prediction accuracy and strong in generalization capability, but is more suitable for small sample prediction, is not suitable for the characteristics of large data of a cement process, and cannot directly predict the specific surface area of a cement product.

Disclosure of Invention

In order to solve the defects, the invention provides a cement product specific surface area prediction method based on a convolution simple cycle unit network, which not only solves the problem that a traditional linear prediction model is difficult to accurately predict, but also solves the problem that the existing method is not suitable for small sample prediction and the cement product specific surface area is directly predicted.

In order to solve the technical problems, the invention adopts the technical scheme that: the cement product specific surface area prediction method based on the convolution simple cycle unit network comprises the following steps:

step S1: analyzing the whole cement mill production process, selecting 8 input variables related to the specific surface area of a cement finished product, wherein the input variables are respectively that firstly, selected time variable data are sorted according to a time sequence, and then, the arranged data are normalized;

step S2: the normalized data sequentially passes through an input layer, a convolution layer, a pooling layer, a full-link layer and an output layer of a convolution network to obtain output data;

step S3: inputting training data processed by a convolution network into a simple cycle unit network model for training, firstly, simply linearizing the input data, calculating through a forgetting gate, a resetting gate and a state gate, and finally, calculating a final state of a unit state to finish the forward propagation of the simple cycle unit network;

step S4: adopting a time-based back propagation algorithm, reversely carrying out error items of each neuron in the simple cyclic unit network neural network, wherein the error items can carry out back propagation in a time-delay manner, calculating the error items of the previous time from the current time, the error items are propagated to the upper layer, the error items are transmitted to an output layer from input data and weights, finally, a predicted value is solved and forms a loss function with a target value, in the back propagation process, the partial derivative of each node is solved from the loss function, then, the gradient of the weights is calculated according to the corresponding error items, the weight parameters and the bias parameters are updated by using an adaptive matrix estimation method, the processes are repeated until the requirements are met or the iteration times are reached, if the error items are solved and are smaller than a threshold value, the update of the weight matrix and the bias items is carried out until the set error threshold value is reached or the iteration times are maximized, completing model training and exiting the cycle;

and 5: and (3) the process variable data in the cement grinding process after the convolution network processing is sent to a trained simple cycle unit network model, so that the on-line prediction of the specific surface area of the cement finished product is realized.

The technical scheme of the invention is further improved as follows: in step S1, the 8 input variables include feed feedback P1, mill main machine current P2, mill tail dust collection fan baffle opening feedback P3, mill circulating fan frequency conversion feedback P4, mill hopper lifting current feedback P5, mill current feedback P6, mill rotational speed feedback P7, and ball mill semi-closed circuit material distribution baffle feedback P8.

The technical scheme of the invention is further improved as follows: in the step S2, performing data feature extraction by using convolution, taking the obtained normalized data as input data of a convolution network, and sequentially passing through an input layer, a convolution layer, a pooling layer, a full-link layer and an output layer of the convolution network to obtain an output sequence of the convolution network;

the method aims to avoid overfitting and accelerated training, and layer normalization is added into the convolutional layer, wherein a layer normalization formula is as follows:

wherein H is a convolutional layer input neuron, x_iAs a variable, μ_tFor translation parameter at time t, σ_tA scaling parameter at time t;

secondly, extracting the characteristics of the variable data related to the specific surface area of the cement by one-dimensional convolution, wherein the calculation formula of the convolution layer after adding layer normalization is as follows:

in formula, an is the matrix product symbol,is the ith characteristic of the output value of the ith layer,a weight matrix that is the ith convolution kernel of the l-th layer; g_iThe dimension is the same as H; operators represent convolution operations; x is the number of^(l-1)Is the output of layer l-1;is a bias term; the function f is an output activation function and is a linear rectifying unit (ReLU);

compressing data through a pooling layer, simplifying network computation complexity, extracting main data characteristics, and obtaining a formula of the pooling layer as follows:

in the formula (I), the compound is shown in the specification,element in ith feature map representing pooled layer l +1, D_jFor the jth pooling zone, the number of pooling zones,elements of the ith feature map of the ith layer in the range of the pooling core;

and fourthly, the full connection layer connects all the extracted data signs for output, and data distortion is prevented.

