CN109147878B - Soft measurement method for free calcium of cement clinker - Google Patents

Abstract

The invention discloses a soft measurement method for free calcium of cement clinker, which comprises the following steps: selecting 10 variables as auxiliary variables of clinker fCaO soft measurement according to a cement process, using a time sequence of each variable as model input, and performing normalization processing on each selected variable time sequence; according to the characteristics of a time sequence in the cement burning process, a clinker fCaO soft measurement model based on a multivariable time sequence convolution neural network is established; determining initial parameters of an MT-CNN model, and carrying out forward training on a network; and carrying out supervised training by utilizing reverse error fine tuning, optimizing the weight w and the bias b in the MT-CNN by correcting the error, and predicting the fCaO of the cement clinker in real time by utilizing the trained MT-CNN model. The method avoids calculating the time delay between each variable and clinker fCaO, and reduces the computation amount required by time sequence matching; the invention improves the convergence speed, precision and generalization capability of the model; the method can well predict the fCaO content of the cement clinker, improve the quality of the cement clinker and reduce the production energy consumption.

Description

Soft measurement method for free calcium of cement clinker

Technical Field

The invention relates to the field of monitoring of free calcium of cement clinker, in particular to a soft measurement method for the free calcium of the cement clinker.

Background

The content of free calcium (fCaO) of cement clinker is an important index for measuring the quality of the clinker in the novel dry-process cement production. The fCaO content in the clinker not only affects the stability of the cement and the strength of the clinker, but also is directly related to the energy consumption for cement burning. At present, the fCaO content of cement clinker is difficult to monitor on line, the fCaO content is mainly sampled once per hour manually and is tested and measured by laboratory test, the guidance of the cement burning process by an off-line measurement result has obvious hysteresis, and the real-time control and optimization of the cement burning process are difficult to realize. The cement clinker sintering process has the characteristics of large inertia, large time lag, multiple coupling and the like, so that an accurate cement clinker fCaO prediction model is difficult to establish. In response to the above problems, some scholars have used different soft-metric modeling methods to study the clinker fcoa prediction model. And selecting five related variables burnt by the cement clinker in Zhao and Peng courses and the like to establish a multi-core LSSVM cement clinker fCaO prediction model. The method does not consider the time delay between each variable and the cement clinker fCaO, although the convergence rate of the LSSVM prediction model is high, the method is more suitable for small-scale data samples, and the change rule among the variables in large data is difficult to find. Weitao Li et al have implemented clinker fCaO soft measurement methods using improved neural networks, using data from compressed feature vectors. While the method in the literature considers the variable information within one hour, only one characteristic value is extracted from the data within one hour of each variable, so that the characteristic information influencing fCaO is greatly reduced. Due to the complexity and the variability of the cement sintering process, if a more effective soft measurement modeling method for the cement clinker fCaO is to be obtained, the time delay between the multivariable and the clinker fCaO must be considered, and meanwhile, the characteristic information of variable data can be fully extracted.

Disclosure of Invention

Aiming at the existing problems, the invention provides a cement clinker free calcium soft measurement modeling method based on a multivariable time series convolution neural network (MT-CNN for short), thereby eliminating the influence of variable time delay on clinker fCaO soft measurement.

In order to solve the technical problems, the invention is realized by the following technical scheme:

a soft measurement method for free calcium of cement clinker is characterized by comprising the following steps:

step S1: selecting 10 variables as auxiliary variables of clinker fCaO soft measurement according to a cement process, using a time sequence of each variable as model input, and performing normalization processing on each selected variable time sequence;

step S2: according to the characteristics of a time sequence in the cement burning process, a clinker fCaO soft measurement model based on a multivariable time sequence convolution neural network is established;

step S3: and determining initial parameters of the MT-CNN model and carrying out forward training on the network. The initial parameters comprise the number of convolution layers and the number of pooling layers of the MT-CNN, the learning rate, the weight w and the bias b of each hidden layer, all-connected layer and output layer, and the number and the size of convolution kernels and pooling kernels;

step S4: carrying out supervised training by utilizing error reverse fine tuning, and optimizing a weight w and a bias b in the MT-CNN by correcting the error;

step S5: and (5) predicting the fCaO of the cement clinker in real time by using the trained MT-CNN model.

