Method for Predicting Burning Through Point Based on Encoder-Decoder Network

Abstract

A method for predicting burning through point (BTP) based on an encoder-decoder network is provided, which belongs to a field of soft-sensing modeling in an industrial process. A BTP prediction model based on the encoder-decoder network with a temporal attention mechanism and a spatial attention mechanism is developed according to data acquired during an operation of a sintering machine, where the temporal attention mechanism is used to characterize temporal dynamics of samples, and the spatial attention mechanism is used to capture a correlation between an object variable and an advanced feature, to improve accuracy and robustness of the model. With the model, BTP in a sintering process can be predicted in real time, which has great practical significance for on-site process guidance and parameter adjustment.

Claims

1. A method for predicting burning through point (BTP) based on an encoder-decoder network, comprising: a first step determining auxiliary variables related to BTP as input features, reading and preprocessing data of a sintering process from a database; reading data of exhaust-gas temperatures in bellows from the database, and calculating BTP and burning rising point (BRP) with a polynomial fitting method; a second step segmenting the data with a sliding window method based on the input features to construct training samples, verification samples, and test samples; a third step establishing a BTP prediction model based on the encoder-decoder network, and training the model by means of the training samples; and a fourth step reading, at current time k, on-line historical data from time k - t.sub.h to time k in real time from a sensor and the database, collecting and preprocessing the auxiliary variables; reading the data of the exhaust-gas temperatures in the bellows from the time k -t.sub.h to the time k, calculating BTP and BRP with a least square method; segmenting the data with the sliding window method to obtain data segments and establish a many-to-many sequence data set from the time k - t.sub.h to the time k; and inputting the many-to-many sequence data set into the trained BTP prediction model to obtain a BTP prediction result within a next prediction time length t.sub.f from the time k.

2. The method for predicting BTP based on the encoder-decoder network according to claim 1, wherein in the first step, the auxiliary variables are selected as: a solid fuel ratio, a quicklime ratio, a limestone ratio, a dolomite-water ratio, a water content after a second mixing, a material thickness, an ignition temperature, air permeability, a negative pressure of a main fan, a pallet velocity, an exhaust-gas temperature of a large flue, and BRP, wherein all the auxiliary variables except BRP are obtained from the data of the sintering process stored in the database; the auxiliary variables are taken as the input features, and calculated BTP is taken as an output label.

3. The method for predicting BTP based on the encoder-decoder network according to claim 1, wherein, in the first step, reading the data of the exhaust-gas temperatures in the bellows from the database and calculating BTP and BRP with the polynomial fitting method comprises: regarding the exhaust-gas temperature T.sub.i and a position x.sub.i of the bellows at a vicinity of BTP as a quadratic relation which satisfies a first formula: $T_i=\text{a}{x}_i^2+b{x}_i+c\left(i=1,2,.Math.,m\right)$ substituting the positions and the exhaust-gas temperatures, (x.sub.i,T.sub.i of last three bellows into the first formula to obtain a linear equation set of the exhaust-gas temperatures and the positions of the bellows, wherein a subscript i represents an i.sup.th bellows to a last bellows; and solving the linear equation set to obtain a: $a=\frac{\frac{T_1-T_2}{x_1-x_2}-\frac{T_2-T_3}{x_2-x_3}}{x_1-x_3}$ then solving the linear equation set to obtain b: $b=\frac{T_1-T_2}{x_1-x_2}-a\left(x_1+x_2\right)$ then: $\text{c}=T_i-\text{a}{x}_i^2-b{x}_i$ obtaining BTP by means of the equations as follows: $x_{max}=-\frac{b}{2a}$ wherein, BRP refers to a position where the exhaust-gas temperature rise in a length direction of a sintering machine, and the position x.sub.k corresponding to the exhaust-gas temperature T.sub.k of 180° C. is solved based on a following formula: $T_k=\text{a}{x}_k^2+b{x}_k+c.$ .

4. The method for predicting BTP based on the encoder-decoder network according to claim 1, wherein, in the second step, sampling is performed with a sliding time window segment method, and each input segment sample is expressed as a matrix: $X\in R^{T_h\times f}$ wherein, T.sub.h represents a number of frames of an observation segment, f represents a number of features of the segment; and an output sample Y is set to correspond to each input sample X: $Y\in R^{T_f\times f}.$ .