The technical scheme of the invention is further improved as follows: in step S3, a simple cyclic unit network neural network is used for prediction, each neural unit of the simple cyclic unit network is a processing unit, each processing unit includes a plurality of thresholds, and the thresholds are used for controlling information flow;

firstly, a forgetting gate is used for determining how much information needs to be transmitted continuously at the current time, the output of the forgetting gate is determined by the input of the current time, and the forgetting gate has the calculation formula as follows:

f_t＝σ(W_fx_t+b_f)

in the formula, W_fIs the weight matrix of the forgetting gate, x_tFor the input at the current moment, σ is the forgetting gate activation function Sigmoid, b_fBiasing the item for the forgetting gate;

resetting the gate to determine how much information needs to be forgotten, wherein the output of the resetting gate is determined by the input of the current moment, and the formula of the resetting gate is as follows:

r_t＝σ(W_rx_t+b_r)

in the formula, W_rIs a weight matrix of reset gates, x_tFor the input of the reset gate, σ is the reset gate activation function Sigmoid, b_rA reset gate bias term;

the current input unit state calculation formula:

c_t＝f_t⊙c_t-1+(1-f_t)⊙Wx_t

in the formula (f)_tIs forgetting to output at the present moment, c_t-1Is the output of the state gate of the cell at the last moment, W is the weight matrix linearized by the input cell, x_tInput for the current time;

the final output of the hidden layer is determined by the current time output of the reset gate, the current time output of the unit state gate and the current time input together, and the state calculation formula of the hidden layer is as follows:

h_t＝r_t⊙Tanh(c_t)+(1-r_t)⊙x_t

in the formula, h_tFor the final output of the hidden layer, c_tIs the cell state at time t, r_tFor resetting the output of the gate at the current moment, Tanh is that the hidden layer activation function is a hyperbolic tangent activation function, x_tInput for the current time;

the final output sequence is as follows:

y_t＝σ(W_yh_t+b)

in the formula, y_tFor the predicted output of the current time t, h_tFor final output of the hidden layer, W_yFor predicting the weight matrix of the output layer, b isAnd measuring the offset vector of the output layer.

The technical scheme of the invention is further improved as follows: in step S4;

firstly, an error term is propagated in two directions, one is propagated along the time direction, and the error term of each time is calculated from the time t; the other direction is to propagate the error spatially one layer up, defining a loss function:

in the formula (I), the compound is shown in the specification,and y_tRespectively a predicted output value and an expected output at the time t;

at time t, the hidden layer output of the simple cyclic unit network is h_tDefining the error term delta at time t_tComprises the following steps:

the state unit information c can be known from the calculation chart of the simple cycle unit network and the chain derivation method_tThe gradient of (d) is:

wherein L is a loss function, r_tIs the reset gate output;

the forgetting gate gradient is:

the reset gate gradient is:

the formula for back propagation along the time sequence is:

the error term propagates to the upper layer: assuming that the error term defining l-1 for the current layer as l is the derivative of the weighted input of the error function to l-1, then there is

In the formula (I), the compound is shown in the specification,is an error term of L-1 layer, L is an error function,is a weighted input for level l-1;

calculating weight gradient:

the bias gradients corresponding to the weights are:

updating the weight:

in the formula, eta is the learning rate of the model, in order to increase the applicability of the formula, the formula is generalized, W represents the network node weight, and b represents the node corresponding bias term.

Due to the adoption of the technical scheme, the invention has the technical progress that: according to the cement finished product specific surface area online prediction model established by the invention, related variables are arranged according to a time sequence to serve as initial input data of the model, and a convolution-simple cycle neural network model is constructed corresponding to the product specific surface area at a corresponding moment, so that the influence of the time-varying delay characteristic of cement data on the cement finished product specific surface area prediction is eliminated.

The cement product specific surface area online prediction model established by the invention fully utilizes the time sequence characteristics between the relevant variables and the prediction indexes, has the memory function of a Recurrent Neural Network (RNN) model, removes the dependence on a hidden layer at the last moment, reduces the training time, and overcomes the problems of gradient disappearance and gradient explosion of the RNN.

