CN113593657B - Cement free calcium soft measurement system with quality target as guiding semi-supervised learning - Google Patents
Cement free calcium soft measurement system with quality target as guiding semi-supervised learning Download PDFInfo
- Publication number
- CN113593657B CN113593657B CN202110845983.9A CN202110845983A CN113593657B CN 113593657 B CN113593657 B CN 113593657B CN 202110845983 A CN202110845983 A CN 202110845983A CN 113593657 B CN113593657 B CN 113593657B
- Authority
- CN
- China
- Prior art keywords
- model
- semi
- layer
- supervised
- self
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000005259 measurement Methods 0.000 title claims abstract description 45
- 239000004568 cement Substances 0.000 title claims abstract description 43
- OYPRJOBELJOOCE-UHFFFAOYSA-N Calcium Chemical compound [Ca] OYPRJOBELJOOCE-UHFFFAOYSA-N 0.000 title claims abstract description 20
- 229910052791 calcium Inorganic materials 0.000 title claims abstract description 20
- 239000011575 calcium Substances 0.000 title claims abstract description 20
- 238000000034 method Methods 0.000 claims abstract description 56
- 238000012549 training Methods 0.000 claims abstract description 29
- 238000012545 processing Methods 0.000 claims abstract description 7
- 230000008569 process Effects 0.000 claims description 37
- 238000013528 artificial neural network Methods 0.000 claims description 31
- 230000006870 function Effects 0.000 claims description 29
- 238000004519 manufacturing process Methods 0.000 claims description 15
- BRPQOXSCLDDYGP-UHFFFAOYSA-N calcium oxide Chemical compound [O-2].[Ca+2] BRPQOXSCLDDYGP-UHFFFAOYSA-N 0.000 claims description 10
- ODINCKMPIJJUCX-UHFFFAOYSA-N calcium oxide Inorganic materials [Ca]=O ODINCKMPIJJUCX-UHFFFAOYSA-N 0.000 claims description 10
- 239000000292 calcium oxide Substances 0.000 claims description 10
- 238000004422 calculation algorithm Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 6
- 238000009826 distribution Methods 0.000 claims description 5
- 239000003245 coal Substances 0.000 claims description 4
- 238000003062 neural network model Methods 0.000 claims description 4
- 238000004458 analytical method Methods 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 230000002159 abnormal effect Effects 0.000 claims description 2
- RZSCFTDHFNHMOR-UHFFFAOYSA-N n-(2,4-difluorophenyl)-2-[3-(trifluoromethyl)phenoxy]pyridine-3-carboxamide;1,1-dimethyl-3-(4-propan-2-ylphenyl)urea Chemical compound CC(C)C1=CC=C(NC(=O)N(C)C)C=C1.FC1=CC(F)=CC=C1NC(=O)C1=CC=CN=C1OC1=CC=CC(C(F)(F)F)=C1 RZSCFTDHFNHMOR-UHFFFAOYSA-N 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 claims description 2
- 230000004580 weight loss Effects 0.000 claims description 2
- 238000001514 detection method Methods 0.000 abstract 1
- 230000000875 corresponding effect Effects 0.000 description 10
- 210000002569 neuron Anatomy 0.000 description 9
- 230000002596 correlated effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000001816 cooling Methods 0.000 description 3
- 238000010168 coupling process Methods 0.000 description 3
- 238000005859 coupling reaction Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000010304 firing Methods 0.000 description 3
- 239000002994 raw material Substances 0.000 description 3
- 230000008878 coupling Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 230000007613 environmental effect Effects 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000000691 measurement method Methods 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000005245 sintering Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- UGFAIRIUMAVXCW-UHFFFAOYSA-N Carbon monoxide Chemical compound [O+]#[C-] UGFAIRIUMAVXCW-UHFFFAOYSA-N 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000001354 calcination Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000004134 energy conservation Methods 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 239000003546 flue gas Substances 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000008685 targeting Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/30—Prediction of properties of chemical compounds, compositions or mixtures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
- G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
- G16C20/70—Machine learning, data mining or chemometrics
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Chemical & Material Sciences (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Crystallography & Structural Chemistry (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Image Analysis (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a cement free calcium soft measurement system taking a quality target as a guide for semi-supervised learning, which relates to the technical field of cement clinker quality soft measurement and detection and comprises the following steps: determining auxiliary variables and performing data processing; constructing a semi-supervised model and determining model parameters at the same time; variable weighting AE hierarchical pre-training; constructing a depth network model; reversely fine-tuning the weight; and utilizing the model to predict in real time on line. The invention changes the traditional two-step modeling method into the one-step modeling method, thereby improving the prediction precision of the soft measurement model.
Description
Technical Field
The invention relates to the technical field of soft measurement and monitoring of quality of industrial cement clinker, in particular to a soft measurement system of free calcium of cement, which takes a quality target as a guide for semi-supervised learning.