In the above technical solution, in step S2, a two-dimensional array composed of time series of 10 variables is used as a model input, where one column represents sampling data of a single variable in a certain time period, a one-dimensional convolution pooling manner is used to extract features of each column, and then all feature information is integrated by a full connection layer to construct a clinker fcoa soft measurement model based on MT-CNN.

In the above technical solution, in step S4, the supervised reverse fine tuning is to optimize the weight w and the offset b layer by using a BP reverse error correction algorithm, where the reverse training in the MT-CNN is supervised training.

In the above technical scheme, in step 1, 10 variables of the cement process are as follows: the coal feeding amount of the decomposing furnace, the rotating speed of a high-temperature fan, the outlet temperature of the decomposing furnace, the outlet temperature of a grate cooler, the rotating speed of an EP fan, the negative pressure of the kiln tail, the negative pressure of the kiln head, the secondary air temperature, the kiln current and the coal feeding amount of the kiln head.

Compared with the prior art, the invention has the following beneficial effects:

1. the MT-CNN cement clinker fCaO soft measurement method established by the invention solves the problem of uncertainty of time delay between the selected 10 variables and the clinker fCaO, avoids calculating the time delay between each variable and the clinker fCaO, and reduces the computation amount required by time sequence matching.

2. According to the invention, the MT-CNN model is constructed according to the characteristics of cement multivariable time sequences, and compared with other traditional deep neural network algorithms, the convergence speed, the precision and the generalization capability of the model are improved.

3. The method can well predict the fCaO content of the cement clinker, and plays an important guiding role in improving the quality of the cement clinker and reducing the energy consumption of production.

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 described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.

FIG. 1 is a diagram of a soft measurement scheme for clinker fCaO in accordance with the present invention;

FIG. 2 is a diagram of a soft measurement model of clinker fCaO designed by the invention;

FIG. 3 is a diagram illustrating the MT-CNN convolution process;

FIG. 4 is a diagram of MT-CNN convolution pooling process parameters.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention.

The invention provides a multivariate time series convolution neural network-based cement clinker free calcium soft measurement modeling method, and the design scheme of soft measurement is shown in figure 1. Firstly, variable selection is carried out, related variables of clinker fCaO are obtained according to cement process analysis, and a time sequence of soft measurement modeling is determined. A time sequence containing characteristic information of each variable is used as modeling data, the characteristics are extracted in a one-dimensional convolution pooling kernel mode according to the characteristics of the time sequence, then all the characteristic information is synthesized by utilizing a full connection layer, and a clinker fCaO soft measurement model based on MT-CNN is constructed and is shown in figure 2. And finally, extracting features by adopting a forward algorithm, and carrying out supervised parameter fine adjustment by adopting a back propagation algorithm to complete the construction of the MT-CNN soft measurement model. The method comprises the following steps:

step S1: the comprehensive analysis of the cement process selects 10 variables related to the clinker fCaO, and a data sequence of each variable in a certain time period is used as the input of a soft measurement model.

In the step 1, analyzing a mechanism of fCaO generation in a cement burning process and main reasons influencing the fCaO content of clinker, and selecting 10 variables closely related to the fCaO content of the clinker as auxiliary variables of soft measurement modeling of the fCaO of the clinker; in order to eliminate the influence of time delay between different variables and clinker fCaO on the prediction accuracy of the clinker fCaO, a time sequence containing data characteristics is selected as the input of a soft measurement model; in order to improve the convergence speed of model training, the raw data of each variable is subjected to normalization processing in sequence.

It is known from cement technology that cement raw materials are calcined in a rotary kiln at high temperature to form a sintering zone, a sintering reaction is carried out, solid particle materials obtained by cooling are called cement clinker, and a solidified body contains a small amount of uncombined calcium oxide called free calcium (fCaO). Too high a free calcium content will decrease the stability of the cement and too low a free calcium content will increase the energy consumption for cement firing, so that the fCaO needs to be controlled within a reasonable range. In the cement calcination process, all parameters of a burning zone play a crucial role in the content of the clinker fCaO, so that the parameters of a burning system are main factors for realizing soft measurement of the clinker fCaO. The heat source of the burning zone is the coal feeding amount of the decomposing furnace, the coal feeding amount of the kiln head and secondary air recycled into the kiln from the grate cooler, and the temperature of the burning zone influences the content of calcium oxide generated in the decomposition process of raw materials and the absorption condition of the generated calcium oxide by other compounds. The high-temperature fan and the EP fan enable huge air pressure difference to be generated in the kiln, so that the air passage of a cement burning system is guaranteed to be smooth, and the pressure in the kiln is kept stable. When the rotary kiln rotates, a kiln motor is required to provide power, the uniformity of chemical reaction of materials in the rotary kiln is ensured, and the higher the current of a kiln main machine is, the higher the viscosity of the materials in the kiln is, and the higher the temperature in the kiln is.