The cement product specific surface area on-line prediction model established by the invention effectively combines the convolution network and the simple cycle unit network neural network, improves the accuracy of the cement product specific surface area on-line prediction, improves the training speed of the model, and is more suitable for the big data characteristics of the cement industry.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts;

FIG. 1 is a diagram of a prediction model architecture according to the present invention;

FIG. 2 is a diagram of the convolution network architecture of the present invention;

FIG. 3 is a diagram of a simple cyclic unit neural network architecture according to the present invention;

FIG. 4 is a flow chart of a predictive model of the present invention;

Detailed Description

The present invention will be described in further detail with reference to the following examples:

the invention provides a cement finished product specific surface area prediction method based on a convolution simple cycle unit network, as shown in figures 1 to 4, firstly 8 variables related to the specific surface area of a cement finished product are selected from a database of an existing cement grinding system as input variables of a model, the 8 variables are respectively sequenced according to a time sequence, the variable parameter data are processed by adopting the convolution network, an error item of each node of the simple cycle unit network neural network is solved by adopting a time-based back propagation technology, the weight value of the model is updated by adopting an adaptive matrix moment estimation algorithm, and the minimum error is obtained by repeated training to obtain the optimal model parameter.

Step S1, analyzing the whole cement mill production process, combining the work experience of the field engineer and the measurement process of the specific surface area of the cement, selecting 8 process variables related to the specific surface area of the cement as input variables of the model, and sequencing the data of the selected variables according to a time sequence, wherein the 8 input variables are respectively as follows: feeding feedback P1, A mill main machine current P2, mill tail dust collection fan baffle opening feedback P3, A mill circulating fan frequency conversion feedback P4, A mill hopper current feedback P5, powder concentrator current feedback P6, powder concentrator rotating speed feedback P7, and ball mill semi-closed material distribution baffle feedback P8. The cement grinding system data base is used for obtaining the cement grinding system data, the cement grinding system data base is used for obtaining the cement grinding system data base, the cement grinding system data base is used for obtaining the cement grinding system data, and the cement grinding system data base is used for obtaining the cement grinding system data.

Step S2: and taking the normalized data as the input of the convolutional network to effectively process the data, wherein the data needs to be processed by an input layer, a convolutional layer, a pooling layer, a full-link layer and an output layer of the convolutional network in sequence.

And performing data feature extraction by using convolution, taking the obtained normalized data as input data of the convolution network, and sequentially passing through an input layer, a convolution layer, a pooling layer, a full-link layer and an output layer to obtain output data of the convolution network.

To avoid overfitting and accelerate training, Layer Normalization (Layer Normalization) is added to the convolutional Layer, and the Layer Normalization formula is as follows:

wherein H is a convolutional layer input neuron, x_iAs a variable, μ_tFor translation parameter at time t, σ_tIs the scaling parameter at time t.

Secondly, extracting the characteristics of the variable data related to the specific surface area of the cement by one-dimensional convolution, wherein the calculation formula of the convolution layer after adding layer normalization is as follows:

in formula, an is the matrix product symbol,is the ith characteristic of the output value of the l-th layer;a weight matrix that is the ith convolution kernel of the l-th layer; g_iThe dimension is the same as H; operator representation of convolution operationCalculating; x is the number of^(l-1)Is the output of layer l-1;is a bias term; the function f is an output activation function which is a linear rectifier unit (ReLU).

Compressing data through a pooling layer, simplifying network computation complexity, extracting main data characteristics, and obtaining a formula of the pooling layer as follows:

in the formula (I), the compound is shown in the specification,elements in the ith feature map representing the pooled layer l + 1; d_jIs the jth pooling area;the ith feature map of the ith layer is the element in the range of the pooling kernel.

And fourthly, the full connection layer connects all the extracted data signs for output, and data distortion is prevented.

Step S3: and (4) sending the data processed by the convolutional network to a simple cyclic unit network as input data to carry out sample training. Firstly, input data is simply linearized, and is calculated through a forgetting gate, a resetting gate and a state gate, and finally, a final state is calculated from the unit states, so that the forward propagation of the simple cycle unit network is completed.

Each neural unit of the simple cyclic unit network is a processing unit, each processing unit comprises a plurality of thresholds, the thresholds are used for controlling information flow, and the flow is as follows:

firstly, a forgetting gate is used for determining how much information needs to be transmitted continuously at the current time, the output of the forgetting gate is determined by the input of the current time, and the forgetting gate has the calculation formula as follows:

f_t＝σ(W_fx_t+b_f) (5)

in the formula, W_fIs the weight matrix of the forgetting gate, x_tFor the input at the current moment, σ is the forgetting gate activation function Sigmoid, b_fTo forget the gate bias term.