Background
In modern flow industrial processes, in order to make industrial production quality indexes meet corresponding environmental protection and energy saving requirements, it is very important to timely and effectively monitor important process parameters (such as quality indexes) related to product quality. Therefore, the soft measurement technology based on data driving is straightforward, and the advantages of the soft measurement technology such as strong computing power, accurate expression of industrial processes and the like are realized, so that the soft measurement technology is a new research method for each flow industry. Of these, the most typical is the use of a soft measurement model between the process variables and free calcium oxide (Free Calcium Oxide, f-CaO) in the cement clinker firing process. In the cement process industry, the content of free calcium oxide (Free Calcium Oxide, f-CaO) of cement clinker is an important index for measuring the quality of cement, and the cement quality is reduced due to the excessively high content of f-CaO and the energy consumption is increased due to the excessively low content of f-CaO. Therefore, the on-line monitoring of the f-CaO content of the clinker is of great significance for real-time control and optimization of implementation process variables, improvement of product quality, energy conservation, environmental protection and the like. At present, in cement production enterprises, clinker f-CaO content values are usually obtained by a method of on-site manual sampling and off-line testing, but the method of on-site manual sampling and then manual testing has serious hysteresis, and real-time prediction and control in cement production process are difficult to realize. Because of the characteristics of large inertia, multiple coupling and the like of variables in the industrial process collection of the cement process, the establishment of an accurate f-CaO content prediction model is very difficult. Data-driven based soft measurement methods are increasingly being applied to the cement process industry by using readily available auxiliary variables to predict free calcium oxide (f-CaO) content. Li and the like enhance the generalization capability of the model based on a fuzzy entropy compression feature vector method, and realize the online prediction of the free calcium oxide of the cement clinker by adopting a neural network. Yuan X et al propose a self-coding method targeting quality, which performs stack self-coding by supervised learning to make the extracted features have quality features. Zhao Pengcheng and the like select five variables related to cement clinker firing to establish a polynuclear LSSVM cement clinker f-CaO prediction model. The soft measurement method has the defects that the existing soft measurement model based on deep network learning is mostly supervised learning, but in the cement flow process, the acquisition of cement free calcium label samples is time-consuming and labor-consuming, the label samples are limited, and meanwhile, the models are mostly shallow network or machine learning, so that the f-CaO content prediction problem under the complex cement working condition can not be solved. Meanwhile, compared with an unsupervised learning method, the supervised learning method can learn important features more effectively, but when the tag data sample is less, the mode of combining the supervised learning and the unsupervised learning, namely the semi-supervised learning, is a very effective mode. At present, mostly, feature extraction is performed by hierarchical pre-learning in a self-coding learning algorithm, and then fine adjustment is performed by a supervised learning method, however, only features related to input data are extracted from a self-coding network, and although the original input data can be well represented, the features possibly neglect feature information related to quality influence, so that soft measurement accuracy is influenced. Meanwhile, in the label learning, only limited label data are used when parameters are adjusted, and the model is easy to be fitted, so that the final adjusted model has poor effect. Therefore, how to organically combine the two is the key of semi-supervised network research and application. Meanwhile, deep network feature extraction is performed layer by layer, and it is extremely important to extract features related to output prediction height. Therefore, how to emphasize the expression of the input features is also a considerable work. Meanwhile, designing a soft measurement system capable of accurately predicting f-CaO is an important subject in the cement industry.
Disclosure of Invention
The invention aims to solve the technical problem of providing a cement free calcium soft measurement system which takes a quality target as a guide and semi-supervised learning, and changes the traditional two-step modeling method into a one-step modeling method, thereby improving the prediction precision of a soft measurement model.
In order to solve the technical problems, the invention adopts the following technical scheme:
the cement free calcium soft measurement system with the quality target as the guiding semi-supervised learning comprises the following steps:
step 1: determining auxiliary variables and performing data processing: through analysis of the industrial process flow, primarily selecting an easily-measured variable related to the difficult-to-measure variable as an auxiliary variable of the soft measurement model, and collecting a time sequence of the auxiliary variable and the difficult-to-measure parameter; thirdly, adopting a 3 sigma criterion to propose an abnormal value of the acquired time sequence, and carrying out normalization processing on the data before training;
step 2: constructing a semi-supervised model, and simultaneously determining model parameters: constructing a semi-supervised prediction model by using the input variables selected in the step 1, and determining initial parameters, wherein the initial parameters comprise a learning rate mu, a regularization coefficient lambda and a balance coefficient alpha; the semi-supervised self-encoder model combines a self-encoding and neural network model, wherein the self-encoding and the neural network share an input layer and a hidden layer, the output layer is divided into a self-encoding output layer and a neural network output layer, and a balance coefficient is utilized to combine a self-encoding reconstruction error and a neural network prediction error to jointly form a loss function of the model;
step 3: variable weighted AE hierarchical pre-training: the aim of the layered pre-training is to provide better initial weights for the later constructed stacked self-encoders to better highlight process variables more relevant to target output, and the correlation coefficients of the input variables and the target variables are calculated by using the marking data to give weight values of corresponding degrees;
step 4: building a depth network model: the semi-supervised model in the step 2 is utilized to remove the respective coding output layers, and encoder hidden layers are stacked to form a deep network; simultaneously minimizing the weight loss function in the step 3, and initializing the weight value of each layer;
step 5: reverse fine tuning of weights: adding an output layer to the overall fine tuning weight at the top end of the model, and realizing error readjustment and distribution of updated weights of all layers through a BP algorithm;
step 6: real-time online prediction using a model: and carrying out real-time online prediction on the variables acquired in real time by using the model constructed in the prior art.