5. The method for predicting BTP based on the encoder-decoder network according to claim 1, wherein, in the third step, establishing the BTP prediction model based on the encoder-decoder network comprises: establishing the model by means of a encoder-decoder framework, wherein an encoder is established by means of a gated recurrent unit (GRU), and feature data are input in time series to obtain an output, namely an advanced feature, of the encoder; then calculating a correlation between a hidden state vector and an advanced feature vector of a decoder by means of a temporal attention mechanism to obtain a weight coefficient between them; and calculating a correlation between an output label and the advanced feature by means of a spatial attention mechanism to establish a potential correlation between an object variable and the advanced feature.

6. The method for predicting BTP based on the encoder-decoder network according to claim 1, wherein parameters of the BTP prediction model are adjusted in real time according to real-time data of the sintering process for continuous iteration and optimization, so that the model has high robustness.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to describe embodiments of the present disclosure or technical solutions in the existing art more clearly, drawings required to be used in the embodiments will be briefly introduced below. It is apparent that the drawings in the descriptions below are only some embodiments of the present disclosure. Those of ordinary skill in the art also can obtain other drawings according to these drawings without making creative work.

[0021] FIG. 1 shows a construction and application diagram of a BTP prediction model based on an encoder-decoder network;

[0022] FIG. 2 shows a classification chart of variables in a sintering process;

[0023] FIG. 3 shows a fitting curve of exhaust-gas temperatures of bellows;

[0024] FIG. 4 shows a schematic diagram illustrating segmentation of data;

[0025] FIG. 5 shows a basic recurrent network and various prediction tasks;

[0026] FIG. 6 shows an encoder-decoder framework;

[0027] FIG. 7 shows an encoder-decoder framework with a temporal attention mechanism and a spatial attention mechanism; and

[0028] FIG. 8 shows curves illustrating comparison results between an encoder-decoder prediction model and other models.

DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present disclosure will be described below clearly and completely with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are merely some rather than all of the embodiments of the present disclosure. All other embodiments derived from the embodiments of the present disclosure by a person of ordinary skill in the art without creative efforts shall fall within the protection scope of the present disclosure.

[0030] An objective of the present disclosure is to provide a method and system for global stabilization control of a hypersonic vehicle that enable the global stabilization control of a non-minimum phase hypersonic vehicle.

[0031] To make the above-mentioned objective, features, and advantages of the present disclosure clearer and more comprehensible, the present disclosure will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0032] The following clearly and more completely describes the technical solution in the embodiments of the present disclosure in combination with the accompanying drawings of the embodiments of the present disclosure. Apparently, the described embodiments are only part of the embodiments of the present disclosure, not all embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present disclosure.

[0033] FIG. 1 shows specific steps of constructing and applying a BTP prediction model based on an encoder-decoder network with a temporal attention mechanism and a spatial attention mechanism. The specific steps include (1)-(4). [0034] (1) Auxiliary variables related to BTP are determined as input features, data of a sintering process are read and preprocessed from a database; data of exhaust-gas temperatures in bellows are read from the database, and BTP and BRP are calculated with a polynomial fitting method; [0035] (2) The data are segmented with a sliding window method based on the input features to construct training samples, verification samples, and test samples; [0036] (3) The BTP prediction model is established based on the encoder-decoder network, and the model is trained by means of the training samples; and [0037] (4) At current time k, on-line historical data from time k - t.sub.h to time k in real time are read from a sensor and the database, the auxiliary variables are collected and preprocessed; the data of the exhaust-gas temperatures in the bellows from the time k - t.sub.h to the time k are read, BTP and BRP are calculated with a least square method; the data are segmented with the sliding window method to obtain data segments and establish a many-to-many sequence data set from the time k - t.sub.h to the time k; and the many-to-many sequence data set is input into the trained BTP prediction model to obtain a BTP prediction result within a next predication time length t.sub.f from the time k. In a specific embodiment of the present disclosure, a time width from the time k - t.sub.h to the time k may be set as 45 minutes (t.sub.h = 45), and the predication time length t.sub.f may be set as 10 minutes (t.sub.f = 10).

[0038] The present disclosure will be further described below in conjunction with specific embodiments.

Analysis of Sintering Mechanism and Classification of Variables

[0039] This experiment is performed with a 360 m.sup.2 sintering machine manufactured by an Iron and Steel Company in Southern China. Sintering is a process in which sintered mixtures are changed from powder to solid blocks at a high temperature. Particularly, iron ore powder, solvents, fuel, and return ore are batched according to a certain proportion, mixed with water to obtain granules to be transported to a mixture bunker by a belt conveyor. Then, the granules are spread on a belt sintering machine by a feeder. The fuel contained in the mixture is ignited by an igniter at a certain negative pressure. Fume is pumped by an exhaust fan from top to bottom. The mixture is melted and burned from top to bottom during movement of a sintering pallet. A large amount of heat is generated by burning the fuel to melt a surface of the powder iron ore to generate a certain amount of liquid, so as to moisten unmelted ore granules, so that the sinter ore granules are bonded into blocks. Finally, burnt-through finished sinter ore is transported to a tail end of the sintering machine and discharged from it, crushed by a single roller, and transported to a granulation system. And except for hearth layers and return ore for sintering, the rest in products are sent into a blast furnace as the finished sinter ore.