From the above analysis, 10 variables closely related to the clinker fcoa content were selected: the coal feeding amount of the decomposing furnace, the rotating speed of a high-temperature fan, the outlet temperature of the decomposing furnace, the outlet temperature of a grate cooler, the rotating speed of an EP fan, the negative pressure of the kiln tail, the negative pressure of the kiln head, the secondary air temperature, the kiln current and the coal feeding amount of the kiln head.

Step S2: according to the characteristics of the time sequence in the cement sintering process, a clinker fCaO soft measurement model based on a multivariate time sequence convolution neural network is established.

In step S2, a two-dimensional array composed of a time series of 10 variables over a certain period of time is input as a model, where a certain column represents the sample data of one variable. Aiming at the continuous time characteristics of cement variables in a period of time, a time sequence containing the characteristic information of each variable is used as modeling data, the characteristic of each variable is extracted in a single-dimensional convolution pooling mode, then a full-connection layer is connected to synthesize all the characteristic information, and a clinker fCaO soft measurement model based on MT-CNN is constructed.

Preliminary establishment of MT-CNN model

The MT-CNN model comprises a 10-variable time sequence input layer, 2 single-dimensional convolutional layers for extracting data features, 2 single-dimensional pooling layers, a full connection layer and an output layer. Convolution pooling process parameters as in fig. 4, two convolutional layers contain 16 and 32 convolution kernels of size [13,1] and [13,1], respectively, and the pooling kernel of the pooling layer is of size [2,1 ]. The learning rate was 0.001.

Forward training of MT-CNN models

The MT-CNN extracts features through forward training, and then reversely corrects the weight and the bias according to a gradient descent method, so that training errors are reduced. After the input layer of the model reads variables, the features are extracted through the single-dimensional convolution pooling layer, then the local information representing all the features is integrated through the full-connection layer, and finally all the features are integrated and output.

(1) MT-CNN input layer. Let X be the input to MT-CNN, which contains a time series of selected 10 variables, which can be expressed as:

x＝(x₁,x₂,...,x₁₀) (1)

each variable time series contains t sample points:

x_i＝(x_i(1),x_i(2),...,x_i(t)) (2)

in the formula, x_i( i 1, 2.., 10) is a time series of the ith variable。

(2) MT-CNN convolutional layer. And (3) extracting features of the convolution layer of the MT-CNN in a single-dimensional convolution mode according to the characteristics of the cement data. If the layer of input time sequence is convolved with n single-dimensional convolution cores in fig. 2, n different feature vectors are obtained, the weights are shared in the convolution process, and n is a positive integer.

MT-CNN convolution procedure As shown in FIG. 3, if the l-th layer is convolution layer, then the layer inputs x_i ^l-1And output x_i ^lThe expressions are respectively formula (3):

x^l＝f(x^l-1*w^l+b^l) (3)

in the formula, x^l-1Feature vectors, w, representing convolutional layer inputs^lRepresenting a one-dimensional convolution kernel, b^lAnd (4) representing the corresponding bias of the output feature vector, wherein f (-) is an activation function. The MT-CNN model adopts a ReLU function as an activation function, and the expression of the ReLU function is shown as formula (4).

f(x)＝max(0,x) (4)

(3) MT-CNN pooling layer. In order to reduce feature data and simplify network computation complexity, a pooling layer is added in the MT-CNN to realize feature compression. Since the pooling layer does not contain an activation function, the pooling layer outputs p^l+1And the layer input p^lThe relationship between them is expressed as:

in the formula (I), the compound is shown in the specification,

representing pooling level input p^lIs determined, wherein m is the size of the one-dimensional pooling kernel.