The technical scheme of the invention is further improved as follows: in step 3, a semi-supervised model of the output layer of the integrated self-encoder and the output layer of the neural network is constructed, a period of sequence of corresponding variables is extracted, balance coefficients of the loss functions of the two output layers are considered on the loss functions, weight duty ratio of each layer is added to the balance coefficients, and the self-coding model is pre-trained to initialize parameters of the balance coefficients.
The technical scheme of the invention is further improved as follows: in step 4, a deep training model is constructed, an output layer is added at the top end, and model parameters of each layer are updated through reverse training.
The technical scheme of the invention is further improved as follows: in step 5, the model is predicted online.
The technical scheme of the invention is further improved as follows: in step 1, 10 process variables closely related to free calcium oxide are selected as auxiliary variables of a free calcium soft measurement model by analyzing the mechanism and related influencing factors of free calcium oxide production in the clinker.
The technical scheme of the invention is further improved as follows: in step 1, the 10 modeling auxiliary variables selected are: feeding quantity 1 feedback, feeding quantity of decomposing furnace coal feedback, rotating speed feedback of a high-temperature fan, kiln tail temperature, decomposing furnace outlet temperature, kiln current feedback, two-chamber grate lower pressure feedback, secondary air temperature feedback, kiln head negative pressure feedback and kiln head coal feedback.
The technical scheme of the invention is further improved as follows: in step 6, the self-encoder is combined with the neural network aiming at the problems of limited label data and nonlinearity, meanwhile, a nonlinear soft measurement model of the industrial process is finally established by using limited label data and a large amount of label-free data, the newly acquired process data is utilized for prediction, the newly acquired process data is input into the semi-supervised self-encoding model after preprocessing, and the output layer finally outputs corresponding predicted values.
The technical scheme of the invention is further improved as follows: in step 5, the following are included: after the pre-training is finished, adding an output layer to the top end of VE-SAE to fine tune the weight, and using the pre-trained weight to initialize the weight of each hidden layer; obtaining improved weights by applying a back propagation algorithm; according to the loss function calculated in the step 3, through minimizing layer-by-layer back propagation, in the back propagation process, error readjustment and distribution are realized, and weight and bias are corrected again for the distributed error, so that model error is reduced, and the purpose of learning and optimizing is achieved; the back propagation process is sequentially carried out, the weight and the paranoid parameters of the model and the variable weight value of each layer AE newly proposed by the invention are corrected each time, and when the error is reduced to a certain degree or reaches the specified iteration times, the iteration process is stopped; and obtaining derivative formulas of all weights and bias parameters of the whole model, and then continuously and iteratively updating the model parameters through a gradient descent method to obtain the iteration times of which the error meets the expected standard or reaches the stipulation.
By adopting the technical scheme, the invention has the following technical progress:
1. aiming at the problems of rare and nonlinear label samples in the cement industrial process, a semi-supervised self-coding network model based on quality target guidance is established, a semi-supervised cement clinker free calcium soft measurement model based on collaborative learning is researched, a new loss function is formed by combining a self-coding reconstruction error and a neural network, the prediction precision of the neural network soft measurement system under the conditions of rare and nonlinear f-CaO label samples is improved, and finally the effectiveness of the method is tested on a cement industrial data set.
2. Aiming at the characteristics of large inertia, large time lag, multiple coupling and the like in the process of firing cement clinker, a deep structure is adopted to replace a single hidden layer structure, the capability of learning high-order potential interpretation factors of contents is improved, and complex nonlinear characteristics of the cement industry are solved.
3. A new variable weighting idea is introduced to extract output related characteristics in a layering way, firstly important variables are identified from other variables of each self-coding input layer through correlation analysis with the output variables, and then different weights are distributed according to the correlation between the variables and the output variables, so that the problem that irrelevant information exists in the extracted advanced characteristics caused by traditional unsupervised reconstruction is solved.
4. The invention can well solve the problem that the f-CaO is difficult to implement and estimate in the cement process industry, and the established soft measurement system also has good generalization capability, can provide references for staff, and simultaneously provides corresponding effective guidance for subsequent intelligent control.
Drawings
FIG. 1 is a flow chart of the overall model of free calcium cement of the present invention;
FIG. 2 is a block diagram of a single semi-supervised self-encoding model proposed by the present invention;
FIG. 3 is a block diagram of a variable weighted stacked self-encoder model in accordance with the present invention;
FIG. 4 is a schematic structural view of the overall model;
FIG. 5 is a flow chart of the modeling of the online measurement of the present invention;
FIG. 6 is a graph of the predicted outcome of this invention;
fig. 7 is a diagram of the prediction result of the conventional model.