Construction of Variables in Sintering Process

[0040] The sintering process is a dynamic time-varying process which has complex mechanisms, numerous influencing factors, indeterminacy, strong nonlinearity, large lagging property, and high coupling. For better understanding a relation between the sintering process and the variables, the variables in the sintering process are systematically induced as shown in FIG. 2.

[0041] The sintering process of the iron ore may be regarded as a system: once some operating parameters and material parameters have an effect on machine parameters, some index parameters and state parameters will correspond to these parameters, and their relation is expressed as follows:

[0042] (operating parameters, material parameters, machine parameters)-( state parameters, index parameters).

[0043] The material parameters include: an iron-containing material ratio, a fuel ratio and a solvent ratio.

[0044] The operating parameters include: water addition in the first mixing, water addition in the second mixing, a pallet velocity and a material thickness.

[0045] The machine parameters include: exhaust capacity and an exhaust area of a large fan, air leakage rate of the sintering machine, and the like.

[0046] The state parameters include: air permeability of a material layer, a negative pressure of a main tube, the exhaust-gas temperatures of the bellows, BTP, and the like.

[0047] The index parameters include: output of the sinter ore, chemical component, mechanical strength, reducibility, and the like.

[0048] By means of analysis of the sintering mechanism, literature research and statistical analysis of data, it may be determined that factors affecting BTP include: a solid fuel ratio, a limestone fuel ratio, the pallet velocity of the sintering machine, a temperature of the bellows at BRP of a sintered exhaust-gas, an ignition temperature, the material thickness, a water content after the second mixing, current BTP, and the like, as shown in table 1.

TABLE-US-00001 Input parameters of the model and BTP Variable Name X.sub.1 Solid fuel ratio/% X.sub.2 Quicklime ratio/% X.sub.3 Limestone ratio/% X.sub.4 Dolomite-water ratio/% X.sub.5 Water content after the second mixing/% X.sub.6 Material thickness/mm X.sub.7 Ignition temperature/°C X.sub.8 Air permeability X.sub.9 Negative pressure of a main fan /kPa X.sub.10 Pallet velocity (m/min) X.sub.11 exhaust-gas temperature of a large flue/°C X.sub.12 BRP/m Y BTP/m

Calculation of BTP

[0049] The exhaust-gas temperatures of the bellows are read from the database. A soft-sensing model is established according to a mathematical relation between the exhaust-gas temperatures of the bellows and BTP. Considering that the maximum exhaust-gas temperature is reached when the mixtures are just burned through during the sintering process, BTP can be found based on the exhaust-gas temperature of the bellows at the tail end of the sintering machine. The curve of the exhaust-gas temperatures of the bellows is shown in FIG. 3. The exhaust-gas temperature T.sub.i and a position x.sub.i of a bellows at a vicinity of BTP are approximately regarded as a quadratic relation which satisfies a formula (1):

[00009]$T_i=\text{a}{x}_i^2+b{x}_i+c\left(i=1,2,.Math.,m\right)$

[0050] The positions and the exhaust-gas temperatures, (x.sub.i, TL), of the last three bellows are substituted into the formula (1) to obtain a linear equation set of the exhaust-gas temperatures and the positions of the bellows, where a subscript i represents the i.sup.th bellows to the last bellows. The linear equation set is solved to obtain a:

[00010]$\text{a}=\frac{\frac{T_1-T_2}{x_1-x_2}-\frac{T_2-T_3}{x_2-x_3}}{x_1-x_3}$

[0051] Then the linear equation set is solved to obtain b:

[00011]$\text{b}=\frac{T_1-T_2}{x_1-x_2}-a\left(x_1+x_2\right)$

[0052] Then:

[00012]$\text{c}=T_i-\text{a}{x}_i^2-b{x}_i$

[0053] BTP is obtained by means of the equation as follows:

[00013]$x_{max}=-\frac{b}{2a}$

[0054] At a sintering site, sealing measures of the bellows at the tail end of the sintering machine are incomplete due to air leakage, causing that a measured value is less than an actual value of the exhaust-gas temperature of the bellows. In view of this, a correction coefficient, namely a feedback coefficient of a large flue, is introduced to guarantee accuracy in the calculation of BTP as follows:

[00014]$BT{P}_m=BT{P}^{\prime }-\alpha \text{Δ}\text{T}$

[0055] Where, BTP.sub.m represents a corrected value of BTP, BTP′ represents a calculated value of BTP, AT represents a temperature deviation between the measured value and actual value of the exhaust-gas temperature, a represents the correction coefficient and is generally set as 0.02.