Resetting the gate to determine how much information needs to be forgotten, wherein the output of the resetting gate is determined by the input of the current moment, and the formula of the resetting gate is as follows:

r_t＝σ(W_rx_t+b_r) (6)

in the formula, W_rIs a weight matrix of reset gates, x_tFor the input of the reset gate, σ is the reset gate activation function Sigmoid, b_rThe gate bias term is reset.

The current input unit state calculation formula:

c_t＝f_t⊙c_t-1+(1-f_t)⊙Wx_t (7)

in the formula (f)_tIs forgetting to output at the present moment, c_t-1Is the output of the state gate of the cell at the last moment, W is the weight matrix linearized by the input cell, x_tIs input at the current moment.

The final output of the hidden layer is determined by the current time output of the reset gate, the current time output of the unit state gate and the current time input together, and the state calculation formula of the hidden layer is as follows:

h_t＝r_t⊙Tanh(c_t)+(1-r_t)⊙x_t (8)

in the formula, h_tFor the final output of the hidden layer, c_tIs the cell state at time t, r_tFor resetting the output of the gate at the current moment, Tanh is that the hidden layer activation function is a hyperbolic tangent activation function, x_tIs input at the current moment.

The final output sequence is as follows:

y_t＝σ(W_yh_t+b) (9)

in the formula, y_tFor the predicted output of the current time t, h_tFor final output of the hidden layer, W_yB is a weight matrix of the prediction output layer, and b is a bias vector of the prediction output layer.

And (c) completing forward propagation of the simple circulation unit network neural network model from the formula (c) to the formula (c).

And 4, step 4: and (2) adopting a time-based back-propagation time (BPTT) algorithm, reversely carrying out error terms of each neuron in the simple cyclic unit network neural network, carrying out reverse propagation by delaying time on the error terms, calculating the error terms of the previous time from the current time, carrying out propagation on the error terms to the upper layer, starting from input data and weight, transmitting the error terms to an output layer, and finally solving a predicted value to form a loss function with a target value. In the back propagation process, the partial derivative of each node is obtained from the loss function, then the gradient of the weight is calculated according to the corresponding error item, the weight parameter and the bias parameter are updated by using an adaptive matrix estimation method, the process is repeated until the requirement is met or the iteration times are reached, and the error can be minimized by repeated training.

Firstly, an error term is propagated in two directions, one is propagated along the time direction, and the error term of each time is calculated from the time t; the other direction is to propagate the error spatially one layer up.

Defining a loss function:

in the formula (I), the compound is shown in the specification,and y_tRespectively, the predicted output value and the desired output at time t.

At time t, the hidden layer output of the simple cyclic unit network is h_tDefining the error term delta at time t_tComprises the following steps:

the state unit information c can be known from the calculation chart of the simple cycle unit network and the chain derivation method_tThe gradient of (d) is:

wherein L is a loss function, r_tThe gate output is reset.

The forgetting gate gradient is:

the reset gate gradient is:

the formula for back propagation along the time sequence is:

the error term propagates to the upper layer: assuming that the error term defining l-1 for the current layer as l is the derivative of the weighted input of the error function to l-1, then:

in the formula (I), the compound is shown in the specification,is an error term of L-1 layer, L is an error function,is a weighted input for the l-1 layer.

Calculating weight gradient:

the bias gradients corresponding to the weights are:

and thirdly, updating the weight, wherein eta is the learning rate of the model, the formula is generalized to increase the applicability of the formula, W represents the weight of the network node, and b represents the corresponding bias term of the node.

And finishing one-time backward propagation, updating each part through iteration, updating the weight matrix and the bias item if the error item is smaller than the threshold value, and exiting the loop until the set error threshold value is reached or the iteration frequency is maximum, thereby finishing the model training.

And 5: and (3) the process variable data in the cement grinding process after the convolution network processing is sent to a trained simple cycle unit network model, so that the on-line prediction of the specific surface area of the cement finished product is realized.

The above-mentioned embodiments are merely illustrative of the preferred embodiments of the present invention, and do not limit the scope of the present invention, and various modifications and improvements of the technical solution of the present invention by those skilled in the art should fall within the protection scope defined by the claims of the present invention without departing from the spirit of the present invention.