Detailed Description
The invention is further illustrated by the following examples:
as shown in fig. 1, 2, 3, 4, 5 and 6, a cement free calcium soft measurement system with a quality target as a guide for semi-supervised learning is provided, the overall system flow chart is shown in fig. 1, and the overall system flow chart is firstly analyzed according to the production process, and a process variable related to the f-CaO content is selected as an auxiliary variable used by a soft measurement model. According to the Laida rule, outliers are marked, and the average of the variables is used to replace outliers and missing values in the variables. Next, a semi-supervised model is built by combining a traditional self-encoder with a single-output neural network, an automatic encoder part is used for learning unlabeled sample information, and a neural network part is used for learning labeled samples. Both share an input layer and a hidden layer. The marking data is used to obtain corresponding variable weights to act on each hidden layer so as to highlight variable characteristics highly related to output variables. And simultaneously adjusting the semi-supervised model loss function to include both reconstruction errors from the encoder and prediction errors from the neural network layer, and initializing the weight parameters of the respective encoder structure in a manner that minimizes the overall loss function. The output layers of the encoder structure are removed, and the upper hidden layer output is used as the next input layer to construct a deep learning network model so as to improve the nonlinearity and multi-coupling performance of the model. And finally, finely adjusting and updating the weights of all hidden layers by adopting a back propagation mode. Finally, the newly acquired data is utilized to predict the soft measurement model on line, so that the f-CaO content is predicted on line in real time, and the overall structure system schematic diagram is shown in figure 4. The content comprises the following steps:
step 1, determining auxiliary variables and performing data processing
By analyzing the industrial process flow, preliminarily selecting an easily-measured variable related to the difficult-to-measure variable as an auxiliary variable of the soft-measurement model, and collecting a time sequence of the auxiliary variable and the difficult-to-measure variable;
the cement sintering process comprises three stages of raw material preheating, clinker calcining and clinker cooling, wherein the raw material enters a preheater from a homogenizing warehouse to be fully preheated by heat exchange with flue gas, f-CaO is generated after the raw material reaches a certain temperature and enters a decomposing furnace, the undissolved material enters a rotary kiln to be continuously decomposed, part of f-CaO is absorbed along with the temperature rise, and f-CaO which is not absorbed and exists in a free state in clinker is used as clinker fCaO. In addition, if the cooling rate is too slow during the clinker cooling process, secondary f-CaO is generated, which has a certain effect on the cement quality. By researching the cement sintering process and the generation mechanism of clinker f-CaO, the formation cause of fCaO and the main factors influencing the content of fCaO are mastered, and relevant theoretical basis is provided for the selection of soft measurement auxiliary variables of the clinker fCaO and the establishment of a model.
From the above analysis, 10 variables closely related to the clinker f-CaO content were selected: the outlet temperature of the decomposing furnace, the kiln current feedback, the feeding quantity 1 feedback, the kiln tail temperature, the secondary air temperature feedback, the feeding quantity feedback of the decomposing furnace, the rotating speed feedback of the high-temperature fan, the pressure feedback under the two-chamber grate, the rotating speed of the EP fan and the negative pressure of the kiln head.
The input variables are then preprocessed and outlier processed on the collected data using the 3σ criterion. The specific method comprises the following steps:
in the method, in the process of the invention,representing the average value of the variables; v (V) i Is the difference between the variable and the variable mean. When |V i When 3 sigma, then consider x i Coarse errors are removed. Wherein the calculation formula of sigma is as follows:
where σ is the mean square error. At the same time, the outliers in each auxiliary variable are marked and the outliers and missing values in each auxiliary variable are replaced with missing values of the modified auxiliary variable.
The maximum and minimum normalization processing is carried out on the data, the dimension of the variables is unified, the training speed is improved, the model precision is improved, and the specific conversion formula is as follows:
wherein x is min Is the minimum value of the variable, x max For the maximum value of the variable, x' is the time series normalized by the variable x.
Step 2, construction of semi-supervised self-encoder with quality target as guide
Quality target oriented semi-supervised self-encoder model as shown in fig. 2, the semi-supervised self-encoder can be seen as a combination of an automatic encoder for learning unlabeled exemplar information and a single output neural network for labeled exemplar learning. Wherein the automatic encoder shares an input layer and a hidden layer with the single output neural network. Let d= { L, U }, where D represents the dataset used for modeling, l= { X, Y } = { (X) 1 ，y 1 )，(x 2 ，y 2 )，...(x |L| ，y |L| ) The data set of labeled samples is represented by L, the number of labeled samples is represented by u= { X } = { X 1 ,x 2 ,x 3 ,…x |U| The label-free sample data set is represented, where each sample time series includes 60 sample points: x is x i ＝{x i(1) ,x i(2) ,x i(3) ...x i(60) }，x i Is the time series of the ith variable, |u| represents the number of unlabeled exemplars. Semi-supervised self-encoder model forward propagation is equivalent to the forward propagation process of the self-encoder model and neural network model.