[0056] BRP refers to a position where the exhaust-gas temperature rises in a length direction of the sintering machine, and the position x.sub.k corresponding to the exhaust-gas temperature of 180° C. (T.sub.k = 180) is solved based on the following formula:

[00015]$T_k=\text{a}{x}_k^2+b{x}_k+c$

Segmentation of Data

[0057] As shown in FIG. 4, sampling is performed with a sliding time window segment method. Adopting a sliding time window in the segment division for sampling can directly solve problems of few data files and short acquisition time, and reduce an influence of errors of individual coordinate data on determining a movement mode. In order to facilitate a data operation, each input segmented sample is expressed as a matrix:

[00016]$X\in R^{T_h\times f}$

where, T.sub.h represents the number of frames of an observation segment, f represents the number of features of the segment; and an output sample Y is set to correspond to each input sample X:

[00017]$Y\in R^{T_f\times f}$

[0058] In this way, a serial data set is established for subsequent model input.

BTP Prediction Model Based on Encoder-Decoder Network

Step 1, Off-Line Modelling

[0059] Step 1.1, 12 key variables, such as material ratio, the pallet velocity, the air permeability of the material layer, BRP, and other key variables, are determined as the input features of the model by means of the analysis of the sintering mechanism. The data are read from the database in real time, and preprocessings such as data filtering, data smoothing and data normalization are performed.

[0060] Step 1.2, the exhaust-gas temperatures in the bellows are read from the database. BTP and BRP are calculated with a polynomial fitting method. The data are segmented with the sliding window method based on the input features to construct the training samples, the verification samples, and the test samples.

[0061] Step 1.3, FIG. 5 illustrates a basic recurrent network and typical prediction tasks, since BTP is a many-to-many sequence prediction model, a encoder-decoder framework is used for modeling. Firstly, an encoder is established by means of a GRU, and feature data are input in time series to obtain an output, namely an advanced feature, of the encoder. Then, a correlation between a hidden state vector and an advanced feature vector of a decoder is calculated by means of the temporal attention mechanism to obtain a weight coefficient between them. A correlation between an output label and the advanced feature is calculated by means of the spatial attention mechanism to establish a potential correlation between an object variable and the advanced feature, as shown in FIG. 6 and FIG. 7. Finally, BTP series are continuously generated by the trained model to fulfill prediction on BTP. Formulas of the temporal attention mechanism are as follows:

[00018]$e_{(t)}^j=score\left(h_{(t-T+f)},s_{(t-1)}\right)=V_l^j\tanh \left(W_t^j\left[h_{\left(t-T+f\right)},s_{(t-1)}\right]+b_l^j\right)$

[00019]${\alpha }_{\left(t\right)}^j=\frac{\exp \left(e_{\left(t\right)}^j\right)}{{\displaystyle {.Math.}_{j=1}^T\exp \left(e_{\left(t\right)}^j\right)}}$

[00020]$x_{\left(t\right)}={\displaystyle \underset{j}{.Math.}{\alpha }_{\left(t\right)}^j{h}_{\left(j-T+1\right)}}$

[00021]$c_{\left(t\right)}=\tanh \left(x_{\left(t\right)},s_{\left(t-1\right)}\right)$

where, h.sub.(t-T+j) represents a hidden unit of the encoder, S.sub.(t-1) represents a hidden unit of the decoder at previous time,

[00022]$V_l^j,W_l^j,and{b}_l^j$

represent parameters that the model needs to learn,

[00023]${\alpha }_{(t)}^j$

represents a value of the temporal attention mechanism, and c.sub.(t) represents the advanced feature or a contextual semantic vector.

[0062] The spatial attention mechanism is expressed as follows:

[00024]$e_{\left(t\right)}^k=score\left(c_{\left(t\right)},y_{\left(t-1\right)}\right)=V_l^k\tanh \left(W_l^k\left[c_{\left(t\right)},y_{\left(t-1\right)}\right]+b_l^k\right)$

[00025]${\beta }_{\left(t\right)}^k=\frac{\exp \left(e_{\left(t\right)}^k\right)}{{\displaystyle {.Math.}_{k=1}^T\exp \left(e_{\left(t\right)}^k\right)}}$

[00026]${\tilde{c}}_{\left(t\right)}=\left({\beta }_{\left(t\right)}^1{c}_t^1,{\beta }_{\left(t\right)}^2{c}_t^2,{\beta }_{\left(t\right)}^3{c}_t^3,.Math.,{\beta }_{\left(t\right)}^n{c}_t^n\right)$

where, flk (t) represents an attention value of each dimension of a latent variable c.sub.(t), and c.sub.(t) represents an advanced variable after an action of the spatial attention mechanism.

Step 2: On-Line Detection

[0063] Step 2.1, on-line data are read in real time from the sensor and the database, and the auxiliary variables are collected and preprocessed. The data of the exhaust-gas temperatures in the bellows are read. BTP and BRP are calculated with the least square method. The data are segmented with the sliding window method to obtain data segments and establish the many-to-many sequence data set. Then, the established prediction model is deployed into a sintering expert system to perform on-line prediction according to the real-time data. Finally, process parameters are adjusted in time according to the BTP prediction result to achieve desired BTP.

Step 3: Update of Model

[0064] The parameters of the encoder-decoder model are adjusted in real time according to the real-time data during the sintering process, and the step 1 is repeated for continuous iteration and optimization so as to fulfill high robustness of the model.

Test of Model Performance

[0065] In order to test effectiveness of the model, 10000 samples are collected from a certain sintering plant with an interval of one minute and are segmented with the sliding window method. After data preprocessing, data segments are divided into 6000 training samples, 1000 verification samples, and 785 test samples. Because it is troublesome to adjust parameters of a deep learning model, the parameters of the model are set as shown in table 2 below through an experiment.

TABLE-US-00002 Hyperparameters of model Parameter Hidden size Learning rate Hidden_layer Dropout Input_size Output_size Value 20 0.001 2 0.1 40 3

[0066] In order to test the performance of the established model, two traditional machine learning models, namely a vector autoregression (VAR) model and an autoregressive integrated moving average (ARIMA) model, and two deep recurrent networks, namely a long short-term memory (LSTM) network and a GRU network, are adopted as comparison models. Evaluation indexes are a determination coefficient, a root-mean-square error, and a mean absolute error, as shown in the following formulas:

[00027]$R^2=\frac{{\displaystyle {.Math.}_{i=1}^n{\left({\tilde{y}}^{\left(i\right)}-\overline{y}\right)}^2}}{{\displaystyle {.Math.}_{i=1}^n{\left(y^{\left(i\right)}-\overline{y}\right)}^2}}$

[00028]$\text{MAE =}\frac{1}{n}{\displaystyle \underset{i=1}{\overset{n}{.Math.}}\left|{\tilde{y}}^{\left(i\right)}-y\right|}$

[00029]$\text{RMSE=}\sqrt{\frac{1}{n}{\left({\tilde{y}}^{\left(i\right)}-y\right)}^2}$

[0067] Table 3 and FIG. 8 show a prediction effect of the encoder-decoder model with the temporal attention mechanism and the spatial attention mechanism. It can be seen that the two traditional machine learning models, the VAR model and the ARIMA model, have poor performance in multi-step prediction, and accuracy of the two models does not reach 50%, which indicates that statistical learning models have certain limitations for a complex industrial process, but the performance of the two deep learning models, the LSTM network and the GRU network, has greatly improved and the accuracy has greater than 70%. However, for an industrial site, the BTP prediction model cannot be used. The encoder-decoder model achieves better performance for the many-to-many sequence prediction and the accuracy is about 90%, which improves the accuracy in prediction on BTP, provides sufficient time for sintering operators to adjust the process parameters, and improves the output and quality of the sinter ore. For sintering plants, advance prediction of BTP will bring great economic benefits to enterprises.

TABLE-US-00003 Comparison of prediction results of the models Models R.sup.2 RMSE MAE VAR 0.4547 2.4577 1.6723 ARIMA 0.4895 2.3780 1.6623 LSTM 0.7244 1.7472 1.1276 GRU 0.7557 1.6550 1.1970 Ours 0.9094 1.0019 0.7322

[0068] The embodiments are described herein in a progressive manner. Each embodiment focuses on the difference from another embodiment, and the same and similar parts between the embodiments may refer to each other. Since the system disclosed in an embodiment corresponds to the method disclosed in another embodiment, the description is relatively simple, and reference can be made to the method description.

[0069] The principle and implementation modes of the present disclosure are described by applying specific examples in the present disclosure. The descriptions of the above embodiments are only intended to help to understand the method of the present disclosure and a core idea of the method. In addition, those ordinarily skilled in the art can make changes to the specific implementation modes and the application scope according to the idea of the present disclosure. From the above, the contents of this specification shall not be deemed as limitations to the present disclosure.