(4) MT-CNN full connection layer. The feature vectors obtained after multiple convolutions and pooling, such as the k-1 layer in fig. 2, are used as input to the fully-connected layer. Input x of the layer^k-1And output y^k-1The relationship between them is as in equation (6).

y^k-1＝f(w^k-1*x^k-1+b^k-1) (6)

In the formula w^k-1、b^k-1Respectively, the weight and bias of the fully connected layer.

(5) MT-CNN output layer. In order to avoid overfitting, a regularization method, namely a data loss (Dropout) technology is adopted before the output layer of the network model, so that the purpose of improving the generalization capability of the network model is achieved. As shown in the kth layer of FIG. 2, the MT-CNN output layer directly calculates the cement clinker fCaO value using a linear weighted summation. Then the layer enters x^kThe calculation formula between the output clinker fCaO value y' is as follows:

y'＝w^kx^k+b^k (7)

in the formula w^kAnd b^kAs are the weights and offsets of the output layer.

Step S3: and determining initial parameters of the MT-CNN model and carrying out forward training on the network. The initial parameters comprise the convolution layer number and the pooling layer number of the MT-CNN, the learning rate, the weight w and the bias b of each hidden layer, all-connected layer and output layer, and the number and the size of convolution kernels and pooling kernels.

Step S4: and carrying out supervised training by utilizing error reverse fine tuning, and optimizing the weight w and the bias b in the MT-CNN by correcting the error.

And the supervised reverse fine adjustment is to optimize the weight w and the offset b layer by adopting a BP reverse error correction algorithm, wherein the reverse training in the MT-CNN is supervised training.

And the MT-CNN realizes parameter updating by adopting a reverse fine adjustment method. In this model, the mean square error MSE (mean Squared error) is chosen as the loss function J, where MSE is shown in equation (8):

wherein y is_iRepresenting the true value, y, of clinker fCaO_i' denotes the predicted output of clinker fCaO and n denotes the number of training samples.

In the MT-CNN, error sensitivities of the output layer, the fully-connected layer, and the convolutional-layer pooling layer connected to the fully-connected layer can all be obtained by a BP algorithm, and a reverse relay algorithm when the convolutional layer is connected to the pooling layer will be described below.

(1) Back propagation of convolutional layers to pooling layers. When the first layer is a pooling layer followed by a convolution layer (l + 1), the error sensitivity of the pooling layer is delta^lCalculate the equation (9) where the input a of the pooling layer output neuron^lAnd an output p^lAre equal.

Wherein

z^l+1＝p^l*w^l+1+b^l+1 (10)

In the formula, delta^l+1Indicating the error sensitivity of the l +1 layer, z^l+1The convolutional layer outputs the input of the neuron, rot180(·) denotes rotating · by 180 degrees.

(2) Reverse propagation of the pooling layer to the convolutional layer. When the convolutional layer l-1 is followed by the pooling layer l, the convolutional layer error sensitivity δ can be known from the chain law^l-1The equation (11) is calculated.

Wherein:

in the formula, z^l-1Outputs inputs to neurons for convolutional layers; x is the number of^l-1The output of the convolution layer is the input of the pooling layer; a is^lRepresenting a pooling operation result; up (-) completes the amplification and the redistribution of the pooled error array, and the size of the array is consistent with that after convolution; f' (. cndot.) is the activation function derivative.

Error sensitivity δ of the convolutional layer obtained from equation (11)^l-1After, convolution kernelGradient Δ w^l-1And bias gradient Δ b^l-1The calculation process is as shown in formula (13) and formula (14).

In the formula, x^l-2Is the convolution process with a convolution kernel w^l-1The characteristic value of the convolution operation is executed, and rot180 (-) indicates that · is rotated by 180 degrees.

Because the pooling layer does not contain weights and biases, the pooling layer does not involve updating of weights and biases. Updating convolutional layer convolution kernel w and bias b, which is calculated as follows:

w^l-1＝w^l-1-η*△w^l-1 (15)

b^l-1＝b^l-1-η*△b^l-1 (16)

in the formula, η is the learning rate of the network.

Step S5: and (5) predicting the fCaO of the cement clinker in real time by using the trained MT-CNN model.

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.