h＝f(W 1 X+b) (4)
Wherein X is the input of a semi-supervised automatic encoder, W 1 Is the weight of the input layer to the hidden layer, b is the bias between the input layer and the hidden layer, W nn Is the weight of hidden layer to label layer, b nn Is bias of hidden layer to label, W ae Is the weight of the hidden layer to the reconstructed output layer, b ae Is the bias of the hidden layer to the reconstructed output layer, h is the output of the semi-supervised automatic encoder hidden layer,is the predictive output of the neural network part, +.>Is the input reconstruction of the automatic encoder, f (·) is the activation function, here the ReLU function is employed as the activation function, whose expression is as follows:
f(x)＝max(0,x) (7)
the loss function of a semi-supervised self encoder is defined as follows:
wherein N is L Is the number of labeled samples, N U Is the number of unlabeled exemplars, E is the overall error of the semi-supervised self encoder, E ae Is the reconstruction error of the semi-supervised self-encoder, E nn Is a nerveThe prediction error of the network, alpha is the weight coefficient,is the reconstruction of the ith unlabeled exemplar, < >>Is the predicted value of the ith javelin sample, and λ is the regularization coefficient. It can be seen from the semi-supervised loss function that when the system model is modeled again, not only limited tagged data but also a large amount of untagged data are used, and the reconstruction error of the self-encoder and the prediction error of the neural network are combined and optimized. The purpose of the final regularization is to prevent the labeled data training from causing overfitting.
Step 3, variable weight AE layered pre-training
The objective of the layered pre-training is to provide better initial weights for later built stacked self-encoders, in order to achieve the layered pre-training, the reconstruction errors of the training samples for the whole input space must be minimized, in order to get the appropriate initial weights. However, consider further the semi-supervised self encoder of FIG. 2. In the application of conventional self-encoders, its purpose is to reconstruct the input data, i.e. reconstruct the outputShould be as close as possible to the original input x, we therefore reconstruct the error +.>The term is expanded at each layer:
in the semi-supervised self encoder of the present invention, the target output y of each neural network output layer should also be considered, but it is also considered that in soft measurement applications, not all input process variables are related to the corresponding target output, and at the same time, the degree of influence of different variables on the target output is different. More considering that certain dimension elements in certain variables may not have a large correlation with the final output variable, but they have a large correlation with other dimensions in the reconstruction from the encoder, such variable information should also be propagated forward into the advanced feature layer. Whereas for those irrelevant information present in the advanced layers of the forward propagation it is mainly due to the unsupervised reconstruction of the input.
While our invention's variable weighted self-encoder (shown in fig. 3) fully considers the features highly correlated with the output, we should reconstruct more accurately for those features highly correlated with the output, and should weaken or ignore for those features less correlated with the output. Therefore, it is necessary to give different weights to different dimensions in the reconstructed object according to their correlation with the output variable. To train VW-AE, we should first obtain the corresponding variable weights from the tag data. Let us assume that we mark training data { X ] L ,Y L }＝{(x 1 ,y 1 ),(x 2 ,y 2 ),...，(x NL ，y NL ) Where NL is the number of marked samples, the importance of the input variables is determined by their relevance to the target variable, and therefore the correlation coefficient of the d-th variable needs to be calculated from the marked data.
Wherein X is l(d) Is the aggregate set of the d-th variable of the tag data, namely X l(d) ＝{x 1(d) ,x 2(d) ,...,x NL(d) }。
The covariance and correlation of the variances are calculated as follows
Wherein the method comprises the steps ofAnd->The mean of the d-th input and the target variable, respectively. If the absolute value of the correlation coefficient is large, it means that the variable is more correlated with the target variable. It should be given a greater weight value and vice versa. The weight of the d-th input variable should be a monotonically increasing function, its correlation coefficient being an absolute value, and the variable weight used in the present invention is defined as
Training VW-AE in semi-supervised manner in step 1 using variable weights
Step 4, constructing a deep network model
The Stacked self-encoders (SAE) of the present invention are implemented by removing each decoding layer of the semi-supervised self-encoding, and stacking the encoding layers to form a deep network. The SAE model is trained layer by layer and finally uniformly fine-tuned. For input data X L ＝{x 1 ,x 2 ...x 10 }，x i Is the time sequence of the ith variable, and the corresponding marker output data set Y L ＝{y 1 ,y 2 ,...,y j ...y NL -in a semi-supervised mannerThe first layer AE is trained. The pre-training weight is { W ] 1 ，b 1 First build linear decoder model:
H 1 ＝f(W 1 X+b) (18)
for unlabeled data; />Data for marking.
H 2 ＝f(W 2 f(W 1 X+b)) (19)
For the second layer AE, the hidden features h of the first layer AE i Is provided to the input layer. h is a i Is weighted by the marked dataAnd target output Y L To calculate. Then by minimizing +.>The loss function trains the AE. After training is completed, a weight { W } of the pre-training is obtained 2 ，b 2 }. Furthermore, the hidden feature of the second layer AE is denoted +.>And->
In a similar manner as in step 2, the pre-training weights of all AEs are { W }, using the loss function minimization with variable weights added in step 3, proceeding layer by layer in a semi-supervised pre-training method until the last AE is obtained k ，b k } k＝1,2,...,l 。
Step 5, reverse training fine tuning weight
After the pre-training is finished, the output layer is added to the top of VE-SAE to fine tune the weights,weight { W for pre-training k ，b k } k＝1,2,...,l Is used to initialize the weights of each hidden layer. By applying a back propagation algorithm, improved weights can be obtained. According to the loss function calculated in the step 3, through minimizing layer-by-layer back propagation, in the back propagation process, error readjustment and distribution are realized, and weight and bias are corrected again for the distributed error, so that model errors are reduced, and the purpose of learning optimization is achieved. The back propagation process is sequentially performed, the weight and the paranoid parameters of the model and the variable weight value of each layer AE newly proposed by the invention are corrected each time, and when the error is reduced to a certain degree or reaches the specified iteration times, the iteration process is stopped. The following gives the partial derivative derivation formula for each parameter according to the BP algorithm.
