CN114994294B - Soft measurement method for free calcium of cement clinker based on attention and window gating mechanism - Google Patents
Soft measurement method for free calcium of cement clinker based on attention and window gating mechanism Download PDFInfo
- Publication number
- CN114994294B CN114994294B CN202210553703.1A CN202210553703A CN114994294B CN 114994294 B CN114994294 B CN 114994294B CN 202210553703 A CN202210553703 A CN 202210553703A CN 114994294 B CN114994294 B CN 114994294B
- Authority
- CN
- China
- Prior art keywords
- time
- variable
- variables
- window
- data
- 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
- 239000004568 cement Substances 0.000 title claims abstract description 79
- 230000007246 mechanism Effects 0.000 title claims abstract description 46
- OYPRJOBELJOOCE-UHFFFAOYSA-N Calcium Chemical compound [Ca] OYPRJOBELJOOCE-UHFFFAOYSA-N 0.000 title claims abstract description 33
- 229910052791 calcium Inorganic materials 0.000 title claims abstract description 33
- 239000011575 calcium Substances 0.000 title claims abstract description 33
- 238000000691 measurement method Methods 0.000 title claims abstract description 10
- 238000000034 method Methods 0.000 claims abstract description 210
- 230000008569 process Effects 0.000 claims abstract description 173
- 238000005259 measurement Methods 0.000 claims abstract description 78
- 238000004519 manufacturing process Methods 0.000 claims abstract description 55
- 230000008878 coupling Effects 0.000 claims abstract description 18
- 238000010168 coupling process Methods 0.000 claims abstract description 18
- 238000005859 coupling reaction Methods 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims abstract description 14
- 230000009471 action Effects 0.000 claims abstract description 6
- 230000015654 memory Effects 0.000 claims description 24
- 238000004364 calculation method Methods 0.000 claims description 16
- 239000000284 extract Substances 0.000 claims description 16
- 238000007781 pre-processing Methods 0.000 claims description 16
- 230000006870 function Effects 0.000 claims description 14
- 238000012549 training Methods 0.000 claims description 13
- 230000004913 activation Effects 0.000 claims description 11
- 230000003044 adaptive effect Effects 0.000 claims description 10
- 230000007774 longterm Effects 0.000 claims description 10
- 230000002159 abnormal effect Effects 0.000 claims description 9
- 230000007787 long-term memory Effects 0.000 claims description 9
- 230000006403 short-term memory Effects 0.000 claims description 9
- 238000011176 pooling Methods 0.000 claims description 7
- 239000002994 raw material Substances 0.000 claims description 7
- 230000008859 change Effects 0.000 claims description 6
- 238000012360 testing method Methods 0.000 claims description 6
- 238000009826 distribution Methods 0.000 claims description 5
- 238000000605 extraction Methods 0.000 claims description 4
- 238000013459 approach Methods 0.000 claims description 3
- 238000001354 calcination Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims description 3
- 238000013136 deep learning model Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 230000000717 retained effect Effects 0.000 claims description 3
- 238000012546 transfer Methods 0.000 claims description 3
- 150000001875 compounds Chemical class 0.000 claims description 2
- 230000009286 beneficial effect Effects 0.000 abstract description 3
- 238000005265 energy consumption Methods 0.000 abstract description 3
- 238000012544 monitoring process Methods 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 5
- 238000010219 correlation analysis Methods 0.000 description 2
- 230000001934 delay Effects 0.000 description 2
- 238000012628 principal component regression Methods 0.000 description 2
- 238000004891 communication Methods 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
- G01N33/38—Concrete; Lime; Mortar; Gypsum; Bricks; Ceramics; Glass
- G01N33/383—Concrete or cement
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Theoretical Computer Science (AREA)
- Chemical & Material Sciences (AREA)
- General Physics & Mathematics (AREA)
- Immunology (AREA)
- Computational Linguistics (AREA)
- Analytical Chemistry (AREA)
- Medicinal Chemistry (AREA)
- Food Science & Technology (AREA)
- Pathology (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Biochemistry (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Ceramic Engineering (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a soft measurement method for cement clinker free calcium based on attention and a window gating mechanism, which belongs to the technical field of soft measurement and comprises the steps of considering the coupling relation among process variables by using an attention decoupling network model, processing the mutual influence among the variables and realizing the decoupling of process data; the window gating module is used for adaptively learning and gating the time period with the most influence on the target variable, and the learned time window is distributed with corresponding weights at different moments, so that the action duration of the process variable is concerned more effectively; and respectively and independently extracting the characteristics of the single variable through a one-dimensional convolution network, and integrating the characteristic information of each variable to construct a soft measurement model meeting the actual requirement so as to realize the measurement of the f-CaO of the cement clinker. According to the invention, by considering the natural characteristics in the cement production industrial process, the measurement performance of the soft measurement model is improved, and the method has guiding significance for quality monitoring and real-time control, and is beneficial to reducing energy consumption and improving production efficiency.
Description
Technical Field
The invention relates to the technical field of soft measurement, in particular to a soft measurement method for free calcium of cement clinker based on attention and a window gating mechanism.
Background
In the industrial process of cement, the content of clinker free calcium (f-CaO) is an index for measuring the quality of cement, and the stability and strength of the cement quality are influenced. Due to the harsh environment of industrial production, the sensors cannot acquire f-CaO values in real time and are typically chemically analyzed in a laboratory in an off-line manner. So that the f-CaO value cannot be monitored in real time and the cement quality cannot be controlled in the cement production process.
Measurement for achieving quality targets in the process industry is of great significance to production, and how to improve the performance of soft measurement models is also a problem that researchers are constantly exploring in recent years. The scholars propose Principal Component Regression (PCR), cross-correlation analysis and partial least squares to analyze the relationship between variables and then fit the inputs and outputs using algorithms, the modeling method described above only analyzes the linear relationship of simple inputs (process variables) to outputs (f-CaO). As for the time delay between the process variable and the f-CaO content, mutual information and a correlation analysis method are generally adopted to estimate the time delay, and the time-varying characteristic of the time delay is easily ignored due to different production conditions and different variation ranges of the time delay. The time delay between process variables is analyzed by adopting a sliding time window method, and the time delay extracted by the method cannot solve the problem of different process variable action time lengths caused by time delay changes generated by different production conditions due to a fixed time window, and cannot determine the time length which really affects a target variable.