Claims (2)

1. A soft measurement method for free calcium of cement clinker is characterized by comprising the following steps:

step S1: selecting 10 variables as auxiliary variables of clinker fCaO soft measurement according to a cement process, using a time sequence of each variable as model input, and performing normalization processing on each selected variable time sequence; the variables are the coal feeding amount of the decomposing furnace, the rotating speed of a high-temperature fan, the outlet temperature of the decomposing furnace, the outlet temperature of a grate cooler, the rotating speed of an EP fan, the negative pressure of a kiln tail, the negative pressure of a kiln head, the temperature of secondary air, the kiln current and the coal feeding amount of the kiln head;

step S2: according to the characteristics of a time sequence in the cement burning process, a two-dimensional array formed by the time sequence of 10 variables is used as model input, one column represents sampling data of a single variable in a certain time period, the time sequence containing characteristic information of each variable is used as modeling data, the characteristic of each variable is extracted in a single-dimensional convolution pooling mode, then a full-connection layer is connected to synthesize all the characteristic information, and a clinker fCaO soft measurement model based on a multivariable time sequence convolution neural network is established; specifically comprises

Step 2.1 preliminary establishment of MT-CNN model

The MT-CNN model comprises 10 variable time sequence input layers, 2 single-dimensional convolutional layers for extracting data characteristics, 2 single-dimensional pooling layers, a full connection layer and an output layer;

step 2.2 Forward training of MT-CNN model

The MT-CNN extracts features through forward training, and then reversely corrects weight and bias according to a gradient descent method to reduce training errors; after the input layer of the model reads variables, extracting features through a single-dimensional convolution pooling layer, then integrating local information representing all the features through a full-connection layer, and finally integrating all the features and outputting the integrated information;

(1) an MT-CNN input layer; let X be the input to MT-CNN, which contains a time series of selected 10 variables, which can be expressed as:

x＝(x₁,x₂,...,x₁₀) (1)

each variable time series contains t sample points:

x_i＝(x_i(1),x_i(2),...,x_i(t)) (2)

in the formula, x_i(i 1, 2.., 10) is a time series of the ith variable;

(2) MT-CNN convolutional layer; extracting features of the convolution layer of the MT-CNN in a single-dimensional convolution mode according to the characteristics of the cement data; carrying out convolution calculation on the input time sequence of the layer by using n single-dimensional convolution cores to obtain n different feature vectors, wherein the weights are shared in the convolution process, and n is a positive integer;

if the first layer is a convolutional layer, then the layer inputs x^l-1And output x^lThe expressions are respectively formula (3):

x^l＝f(x^l-1*w^l+b^l) (3)

in the formula, x^l-1Feature vectors, w, representing convolutional layer inputs^lRepresenting a one-dimensional convolution kernel, b^lThe corresponding offset of the output characteristic vector is represented, wherein f (-) is an activation function, the MT-CNN model adopts a ReLU function as the activation function, and the expression is shown as formula (4):

f(x)＝max(0,x) (4)

(3) an MT-CNN pooling layer; adding a pooling layer in the MT-CNN to realize feature compression; output p of the pooling layer^l+1And the layer input p^lThe relationship between them is expressed as:

in the formula (I), the compound is shown in the specification,

representing pooling level input p^lSumming the m eigenvalues, where m is the size of the one-dimensional pooling kernel;

(4) an MT-CNN full connection layer; the feature vector obtained after the k-1 layer is subjected to convolution and pooling for multiple times is used as the input of the full-connected layer, and the input x of the layer^k-1And output y^k-1The relationship between them is as in formula (6):

y^k-1＝f(w^k-1*x^k-1+b^k-1) (6)

in the formula w^k-1、b^k-1Respectively the weight and the offset of the full connection layer;

(5) an MT-CNN output layer;

step S3: determining initial parameters of an MT-CNN model, and carrying out forward training on the network, wherein the initial parameters comprise the number of convolution layers and the number of pooling layers of the MT-CNN, a learning rate, weights w and offsets b of all hidden layers, all connection layers and output layers, and the number and the size of convolution kernels and pooling kernels;

step S4: carrying out supervised training by utilizing error reverse fine tuning, and optimizing a weight w and a bias b in the MT-CNN by correcting the error;

step S5: and (5) predicting the fCaO of the cement clinker in real time by using the trained MT-CNN model.

2. The method as claimed in claim 1, wherein in step S4, the supervised reverse fine tuning is to optimize the weights w and the offset b layer by using a BP reverse error correction algorithm, wherein the reverse training in MT-CNN is supervised training.

Images