Defining residuals for jth neuron on custom encoded output layer
Wherein,representing the weighted output from the j-th neuron of the encoder output layer, f' (·) represents the derivative of the nonlinear activation function. After the residual from the encoder output layer is found, the derivative of the weights and offsets of the hidden layer to the self-encoded output layer can be found according to the BP algorithm.
Wherein the method comprises the steps ofRepresenting hidden layer kth neuron toThe connection weight of the jth neuron of the output layer from the encoder, where α k Representing the activation function output of the kth neuron of the hidden layer.
The residuals of the j-th neuron of the neural network output layer are defined.
Similarly, the derivative of the weights and biases of the hidden layer to the neural network output layer can be obtained
Next, when defining the residual error of the jth neuron of the hidden layer, it can be seen that the residual error of the hidden layer may be derived from the two-part output, the self-encoder output and the neural network output, and for the labeled sample, there is not only the self-encoded output but also the neural network output, and for the unlabeled sample, since there is no actual label, only the self-encoder output, so that when defining the residual error of the hidden layer, it needs to be clearly defined.
If the input samples are unlabeled samples, defining the residual of the jth neuron of the hidden layer
If the input samples are labeled samples, defining the residual of the jth neuron of the hidden layer
The derivative formulas of all the weights and the bias parameters of the whole model can be obtained, and then the model parameters can be continuously and iteratively updated through a gradient descent method, so that the error is known to accord with the expected standard or the specified iteration times are reached. The update formula is as follows:
wherein W is old Representing weights prior to iterative update, W new Represents the weight after updating, N represents the number of samples, μ represents the learning rate, E i Representing the loss function calculated for the i-th sample. In this way, all weights and offsets can be updated layer by layer in sequence through the formula, and fine adjustment of parameters of the whole network is completed.
Step 6, real-time online prediction by using model
Aiming at the problems of limited tag data and nonlinearity in the actual industry, a self-encoder is combined with a neural network, limited tag data and a large amount of non-tag data are used at the same time, and a nonlinear soft measurement model of an industrial process is finally established. Fig. 5 shows the algorithm flow of the semi-supervised self-encoder model applied to soft measurements. When in prediction, the newly acquired process data is utilized, the process data is input into a semi-supervised self-coding model after being preprocessed, and the output layer finally outputs a corresponding predicted value, so that the real-time online prediction of f-CaO can be realized.
The prediction result is shown in fig. 6, and compared with the traditional soft measurement system without supervision and with supervision, the semi-supervised soft measurement system provided by the invention improves the prediction precision and reduces the prediction loss as shown in fig. 7.
The foregoing description of the preferred embodiments of the present invention is merely illustrative, and not intended to limit the scope of the invention, and various modifications and improvements may be made by those skilled in the art without departing from the spirit of the invention, and the scope of the invention is defined by the appended claims.
Claims (8)
1. Cement free calcium soft measurement system with quality target as guiding semi-supervised learning, characterized in that: the method comprises the following steps:
step 1: determining auxiliary variables and performing data processing: through analysis of the industrial process flow, primarily selecting an easily-measured variable related to the difficult-to-measure variable as an auxiliary variable of the soft measurement model, and collecting a time sequence of the auxiliary variable and the difficult-to-measure parameter; thirdly, adopting a 3 sigma criterion to propose an abnormal value of the acquired time sequence, and carrying out normalization processing on the data before training;
step 2: constructing a semi-supervised model, and simultaneously determining model parameters: constructing a semi-supervised prediction model by using the input variables selected in the step 1, and determining initial parameters, wherein the initial parameters comprise a learning rate mu, a regularization coefficient lambda and a balance coefficient alpha; the semi-supervised self-encoder model combines a self-encoding and neural network model, wherein the self-encoding and the neural network share an input layer and a hidden layer, the output layer is divided into a self-encoding output layer and a neural network output layer, and a balance coefficient is utilized to combine a self-encoding reconstruction error and a neural network prediction error to jointly form a loss function of the model;