Therefore, there is a need to develop a soft measurement method for free calcium of cement clinker based on attention and window gating mechanism to solve the above problems.
Disclosure of Invention
The invention aims at solving the technical problem of providing a soft measurement method for cement clinker free calcium based on attention and a window gating mechanism, aiming at the production characteristics in the cement production industrial flow, simultaneously considering the strong coupling relation of process variables in the industrial flow, time delay variation under different production conditions and time delay of the variables, constructing a soft measurement model with excellent performance, effectively improving the measurement performance of the soft measurement model, having guiding significance for quality monitoring and real-time control in the cement production industrial flow, being beneficial to reducing energy consumption and improving production efficiency.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
a soft measurement method of cement clinker free calcium based on attention and window gating mechanism, aiming at the production characteristics of data coupling in the cement production process, an attention decoupling network model is utilized to consider the coupling relation between process variables, process the mutual influence between the variables and realize the decoupling of the process data; aiming at the time delay change under different production conditions and the time delay of a process variable, a learnable window gating module is used for adaptively learning and gating the time period which has the most influence on the target variable, and the learnt time window is distributed with corresponding weights at different moments, so that the action duration of the process variable is more effectively concerned; the characteristics of the single variables are respectively and independently extracted through a single-dimensional convolution network, the characteristic information of each variable is integrated, a soft measurement model meeting the actual requirement is constructed, and the measurement of the free calcium of the cement clinker is realized.
The technical scheme of the invention is further improved as follows: the measuring method specifically comprises the following steps:
step 2, constructing an attention decoupling network model;
step 3, constructing a learnable window gating mechanism module;
step 4, combining the single-dimensional convolution to construct a soft measurement model;
and 5, carrying out online measurement on the soft measurement model.
The technical scheme of the invention is further improved as follows: in the step 1, selecting the process variables which have the most influence on the target variables on equipment at each position in the cement production process, and selecting the process variables at a raw material inlet, a raw material preheater, a decomposing furnace, a rotary kiln and a grate cooler respectively, wherein the selected process variables comprise the temperature of the decomposing furnace, the feeding amount, the outlet temperature of a primary cylinder, the temperature of a kiln tail, the temperature of secondary air, the current of the kiln and the negative pressure of a kiln head.
The technical scheme of the invention is further improved as follows: in the step 1, downloading process variable and clinker free calcium content data from a cement database for data preprocessing; the data preprocessing flow comprises abnormal value identification and processing and data standardization;
the specific method for identifying and processing the abnormal value comprises the following steps:
setting the ith process variable time series datan is the number of data, the abnormal value identification and processing adopt a 3 sigma criterion, and the specific calculation formula is as follows:
in the formula (I), the compound is shown in the specification,representing time series data x i Represents the time series data x i The variance of (a);
the specific method for data standardization comprises the following steps:
dimensional differences exist among the collected process variable data, the numerical value and the measurement unit of each process variable are different, and the learning and nonlinear fitting capabilities of the deep learning model are seriously influenced by the differences; the standardization of data adopts 0-1 standardization to eliminate dimension difference, and the 0-1 standardization formula is as follows:
in the formula, x min Representing time series data x i Minimum value of (a), x max Represents time-series data x i Of (c) is calculated.
The technical scheme of the invention is further improved as follows: in the step 2, the method specifically comprises the following steps:
2.1, inputting a process variable time sequence into an LSTM unit;
all time series is X = { X i ,i=1,2,3....M}∈R N×M The time series of the process variable isWherein N represents the time series length, M represents the number of process variables, x i Representing the time of the ith process variableA sequence; inputting the time sequence of the process variable into the LSTM, extracting the time sequence characteristics by using the LSTM, and outputting the hidden state h in the model t The formula is as follows:
2.2, decoupling between variables is realized by applying an attention mechanism;
after the data preprocessing is finished, inputting the time sequence of each process variable into the LSTM to obtain the hidden state of each time sequence; process variables are mutually influenced and have strong coupling; by utilizing an attention mechanism, the influence of other process variables on the current variable is noticed, the weight distribution of the other variables on the current variable is found, and the finally output current variable contains the time sequence characteristics of other variables, so that the decoupling purpose is achieved; the decoupled attention configuration is as follows; mapping a current variable to a vector eta with full concatenation by combining the hidden state of the current variable and the hidden states of the remaining variables t :
The weights calculated are:
wherein w, u, v are learnable parameters,a hidden state representing a current variable; combining the hidden state of the current variable with the hidden states of the rest variables; />Multiplication of representative elements; the resulting weight vector is multiplied by the time series of the original input:
extracting time sequence characteristics of each process variable through LSTM, processing the obtained hidden state through attention mechanism to obtain the influence of other variables on the current variable, distributing different weights, and finally outputting a vector v i Is the current variable that contains the remaining variable information.
The technical scheme of the invention is further improved as follows: in the step 3, the method specifically comprises the following steps:
3.1, determining the center of a learnable time window;
the accuracy of the model is seriously influenced by time delay existing in cement production, the calcination of cement is a complex and variable chemical reaction process, the time delay between process variables and the time delay between the process variables and the content of free calcium are all the results of equipment, and the variables are interacted and mutually influenced; the time delay is difficult to determine due to the complex industrial process; aiming at variable time delay in the cement industry process, a learnable window gating mechanism, a time period for adaptively learning each process variable, and a soft measurement model fused with a window gating mechanism module, wherein the time delay characteristic in the cement production process and the time delay change under different production conditions are considered at the same time;
v i is the output of the attention decoupling module, representing the process variable for which the decoupling is completed; c. C n W is a memory unit of the long-term and short-term memory network, which represents the long-term state in the network structure, the long-term state is a special hidden state, the long-term and short-term memory network can extract important information in a time sequence and store the characteristics in the long-term state, a calculation formula of an adaptive window center is provided according to the characteristics of the memory unit, and the information retained in the memory unit is extracted in the network learning process; the time window center is adaptively learned by utilizing the time sequence characteristics; the specific formula is as follows:
p i =T*sigmod(w p *c n _w+b p ) (7)
where i ∈ {1,2,3.. M }, M is the number of process variables, T represents the length of the sample sequence, p represents i The center of the window representing the ith variable, whose value is a number, w p And b p Weight and bias, respectively, for linear operations; calculating by using the formula, wherein each process variable can learn the corresponding time window center;
3.2, determining the length of the learnable time window;
h n w is the hidden state of the long-term and short-term memory network, the hidden state in the network is to extract partial information in the memory unit, which mainly depends on the state of the output gate, the value of the output gate in the network directly determines how much information the hidden state can extract in the memory unit, and the function is to effectively transfer the information in the memory unit to the hidden state; since the hidden state can learn in the network model and partially extract the information in the memory unit, its characteristics can be mapped to the length calculation of the time window; by using the length of the hidden state self-adaptive learning time window, the specific calculation formula is as follows:
where i is the {1,2,3.. M }, M is the number of process variables, T represents the length of the sample sequence, and l is i The length of the window representing the ith variable is a number, w l And b l Weight and bias, respectively, for linear operations; calculating by using the formula, wherein each process variable can learn the corresponding time window length;
3.3, determining a time window;
the determined window center and the determined window length are required to be applied to the time sequence, and the window center and the window length are combined to calculate a time period on the time sequence; and (3) gating by using an activation function and determining a time period which is most influenced by each variable on the f-CaO, wherein a specific calculation formula is as follows:
α i =sigmod(Relu(l i -abs(t-p i ))) (9)
wherein T is belonged to {1,2,3.. T; t =60, T stands for each time instant in the time series, abs denotes the absolute value of the parameter, p i And l i Respectively window center and window length, alpha i The weight value is distributed to each moment in the time sequence by using an activation function; by utilizing Relu and sigmod activation functions, the weight distributed to the time on the corresponding time period approaches to 1, and the rest times are all set to 0.5; different weight distribution gates the time period that each variable is self-adaptively learned;
each process variable selects a corresponding time period by using the distributed time weight, the time weight is combined with the time series characteristics of the process variable, and the time information which has the most influence on the f-CaO is adaptively gated, wherein the formula is as follows:
wherein i is an element of {1,2,3.. M }, M is the number of process variables,denotes multiplication of elements, alpha i Represents the weight value assigned in the time series in the ith process variable, <' >>Output of the long-short term memory network, Y, representing the ith process variable i Is the final output of the adaptive learning window gating module, representing the time window gated by each variable.
The technical scheme of the invention is further improved as follows: in the step 4, the method specifically comprises the following steps:
4.1, constructing a single-dimensional convolution model;
the single-dimensional convolution is different from the two-dimensional convolution, the single-dimensional convolution only extracts the characteristics in one dimension, and the characteristics of the single-dimensional convolution are utilized to process the cement process data of the time sequence;
x in is the input of a single-dimensional convolutionData, x median Is the output of a single-dimensional convolution, x out Is the final output result after maximum pooling, w is the weight magnitude in the convolution kernel, b is the bias corresponding to the weight in the single-dimensional convolution layer, relu is a nonlinear activation function, and the simple formula of the single-dimensional convolution layer is as follows:
Relu=max(0,x) (11)
x median =Relu(w·x in +b) (12)
x out =MAX(x median ) (13)
4.2, constructing a final soft measurement model;
the time sequence passing through the window gating module and the time window of the adaptive gating process variable distribute weights to all moments on the time sequence of the process variable, the gated time period contains important time sequence characteristics, and deep extraction of the time sequence characteristics is realized by utilizing single-dimensional convolution; by increasing the number of the layers of the convolutional layers and continuously extracting the time sequence characteristics influencing free calcium more deeply, a soft measurement model meeting the requirements is developed;
the 3 layers of single-dimensional convolutional layers and the corresponding 3 layers of maximum pooling layers are utilized to extract the characteristics of the time sequence with time weight more deeply, the independence among all variables is enhanced through the time sequence processed by the attention decoupling module, the variables are not influenced mutually, and better time sequence data are provided for a later self-adaptive learning time window gating module; according to the single-dimensional convolution proposed later, due to the characteristics of the single-dimensional convolution, features can be extracted from each variable, and the variables cannot be influenced mutually; the characteristic ensures that the process variable characteristics are extracted, the relative independence of the variables is further enhanced, the multilayer convolution layers extract the characteristics of each process variable, the deep-level characteristics extracted by each variable are fused together through full connection, the finally output characteristic vector is the characteristic set of each process variable, and the characteristic set is used for measuring the f-CaO content;
4.3, training a soft measurement model;
selecting sample data in a cement database, dividing the sample data into a training set and a testing set, inputting the processed training data into a soft measurement model, and training the soft measurement model by using the sample data; and finally, extracting the characteristics among the process variables through convolution after the treatment of each proposed module, fusing the characteristics of each process variable, and measuring the content value of free calcium in the cement clinker.
The technical scheme of the invention is further improved as follows: step 5, storing parameters of the finally trained soft measurement model, and constructing a soft measurement model for measuring the content of free calcium in the cement clinker; and (3) preprocessing the data sample acquired in real time, inputting the preprocessed data sample into the constructed soft measurement model, and measuring the value of free calcium of the cement clinker in real time through the model.
Due to the adoption of the technical scheme, the invention has the technical progress that:
1. the attention decoupling network module combining the LSTM and the attention mechanism can be used for paying attention to and finding the relation among all variables by the attention mechanism aiming at the strong coupling relation among the process variables in the cement process, decoupling the process variables, eliminating the influence among the variables, enabling the process variables to be relatively independent and improving the measurement performance of the model.
2. The window gating mechanism network module provided by the invention adaptively learns and gates corresponding time periods, the time periods are process data directly influencing the f-CaO content, the module has adaptive learning characteristics, different action durations of process variables caused by time delays of different variables in a cement production process can be learned, and the measurement accuracy and the generalization capability of the model are effectively improved.
3. According to the soft measurement model of the cement clinker free calcium based on the attention and window gating mechanism, aiming at inherent characteristics in a cement flow, the two methods of the attention decoupling network module and the window gating mechanism network module which are combined by the LSTM and the attention mechanism are combined, and the strong coupling relation between two most main production characteristics in the cement flow, namely process variables and the time-varying delay of the process variables are considered, so that the measurement performance of the soft measurement model is effectively improved.
4. The invention effectively improves the measurement performance of the soft measurement model by considering the natural characteristics in the cement production industrial process, has guiding significance for quality monitoring and real-time control in the cement production industrial process, and is beneficial to reducing energy consumption and improving production efficiency.
Drawings
FIG. 1 is a flow chart of a soft measurement model implemented in an embodiment of the present invention;
FIG. 2 is a diagram of an implementation of the method for attention decoupling according to an embodiment of the present invention;
FIG. 3 is a diagram illustrating an exemplary implementation of a learnable window gating mechanism in an embodiment of the present invention;
FIG. 4 is a diagram illustrating a single-dimensional convolution implementation in a soft measurement model according to an embodiment of the present invention;
FIG. 5 is a soft measurement model combining two methods in an embodiment of the present invention;
FIG. 6 is a diagram of an overall soft measurement model in an embodiment of the present invention;
FIG. 7 is a graph of test results in an embodiment of the present invention.
Detailed Description
The embodiment of the application provides a soft measurement method for free calcium of cement clinker based on attention and a window gating mechanism, solves the problems that in the prior art, the acting time of process variables is different due to time delay changes generated by different production conditions, and the time length which really affects target variables cannot be determined, and provides an attention decoupling method for considering the coupling between variables and eliminating the coupling relation aiming at two production characteristics (strong coupling relation between process variables and time-varying time delay of the process variables) which have the largest influence on the f-CaO content in cement production; the method is characterized in that a learnable window gating mechanism is used for learning a corresponding time window, a time period which has the most influence on the f-CaO content value of the cement clinker in the variables in the learning process is combined with an attention decoupling method and the learnable window gating mechanism, and the time characteristics of each variable are extracted through one-dimensional convolution, so that the measurement performance of the soft measurement model is improved.
The invention is described in further detail below with reference to the following figures and examples:
as shown in fig. 1, a soft measurement method for cement clinker free calcium based on attention and window gating mechanism, aiming at the production characteristics of data coupling in the cement production process, utilizes an attention decoupling network module to analyze and consider the mutual influence among variables, decouples the process variable data with strong coupling relation, and improves the independence among all variables; secondly, aiming at the time delay change under different production conditions and the time delay of the process variable, a learnable window gating mechanism is utilized to gate the time period with the maximum influence of the process variable on the target variable, and the two methods are combined to construct a soft measurement model meeting the actual requirement through single-dimensional convolution. And finally, constructing a soft measurement model, simultaneously considering the two most important cement production characteristics (the strong coupling relation between the process variables and the time-varying delay of the process variables), extracting the characteristics of each process variable by utilizing single-dimensional convolution, finally fully connecting and fusing the extracted variable characteristics, and measuring the content value of the f-CaO in the cement clinker.
The method specifically comprises the following steps:
specifically, the method comprises the following steps: through on-site learning and communication with control personnel, operators explain the process variables most influencing the f-CaO content, the process variables are respectively distributed at each position in the cement process equipment, and the position variables reflect the production conditions of each link in the process, so that the production quality of cement is directly influenced. Therefore, the position variable in the cement production facility is selected. Selecting the process variables which have the most influence on the target variables on equipment at each position in the cement process, selecting the process variables at a raw material inlet, a raw material preheater, a decomposing furnace, a rotary kiln and a grate cooler respectively, and finally selecting 7 position process variables: raw material feeding quantity, primary barrel outlet temperature, decomposing furnace outlet temperature, kiln head negative pressure, rotary kiln current and kiln tail temperature.
Downloading process variable and clinker free calcium (f-CaO) content data from a cement database for data preprocessing; the data preprocessing flow comprises abnormal value identification and processing and data standardization.
The specific method for identifying and processing the abnormal value comprises the following steps:
let i process variable time series datan is the number of data, the abnormal value identification and processing adopt a 3 sigma criterion, and the specific calculation formula is as follows:
The specific method for data standardization is as follows:
dimensional differences exist among the acquired variable data, the numerical value and the measurement unit of each variable are different, and the differences can seriously affect the learning and nonlinear fitting capabilities of the deep learning model. The standardization processing of the data adopts 0-1 standardization to eliminate dimension difference, and the 0-1 standardization formula is as follows:
in the formula, x min Represents time-series data x i Minimum value of (1), x max Represents time-series data x i Is measured.
Step 2, attention is paid to specific realization of a decoupling model;
the LSTM is a network module for typically processing a time sequence, important time information in the time sequence is extracted, the proposed attention decoupling module is combined with the LSTM, time sequence characteristics of each process variable are extracted through the LSTM, the obtained hidden state is subjected to attention control to obtain influences of other variables on the current variable, different weights are distributed, the final output vector is the current variable containing the information of the other variables, the attention control is utilized to notice the mutual influences among the variables, and the calculated weights are applied to the original input time sequence to decouple the input process variables.
2.1, inputting the process variable time sequence into an LSTM unit;
all time series samples are X = { X = ×) i ,i=1,2,3....M}∈R N×M The time series of the process variable isWherein N represents the time series length, M represents the number of process variables, x i Represents a time series of the ith process variable. Inputting the time sequence of the process variable into the LSTM, extracting the time sequence characteristics by using the LSTM, and outputting the hidden state h in the model t The formula is as follows:
2.2, decoupling between variables is realized by applying an attention mechanism;
after the data preprocessing is completed, the time series of each process variable is input into the LSTM to obtain the hidden state of each time series. The process variables are mutually influenced and have strong coupling. And (3) by using an attention mechanism, paying attention to the influence of other process variables on the current variable, finding the weight distribution of the other variables on the current variable, and finally outputting the current variable containing the time sequence characteristics of the other variables to achieve the aim of decoupling. The decoupled attention configuration is as follows. Mapping a current variable to a vector eta with full concatenation by combining the hidden state of the current variable and the hidden states of the remaining variables t 。
The calculated weights are:
wherein w, u, v are learnable parameters,a hidden state representing a current variable; combining the hidden state of the current variable with the hidden states of the rest variables; />Multiplication of representative elements; the resulting weight vector is multiplied by the time series of the original input:
extracting time sequence characteristics of each process variable through LSTM, processing the obtained hidden state through attention mechanism to obtain the influence of other variables on the current variable, distributing different weights, and finally outputting a vector v i Is the current variable that contains the remaining variable information. A specific implementation of the proposed attention decoupling method is shown in fig. 2.
Step 3, realizing a learnable window gating mechanism module;
the proposed window gating mechanism learns the time window center and the time window length of the process variable data by using the long-term state and the short-term state in the LSTM, and learns the time period of each process variable by using the time window which has direct influence on the target content in the window center and window length gating time sequence, and the learnable window gating mechanism can learn different action durations caused by different process variable time delays in a self-adaptive manner.
3.1, determining the center of a learnable time window;
the accuracy of the model is seriously influenced by the time delay existing in cement production, the calcination of cement is a complex and variable chemical reaction process, and the time delay between process variables and the time delay between the process variables and the f-CaO content are the results of interaction and mutual influence among all the variables of each device. The time delay is difficult to determine due to the complex industrial process. Aiming at variable time delay in the cement industry process, the invention provides a learnable window gating mechanism, adaptively learns the time period of each process variable, and fuses a soft measurement model of a window gating mechanism module while considering the time delay characteristic in the cement production process and the time delay change under different production conditions. A module for a learnable window gating mechanism is presented as shown in fig. 3.
As shown in fig. 3, v i Is the output of the attention decoupling module and represents the process variable for which the decoupling is complete. c. C n And w is a memory unit of the long-term and short-term memory network, which represents a long-term state in a network structure, the long-term state is a special hidden state, the long-term and short-term memory network can extract important information in a time sequence and store the characteristics in the long-term state, a calculation formula of an adaptive window center is provided according to the characteristics of the memory unit, and the information retained in the memory unit is extracted in the network learning process. The time window center is adaptively learned using these timing characteristics. The specific formula is as follows:
p i =T*sigmod(w p *c n _w+b p ) (7)
where i ∈ {1,2,3.. M }, M is the number of process variables, T represents the length of the sample sequence (samples take one hour of time data), p is i The center of the window representing the ith variable, whose value is a numerical value (scalar quantity), w p And b p Respectively, the weight and the offset of the linear operation. Using this formula, each process variable can learn the corresponding time window center.
3.2, determining the length of the learnable time window;
h n w is the hidden state of the long and short term memory network, the hidden state in the network is to extract partial information in the memory unit, which depends mainly on the state of the output gate, and the value of the output gate in the network is directlyThe function of the memory unit is to effectively transfer the information in the memory unit to the hidden state. Since the hidden state can be learned in the network model and partially extract the information in the memory unit, its characteristics can be mapped to the length calculation of the time window. The length of the time window is adaptively learned by utilizing the hidden state, and the specific calculation formula is as follows:
where i e {1,2,3.. M }, M is the number of process variables, T represents the length of the sample sequence (samples take one hour of time data), and l is i The window length representing the ith variable is a numerical value (scalar quantity), w l And b l Respectively, the weight and the offset of the linear operation. Using this formula calculation, each process variable can learn the corresponding time window length.
3.3, determining a time window;
the determined window center and the window length need to be applied to the time series, and the time period on the time series is calculated by combining the determined window center and the window length. And (3) gating by using an activation function and determining a time period which is most influenced by each variable on f-CaO, wherein a specific calculation formula is as follows:
α i =sigmod(Relu(l i -abs(t-p i ))) (9)
wherein T is belonged to {1,2,3.. T; t =60, T stands for each time instant in the time series, abs denotes the absolute value of the parameter, p i And l i Respectively window center and window length, alpha i Is the weight value assigned to each time instant in the time sequence by the activation function. By utilizing Relu and sigmod activation functions, the weight distributed to the time on the corresponding time period approaches to 1, and all the rest times are set to be 0.5. Different weight assignments gate the time period to which each variable adaptively learns.
Each process variable selects a corresponding time period by using the distributed time weight, the time weight is combined with the time series characteristics of the process variable, and the time information which has the most influence on the f-CaO is adaptively gated, wherein the formula is as follows:
wherein i is an element of {1,2,3.. M }, M is the number of process variables,denotes multiplication of elements, alpha i Represents the weight value assigned in the ith process variable in time series>Output of the long-short term memory network, Y, representing the ith process variable i Is the final output of the adaptive learning window gating module, representing the time window gated by each variable.
Step 4, establishing a final network model by combining the single-dimensional convolution;
due to the characteristic of the single-dimensional convolution, the characteristic can be extracted from each variable, the variables cannot be influenced mutually, the characteristic ensures that the process variable characteristic can be extracted, the relative independence of the variables is further enhanced, the characteristic extraction of the multi-layer convolution layer on each process variable can extract the data characteristic in a time window more deeply and effectively, the single-dimensional convolution combines an attention decoupling method and a window gating mechanism, the strong coupling and time delay characteristic of the process variables in the cement flow are considered at the same time, the measurement performance of a soft measurement model is improved, the extracted deep-level characteristics of each variable are fused together through full connection, the finally output characteristic vector is a characteristic set of each process variable, and the content value of free calcium of cement clinker is measured by utilizing the characteristic information of each process variable.
4.1, constructing a single-dimensional convolution model;
the single-dimensional convolution is different from the two-dimensional convolution, the single-dimensional convolution only extracts the characteristics in one dimension, and the characteristics of the single-dimensional convolution are utilized to process the cement process data of the time sequence. The specific structure of the convolution pooling is shown in FIG. 4.
x in Is input data of a single-dimensional convolution, x median Is the output of a single-dimensional convolution, x out Is the final output result after maximum pooling, w is the weight magnitude in the convolution kernel, b is the bias corresponding to the weight in the single-dimensional convolution layer, relu is a nonlinear activation function, and the simple formula of the single-dimensional convolution layer is as follows:
Relu=max(0,x) (11)
x median =Relu(w·x in +b) (12)
x out =MAX(x median ) (13)
4.2, constructing a final soft measurement model by combining the methods in the steps 2,3 and 4;
the time sequence passing through the window gating module and the time window of the adaptive gating process variable distribute weight to each moment on the time sequence of the process variable, the gated time period contains important time sequence characteristics, and deep extraction of the time sequence characteristics is realized by utilizing single-dimensional convolution. By increasing the number of the layers of the convolution layers, the time sequence characteristics influencing f-CaO are continuously extracted in a deeper layer, and a soft measurement model meeting the requirements is developed.
By utilizing the 3 layers of single-dimensional convolutional layers and the corresponding 3 layers of maximum pooling layers, the characteristics of the time sequence with time weight (gated time window) are extracted more deeply, the independence among variables is enhanced through the time sequence processed by the attention decoupling module, the variables are not influenced mutually, and better time sequence data is provided for a later self-adaptive learning time window gating module. The single-dimensional convolution proposed later can extract features on each variable due to the characteristics of the single-dimensional convolution, and the variables cannot be influenced mutually. The characteristic ensures that the process variable characteristics can be extracted, the relative independence of the variables is further enhanced, the multilayer convolution layers extract the characteristics of the process variables, the deep-level characteristics extracted by each variable are fused together through full connection, the finally output characteristic vector is the characteristic set of each process variable, the f-CaO content is measured by utilizing the characteristic set, and a specific model structure realization diagram combining the one-dimensional convolution is shown in figure 5.
The final structure of the model is shown in fig. 6 by using the concrete implementation of each method and the combination of each method.
4.3, training a soft measurement model;
selecting sample data in a cement database, dividing the sample data into a training set and a testing set, inputting the processed training data into a soft measurement model, and training the soft measurement model by using the sample data. And finally, extracting the characteristics among the process variables through convolution after the treatment of each proposed module, fusing the characteristics of each process variable, and measuring the content value of f-CaO in the cement clinker.
And 5, carrying out online measurement on the model.
And finally, the trained model stores parameters, and a soft measurement model for measuring the f-CaO content is constructed. And (3) preprocessing the data sample acquired in real time, inputting the preprocessed data sample into the constructed soft measurement model, and measuring the value of the f-CaO in real time through the model.
Specifically, data of normally produced cement in 3 months are collected to serve as an integral sample, after data preprocessing, the time series length of each sample is 60, the number of process variables is 7, the dimension of a single sample input by the model is 60 × 7, 2016 sample data are provided in total, 1792 samples in the 2016 sample data are selected to serve as training samples of the model, and the rest 224 samples serve as testing samples; obtaining a soft measurement model which is well trained and tested;
the trained and tested soft measurement model is applied to measurement of clinker free calcium in the actual cement production process, the error between the measured value and the actual value of the constructed soft measurement model is small, the measurement precision of the model is high, the large value and the small value of the f-CaO content can be effectively measured, the overall test condition is good, and the measurement result is shown in FIG. 7.
The above-described embodiments are merely illustrative of the preferred embodiments of the present invention, and not restrictive, and various changes and modifications may be made to the technical solution of the present invention by those skilled in the art without departing from the spirit of the present invention, which is defined by the claims.
Claims (5)
1. A soft measurement method for free calcium of cement clinker based on attention and window gating mechanism is characterized by comprising the following steps: aiming at the production characteristics of data coupling in the cement production process, the coupling relation among process variables is considered by utilizing an attention decoupling network model, the mutual influence among the variables is processed, and the decoupling of the process data is realized; aiming at the time delay change under different production conditions and the time delay of a process variable, a learnable window gating module is used for adaptively learning and gating the time period which has the most influence on the target variable, and the learnt time window is distributed with corresponding weights at different moments, so that the action duration of the process variable is more effectively concerned; respectively and independently extracting the characteristics of the single variables through a one-dimensional convolution network, and integrating the characteristic information of each variable to construct a soft measurement model meeting the actual requirements so as to realize the measurement of the free calcium of the cement clinker;
the measuring method specifically comprises the following steps:
step 1, acquiring process variables in a cement production industrial process and carrying out data preprocessing;
step 2, constructing an attention decoupling network model; the method specifically comprises the following steps:
2.1, inputting a process variable time sequence into an LSTM unit;
all time series samples are X = { X = ×) i ,i=1,2,3....M}∈R N×M The time series of the process variable isWhere N represents the length of the time series, M represents the number of process variables, x i A time series representing the ith process variable; inputting the time sequence of the process variable into the LSTM, extracting the time sequence characteristics by using the LSTM, and outputting the hidden state h in the model t The formula is as follows:
2.2, decoupling between variables is realized by applying an attention mechanism;
after the data preprocessing is finished, inputting the time sequence of each process variable into the LSTM to obtain the hidden state of each time sequence; process variables are mutually influenced and have strong coupling; by utilizing an attention mechanism, the influence of other process variables on the current variable is noticed, the weight distribution of the other variables on the current variable is found, and the finally output current variable contains the time sequence characteristics of other variables, so that the decoupling purpose is achieved; the decoupled attention configuration is as follows; mapping a current variable to a vector eta using full connections by combining the hidden state of the current variable with the hidden states of the remaining variables i :
The weights calculated are:
wherein w, u, v are learnable parameters,a hidden state representing a current variable; combining the hidden state of the current variable with the hidden states of the rest variables; />Multiplication of representative elements; the resulting weight vector is multiplied by the time series of the original input:
extracting time sequence characteristics of each process variable through LSTM, obtaining the influence of other variables on the current variable through an attention mechanism, distributing different weights, and finally outputting a vector v i Is the current variable containing the information of the rest variables;
step 3, a learnable window gating mechanism module is constructed; the method specifically comprises the following steps:
3.1, determining the center of a learnable time window;
the accuracy of the model is seriously influenced by time delay existing in cement production, the calcination of cement is a complex and variable chemical reaction process, the time delay between process variables and the time delay between the process variables and the content of free calcium are all the results of equipment, and the variables are interacted and mutually influenced; the time delay is difficult to determine due to the complex industrial process; aiming at variable time delay in the cement industry process, a learnable window gating mechanism, a time period for adaptively learning each process variable, and a soft measurement model fused with a window gating mechanism module, wherein the time delay characteristic in the cement production process and the time delay change under different production conditions are considered at the same time;
v i is the output of the attention decoupling module, representing the process variable for which decoupling is complete; c. C n The _wis a memory unit of the long-term and short-term memory network, which represents the long-term state in the network structure, the long-term state is a special hidden state, the long-term and short-term memory network can extract important information in a time sequence and store the characteristics in the long-term state, a calculation formula of an adaptive window center is provided according to the characteristics of the memory unit, and the information retained in the memory unit is extracted in the network learning process; the time window center is adaptively learned by utilizing the time sequence characteristics; the specific formula is as follows:
p i =T*sigmod(w p *c n _w+b p ) (7)
where i ∈ {1,2,3.. M }, M is the number of process variables, T represents the length of the sample sequence, p represents i The center of the window representing the ith variable, whose value is a number, w p And b p Weight and bias, respectively, for linear operations; calculating by using the formula, wherein each process variable can learn the corresponding time window center;
3.2, determining the length of the learnable time window;
h n w is the hidden state of the long-term and short-term memory network, the hidden state in the network is to extract partial information in the memory unit, which mainly depends on the state of the output gate, the value of the output gate in the network directly determines how much information the hidden state can extract in the memory unit, and the function is to effectively transfer the information in the memory unit to the hidden state; since the hidden state can learn in the network model and partially extract the information in the memory unit, its characteristics can be mapped to the length calculation of the time window; by using the length of the hidden state self-adaptive learning time window, the specific calculation formula is as follows:
where i ∈ {1,2,3.. M }, M is the number of process variables, T represents the length of the sample sequence, and l represents the length of the sample sequence i The length of the window representing the ith variable is a number, w l And b l Weights and offsets, respectively, for linear operations; calculating by using the formula, each process variable can learn the corresponding time window length;
3.3, determining a time window;
the determined window center and the determined window length are required to be applied to the time sequence, and the window center and the window length are combined to calculate a time period on the time sequence; and (3) gating by using an activation function and determining a time period which is most influenced by each variable on the f-CaO, wherein a specific calculation formula is as follows:
β i =sigmod(Relu(l i -abs(t-p i ))) (9)
wherein T belongs to {1,2,3.. T; t =60}, T representing each of the time seriesAt one time, abs denotes the absolute value of the parameter, p i And l i Respectively window center and window length, beta i Representing the assigned weight values in the ith process variable over a time series; by utilizing Relu and sigmod activation functions, the weight distributed to the time on the corresponding time period approaches to 1, and the rest times are all set to 0.5; different weight distribution gates the time period that each variable has been adaptively learned;
each process variable selects a corresponding time period by using the distributed time weight, the time weight is combined with the time series characteristics of the process variable, and the time information which has the most influence on the f-CaO is adaptively gated, wherein the formula is as follows:
wherein i is an element of {1,2,3.. M }, M is the number of process variables,denotes multiplication of elements, beta i Represents the weight value assigned in the ith process variable in time series>Output of the long-short term memory network, Y, representing the ith process variable i The final output of the adaptive learning window gating module represents the time window gated by each variable; />
Step 4, combining the single-dimensional convolution to construct a soft measurement model;
and 5, carrying out online measurement on the soft measurement model.
2. The method for soft measurement of free calcium in cement clinker based on attention and window gating mechanism as claimed in claim 1, wherein: in the step 1, the process variables which have the most influence on the target variables on the equipment at each position in the cement production process are selected, and the process variables are selected at a raw material inlet, a raw material preheater, a decomposing furnace, a rotary kiln and a grate cooler respectively, wherein the selected process variables comprise the temperature of the decomposing furnace, the feeding amount, the outlet temperature of a primary cylinder, the temperature of a kiln tail, the temperature of secondary air, the current of the kiln and the negative pressure of a kiln head.
3. The method for soft measurement of free calcium in cement clinker based on attention and window gating mechanism as claimed in claim 1, wherein: in the step 1, downloading process variable and clinker free calcium content data from a cement database for data preprocessing; the data preprocessing flow comprises abnormal value identification and processing and data standardization;
the specific method for identifying and processing the abnormal value comprises the following steps:
setting the ith process variable time series datan is the number of data, the abnormal value identification and processing adopt a 3 sigma criterion, and the specific calculation formula is as follows:
in the formula (I), the compound is shown in the specification,represents time-series data x i Represents the time series data x i The variance of (a);
the specific method for data standardization comprises the following steps:
dimensional differences exist among the collected process variable data, the numerical value and the measurement unit of each process variable are different, and the learning and nonlinear fitting capabilities of the deep learning model are seriously influenced by the differences; the standardization of data adopts 0-1 standardization to eliminate dimension difference, and the 0-1 standardization formula is as follows:
in the formula, x min Represents time-series data x i Minimum value of (a), x max Represents time-series data x i Is measured.
4. The method for soft measurement of free calcium in cement clinker based on attention and window gating mechanism as claimed in claim 1, wherein: in the step 4, the method specifically comprises the following steps:
4.1, constructing a single-dimensional convolution model;
the single-dimensional convolution is different from the two-dimensional convolution, the single-dimensional convolution only extracts features in one dimension, and the cement process data of the time sequence is processed by utilizing the characteristics of the single-dimensional convolution;
x in is input data of a single-dimensional convolution, x median Is the output of a single-dimensional convolution, x out Is the final output result after the maximal pooling, w is the weight size in the convolution kernel, b is the bias corresponding to the weight in the one-dimensional convolution layer, relu is a nonlinear activation function, and the simple formula of the one-dimensional convolution layer is as follows:
Relu=max(0,x) (11)
x median =Relu(w·x in +b) (12)
x out =MAX(x median ) (13)
4.2, constructing a final soft measurement model;
the time sequence passing through the window gating module and the time window of the adaptive gating process variable distribute weights to all moments on the time sequence of the process variable, the gated time period contains important time sequence characteristics, and deep extraction of the time sequence characteristics is realized by utilizing single-dimensional convolution; by increasing the number of the layers of the convolutional layers and continuously extracting the time sequence characteristics influencing free calcium more deeply, a soft measurement model meeting the requirements is developed;
the 3 layers of single-dimensional convolutional layers and the corresponding 3 layers of maximum pooling layers are utilized to extract the characteristics of the time sequence with time weight more deeply, the independence among all variables is enhanced through the time sequence processed by the attention decoupling module, the variables are not influenced mutually, and better time sequence data are provided for a later self-adaptive learning time window gating module; according to the single-dimensional convolution proposed later, due to the characteristics of the single-dimensional convolution, features can be extracted from each variable, and the variables cannot be influenced mutually; the characteristic ensures that the process variable characteristics are extracted, the relative independence of the variables is further enhanced, the multilayer convolution layers extract the characteristics of each process variable, the deep-level characteristics extracted by each variable are fused together through full connection, the finally output characteristic vector is the characteristic set of each process variable, and the characteristic set is used for measuring the f-CaO content;
4.3, training a soft measurement model;
selecting sample data in a cement database, dividing the sample data into a training set and a testing set, inputting the processed training data into a soft measurement model, and training the soft measurement model by using the sample data; and finally, extracting the characteristics among the process variables through convolution after the treatment of each proposed module, fusing the characteristics of each process variable, and measuring the content value of free calcium in the cement clinker.
5. The method for soft measurement of free calcium in cement clinker based on attention and window gating mechanism as claimed in claim 1, wherein: step 5, storing parameters of the finally trained soft measurement model, and constructing a soft measurement model for measuring the content of free calcium in the cement clinker; and (3) inputting a data sample acquired in real time into the constructed soft measurement model after data preprocessing, and measuring the value of free calcium of the cement clinker in real time through the model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210553703.1A CN114994294B (en) | 2022-05-20 | 2022-05-20 | Soft measurement method for free calcium of cement clinker based on attention and window gating mechanism |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202210553703.1A CN114994294B (en) | 2022-05-20 | 2022-05-20 | Soft measurement method for free calcium of cement clinker based on attention and window gating mechanism |
Publications (2)
Publication Number | Publication Date |
---|---|
CN114994294A CN114994294A (en) | 2022-09-02 |
CN114994294B true CN114994294B (en) | 2023-04-18 |
Family
ID=83026414
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202210553703.1A Active CN114994294B (en) | 2022-05-20 | 2022-05-20 | Soft measurement method for free calcium of cement clinker based on attention and window gating mechanism |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN114994294B (en) |
Family Cites Families (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109147878B (en) * | 2018-10-08 | 2021-10-15 | 燕山大学 | Soft measurement method for free calcium of cement clinker |
CN109919358B (en) * | 2019-01-31 | 2021-03-02 | 中国科学院软件研究所 | Real-time station flow prediction method based on neural network space-time attention mechanism |
CN110378044B (en) * | 2019-07-23 | 2021-06-11 | 燕山大学 | Multi-time scale convolution neural network soft measurement method based on attention mechanism |
CN111027752B (en) * | 2019-11-19 | 2022-06-21 | 浙江大学 | Crop yield estimation method based on deep spatiotemporal feature joint learning |
CN112365935B (en) * | 2020-10-20 | 2022-08-30 | 燕山大学 | Cement free calcium soft measurement method based on multi-scale depth network |
CN112765229B (en) * | 2020-12-25 | 2022-08-16 | 哈尔滨工程大学 | Air quality inference method based on multilayer attention mechanism |
CN113435666B (en) * | 2021-07-20 | 2023-11-07 | 山东大学 | Commercial vehicle fuel consumption prediction method and system based on vehicle running state |
CN113902202A (en) * | 2021-10-15 | 2022-01-07 | 南京工程学院 | Short-term load prediction model and method based on double attention mechanism and LSTM |
-
2022
- 2022-05-20 CN CN202210553703.1A patent/CN114994294B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN114994294A (en) | 2022-09-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107168402B (en) | Environment of chicken house temperature intelligent monitoring system based on CAN fieldbus | |
CN109685314B (en) | Non-intrusive load decomposition method and system based on long-term and short-term memory network | |
CN110378044B (en) | Multi-time scale convolution neural network soft measurement method based on attention mechanism | |
CN101630376A (en) | Soft-sensing modeling method and soft meter of multi-model neural network in biological fermentation process | |
CN112101480A (en) | Multivariate clustering and fused time sequence combined prediction method | |
CN109389314A (en) | A kind of quality hard measurement and monitoring method based on optimal neighbour's constituent analysis | |
CN110083096A (en) | A kind of silo intelligent monitoring method | |
CN113837464A (en) | Load prediction method of cogeneration boiler based on CNN-LSTM-Attention | |
CN108710964A (en) | A kind of prediction technique of Fuzzy time sequence aquaculture water quality environmental data | |
CN111444947A (en) | Bridge settlement intelligent detection system | |
CN114119273A (en) | Park comprehensive energy system non-invasive load decomposition method and system | |
CN116520772A (en) | Sheet part numerical control machining quality prediction method based on multitasking transfer learning | |
CN103279030B (en) | Dynamic soft measuring modeling method and device based on Bayesian frame | |
CN114994294B (en) | Soft measurement method for free calcium of cement clinker based on attention and window gating mechanism | |
CN108563201B (en) | DMAIC-driven small-batch part machining process quality improvement method | |
CN114239397A (en) | Soft measurement modeling method based on dynamic feature extraction and local weighted deep learning | |
Mei et al. | Soft sensor based on Gaussian process regression and its application in erythromycin fermentation process | |
CN111126827A (en) | Input-output accounting model construction method based on BP artificial neural network | |
CN116109019A (en) | Method and equipment for predicting daily gas production of gas well based on random forest | |
CN114943277A (en) | Method for identifying operation condition and predicting operation performance of comprehensive energy system | |
CN115423146A (en) | Self-adaptive runoff forecasting method based on multi-factor nearest neighbor sampling regression and support vector machine | |
CN109598083A (en) | A kind of glutamic acid fermentation process soft-measuring modeling method based on the twin support vector regression of drosophila algorithm optimization | |
CN113591384B (en) | Cement finished product specific surface area prediction method based on gating convolution network | |
Wang | A new variable selection method for soft sensor based on deep learning | |
Shuaishuai et al. | Prediction model of transpiration rate of strawberry in closed cultivation based on DBN-LSSVM algorithm |
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 |