the semi-supervised self-encoder is a combination of an automatic encoder and a single-output neural network, wherein the automatic encoder part is used for learning unlabeled sample information, and the neural network part is used for learning labeled samples; wherein the automatic encoder shares an input layer and a hidden layer with the single output neural network; let d= { L, U }, where D represents the modeling useData set, l= { X, Y } = { (X) 1 ，y 1 )，(x 2 ，y 2 )，…(x |L| ，y |L| ) The data set of labeled samples is represented by L, the number of labeled samples is represented by u= { X } = { X 1 ,x 2 ,x 3 ,…x |U| The label-free sample data set is represented, where each sample time series includes 60 sample points: x is x i ＝{x i(1) ,x i(2) ,x i(3) …x i(60) }，x i Is the time sequence of the ith variable, |u| represents the number of unlabeled exemplars; semi-supervised self-encoder model forward propagation is identical to the forward propagation process of the self-encoder model and neural network model:
h＝f(W 1 X+b) (4)
wherein X is the input of a semi-supervised automatic encoder, W 1 Is the weight of the input layer to the hidden layer, b is the bias between the input layer and the hidden layer, W nn Is the weight of hidden layer to label layer, b nn Is bias of hidden layer to label, W ae Is the weight of the hidden layer to the reconstructed output layer, b ae Is the bias of the hidden layer to the reconstructed output layer, h is the output of the semi-supervised automatic encoder hidden layer,is the predictive output of the neural network part, +.>Is the input reconstruction of the automatic encoder, f (·) is the activation function, here the ReLU function is employed as the activation function, whose expression is as follows:
f(x)＝max(0,x) (7)
the loss function of a semi-supervised self encoder is defined as follows:
wherein N is L Is the number of labeled samples, N U Is the number of unlabeled exemplars, E is the overall error of the semi-supervised self encoder, E ae Is the reconstruction error of the semi-supervised self-encoder, E nn Is the prediction error of the neural network, alpha is the weight coefficient,is the reconstruction of the ith unlabeled exemplar, < >>Is the predicted value of the ith javelin sample, lambda is the regularization coefficient;
step 3: variable weighted AE hierarchical pre-training: the aim of the layered pre-training is to provide better initial weights for the later constructed stacked self-encoders to better highlight process variables more relevant to target output, and the correlation coefficients of the input variables and the target variables are calculated by using the marking data to give weight values of corresponding degrees;
step 4: building a depth network model: the semi-supervised model in the step 2 is utilized to remove the respective coding output layers, and encoder hidden layers are stacked to form a deep network; simultaneously minimizing the weight loss function in the step 3, and initializing the weight value of each layer;
step 5: reverse fine tuning of weights: adding an output layer to the overall fine tuning weight at the top end of the model, and realizing error readjustment and distribution of updated weights of all layers through a BP algorithm;
step 6: real-time online prediction using a model: and carrying out real-time online prediction on the variables acquired in real time by using the model constructed in the prior art.
2. The mass-targeted semi-supervised learning cement free calcium soft measurement system of claim 1, wherein: in step 3, a semi-supervised model of the output layer of the integrated self-encoder and the output layer of the neural network is constructed, a period of sequence of corresponding variables is extracted, balance coefficients of the loss functions of the two output layers are considered on the loss functions, weight duty ratio of each layer is added to the balance coefficients, and the self-coding model is pre-trained to initialize parameters of the balance coefficients.
3. The mass-targeted semi-supervised learning cement free calcium soft measurement system of claim 1, wherein: in step 4, a deep training model is constructed, an output layer is added at the top end, and model parameters of each layer are updated through reverse training.
4. The mass-targeted semi-supervised learning cement free calcium soft measurement system of claim 1, wherein: in step 5, the model is predicted online.
5. The mass-targeted semi-supervised learning cement free calcium soft measurement system of claim 1, wherein: in step 1, 10 process variables closely related to free calcium oxide are selected as auxiliary variables of a free calcium soft measurement model by analyzing the mechanism and related influencing factors of free calcium oxide production in the clinker.
6. The mass-targeted semi-supervised learning cement free calcium soft measurement system of claim 5, wherein: in step 1, the 10 modeling auxiliary variables selected are: feeding quantity 1 feedback, feeding quantity of decomposing furnace coal feedback, rotating speed feedback of a high-temperature fan, kiln tail temperature, decomposing furnace outlet temperature, kiln current feedback, two-chamber grate lower pressure feedback, secondary air temperature feedback, kiln head negative pressure feedback and kiln head coal feedback.
7. The mass-targeted semi-supervised learning cement free calcium soft measurement system of claim 1, wherein: in step 6, the self-encoder is combined with the neural network aiming at the problems of limited label data and nonlinearity, meanwhile, a nonlinear soft measurement model of the industrial process is finally established by using limited label data and a large amount of label-free data, the newly acquired process data is utilized for prediction, the newly acquired process data is input into the semi-supervised self-encoding model after preprocessing, and the output layer finally outputs corresponding predicted values.
8. The mass-targeted semi-supervised learning cement free calcium soft measurement system of claim 1, wherein: in step 5, the following are included: after the pre-training is finished, adding an output layer to the top end of VE-SAE to fine tune the weight, and using the pre-trained weight to initialize the weight of each hidden layer; obtaining improved weights by applying a back propagation algorithm; according to the loss function calculated in the step 3, through minimizing layer-by-layer back propagation, in the back propagation process, error readjustment and distribution are realized, and weight and bias are corrected again for the distributed error, so that model error is reduced, and the purpose of learning and optimizing is achieved; the back propagation process is sequentially carried out, the weight and the paranoid parameters of the model and the variable weight value of each layer of AE are corrected each time, and when the error is reduced to a certain degree or reaches the specified iteration times, the iteration process is stopped; and obtaining derivative formulas of all weights and bias parameters of the whole model, and then continuously and iteratively updating the model parameters through a gradient descent method to obtain the iteration times of which the error meets the expected standard or reaches the stipulation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110845983.9A CN113593657B (en) | 2021-07-26 | 2021-07-26 | Cement free calcium soft measurement system with quality target as guiding semi-supervised learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110845983.9A CN113593657B (en) | 2021-07-26 | 2021-07-26 | Cement free calcium soft measurement system with quality target as guiding semi-supervised learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113593657A CN113593657A (en) | 2021-11-02 |
| CN113593657B true CN113593657B (en) | 2024-01-09 |
Family
ID=78250127
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110845983.9A Active CN113593657B (en) | 2021-07-26 | 2021-07-26 | Cement free calcium soft measurement system with quality target as guiding semi-supervised learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113593657B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109147878A (en) * | 2018-10-08 | 2019-01-04 | 燕山大学 | A kind of clinker free calcium flexible measurement method |
| CN110444257A (en) * | 2019-08-05 | 2019-11-12 | 燕山大学 | It is a kind of based on unsupervised and supervised learning cement free calcium flexible measurement method |
| CN112365935A (en) * | 2020-10-20 | 2021-02-12 | 燕山大学 | Cement free calcium soft measurement method based on multi-scale depth network |
| WO2021128510A1 (en) * | 2019-12-27 | 2021-07-01 | 江苏科技大学 | Bearing defect identification method based on sdae and improved gwo-svm |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190042952A1 (en) * | 2017-08-03 | 2019-02-07 | Beijing University Of Technology | Multi-task Semi-Supervised Online Sequential Extreme Learning Method for Emotion Judgment of User |
-
2021
- 2021-07-26 CN CN202110845983.9A patent/CN113593657B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109147878A (en) * | 2018-10-08 | 2019-01-04 | 燕山大学 | A kind of clinker free calcium flexible measurement method |
| CN110444257A (en) * | 2019-08-05 | 2019-11-12 | 燕山大学 | It is a kind of based on unsupervised and supervised learning cement free calcium flexible measurement method |
| WO2021128510A1 (en) * | 2019-12-27 | 2021-07-01 | 江苏科技大学 | Bearing defect identification method based on sdae and improved gwo-svm |
| CN112365935A (en) * | 2020-10-20 | 2021-02-12 | 燕山大学 | Cement free calcium soft measurement method based on multi-scale depth network |
Non-Patent Citations (1)
| Title |
|---|
| 基于粒子群算法与连续型深度信念网络的水泥熟料游离氧化钙预测;刘彬;赵朋程;高伟;孙超;刘浩然;;计量学报(第03期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113593657A (en) | 2021-11-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110444257B (en) | Cement free calcium soft measurement method based on unsupervised and supervised learning | |
| CN110378044B (en) | Multi-time scale convolution neural network soft measurement method based on attention mechanism | |
| CN109147878A (en) | A kind of clinker free calcium flexible measurement method | |
| CN106202946A (en) | Clinker free calcium levels Forecasting Methodology based on degree of depth belief network model | |
| CN109647899B (en) | Method for forecasting power consumption of multi-specification rolled pieces in hot rolling and finish rolling process of strip steel | |
| CN113128113B (en) | Lean information building load prediction method based on deep learning and transfer learning | |
| CN112668234A (en) | Intelligent control method for steelmaking endpoint of converter | |
| Li et al. | Dynamic time features expanding and extracting method for prediction model of sintering process quality index | |
| CN103745101A (en) | Improved neural network algorithm based forecasting method of set value of rolling force of medium plate | |
| Liu et al. | Model fusion and multiscale feature learning for fault diagnosis of industrial processes | |
| CN107868979A (en) | A kind of silicon single crystal diameter control method based on permanent casting speed control structure | |
| CN109685283B (en) | Method for predicting submerged arc furnace working condition based on confidence rule base reasoning | |
| CN114692993A (en) | Water conservancy facility deformation prediction method fusing ARIMA and BilSTM in seasons | |
| CN113177358A (en) | Soft measurement method for cement quality based on fuzzy fine-grained feature extraction | |
| CN114841073A (en) | Instant learning semi-supervised soft measurement modeling method based on local label propagation | |
| CN114721263A (en) | Intelligent regulation and control method for cement decomposing furnace based on machine learning and intelligent optimization algorithm | |
| CN113593657B (en) | Cement free calcium soft measurement system with quality target as guiding semi-supervised learning | |
| Ji et al. | Application of the improved the ELM algorithm for prediction of blast furnace gas utilization rate | |
| CN114139439A (en) | Steam turbine optimal initial pressure determination method based on simulated annealing particle swarm algorithm | |
| CN114117852A (en) | Regional heat load rolling prediction method based on finite difference working domain division | |
| CN110033126A (en) | Shot and long term memory network prediction technique based on attention mechanism and logistic regression | |
| CN116483036B (en) | Transformer-based self-encoder soft measurement modeling method | |
| CN109101683B (en) | Model updating method for pyrolysis kettle of coal quality-based utilization and clean pretreatment system | |
| CN114943277A (en) | Method for identifying operation condition and predicting operation performance of comprehensive energy system | |
| CN115392125A (en) | Temperature prediction method for rotary cement kiln |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |