CN112001527A - Industrial production process target data prediction method of multi-feature fusion deep neural network - Google Patents

Industrial production process target data prediction method of multi-feature fusion deep neural network Download PDF

Info

Publication number
CN112001527A
CN112001527A CN202010744152.8A CN202010744152A CN112001527A CN 112001527 A CN112001527 A CN 112001527A CN 202010744152 A CN202010744152 A CN 202010744152A CN 112001527 A CN112001527 A CN 112001527A
Authority
CN
China
Prior art keywords
data
neural network
deep
layer
feature fusion
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.)
Granted
Application number
CN202010744152.8A
Other languages
Chinese (zh)
Other versions
CN112001527B (en
Inventor
曾九孙
欧阳航
丁克勤
蔡晋辉
姚燕
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Jiliang University
Original Assignee
China Jiliang University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by China Jiliang University filed Critical China Jiliang University
Priority to CN202010744152.8A priority Critical patent/CN112001527B/en
Publication of CN112001527A publication Critical patent/CN112001527A/en
Application granted granted Critical
Publication of CN112001527B publication Critical patent/CN112001527B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/04—Manufacturing
    • 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
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • Strategic Management (AREA)
  • Evolutionary Computation (AREA)
  • Human Resources & Organizations (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Economics (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Tourism & Hospitality (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Operations Research (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Game Theory and Decision Science (AREA)
  • Manufacturing & Machinery (AREA)
  • Development Economics (AREA)
  • Quality & Reliability (AREA)
  • Primary Health Care (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Feedback Control In General (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a method for predicting target data in an industrial production process by a multi-feature fusion deep neural network. Acquiring other variable time sequence data related to the key variable in the industrial equipment by using a sensor through equal-interval sampling, and performing predictive analysis on the time sequence data of the key variable in the process industry; inputting the data into a pre-designed and constructed deep convolutional neural network for training; dividing historical data of key variables according to time steps and inputting the segmented historical data into a deep gating cyclic neural network for learning; by utilizing a multi-feature fusion method, the output features obtained by the two networks are fused and then input into a full connection layer, and the network parameters are optimized through back propagation, so that the prediction precision is improved. The method provides reliable and effective target variable parameter prediction for process monitoring in industrial production, and relieves the hysteresis of measurement of key variables such as silicon content in molten iron and the like in industrial production.

Description

Industrial production process target data prediction method of multi-feature fusion deep neural network
Technical Field
The invention belongs to an industrial parameter prediction method in the field of process industrial production, and particularly relates to an industrial production process key parameter prediction method based on multi-feature fusion.
Background
The existing process industrial system gradually goes to the intelligentization, integration and higher automation degree, the function of the whole industrial system is more and more perfect, and therefore the correlation among all variables in the system is more and more compact. Meanwhile, with the appearance of various sensors, more and more parameters can be obtained in the process industry, and a data source is provided for industrial big data processing. However, the measurement of key parameters in the process industry, such as the silicon content in molten iron in blast furnace iron making, still has great hysteresis, and such key parameters are often obtained only by soft measurement methods, and the traditional direct measurement method has great hysteresis.
The current and future states of the whole industrial system can be speculated by predicting the key parameters in the process industry in advance, so that the practitioner can find out which equipment in the industrial system has problems in time conveniently, and the operation is convenient to be carried out in advance, the system is stabilized, and industrial accidents are prevented. The realization of the prediction of the key parameters of the process industry has great significance for stabilizing the whole industrial system and the industrial production safety.
For monitoring key parameters of an industrial process, two common solutions are provided. One method is to use a novel instrument with strong real-time performance and higher accuracy for direct measurement through continuous research and development. The second method is to measure other variables associated with the key parameter to be detected, obtain the nonlinear relationship between the variables by statistical or neural network methods to calculate the key variable, and indirectly obtain the target parameter. Compared with the second method, the first method has the problems of high equipment cost, long research and development period, long analysis period, serious hysteresis and the like, so that an operator cannot perform the analysisThe real-time state of the industrial system can be known in time, so that the production quality is reduced, and the energy waste is serious. The second method, also called soft measurement method, directly measures variables easy to monitor, and establishes a prediction model by learning the potential relationship between the target variable and the known variables[2]Thus, real-time prediction and estimation of the target variable can be achieved, and compared with the first method, the method has the advantages of short development period and low cost.
Chinese utility model patent No. CN110400007A discloses "a molten iron quality prediction method based on an improved gated recurrent neural network", which reduces the complexity of the model by combining the reset gate and the update gate of the gated recurrent neural network into a single disposal gate. However, the gated cyclic neural network used in this method is not good at capturing the data dimension characteristics, so the data dimension information is easily lost, and the prediction effect is not ideal for more complicated process industry problems and working conditions. The invention introduces a neural network structure which is good at learning data dimensionality through a multi-feature fusion technology, so that the problem that data dimensionality information is lost can be solved, and finally, a predicted value is obtained by using the time dimensionality feature and the data dimensionality feature at the same time, so that the prediction precision of the key parameter is further improved.
Disclosure of Invention
In order to solve the technical problems in the background art, the invention provides a flow industry key parameter prediction method based on a multi-feature fusion deep neural network, which can perform data mining on an original time sequence signal in the flow industry, learn the relation between a target key parameter and an easily-measured parameter and predict the target key parameter in real time.
The method is suitable for monitoring the state of a real-time system in the process industry, particularly the target key parameters have strong sequence correlation, the measured state parameters of the industrial system have strong inertia, reliable and effective target variable parameter prediction is provided for process monitoring in industrial production, and the hysteresis of measurement of key variables such as silicon content in molten iron and the like in the industrial production is relieved.
The technical scheme adopted by the invention is that the method specifically comprises the following steps:
the method generally comprises the following steps:
step 1, sampling an industrial production process under normal working conditions at different times at equal time intervals to obtain production state data containing target variable parameters, and preprocessing the production state data to obtain a data set;
and 2, inputting the data set into a multi-feature fusion deep neural network for processing to obtain a prediction result of the target variable parameter.
The method specifically comprises the following steps:
step 1, sampling an industrial production process under normal working conditions at different times at equal time intervals to obtain production state data, wherein the production state data are multivariable parameters obtained through detection of different sensors, dividing the collected production state data according to time and according to the sensors, and inputting time sequence data and sensor classification data into two data sets of a multi-feature fusion deep neural network comprising two neural networks as follow-up data; the signal time series data of the target variable can be divided into sub-data with the same length according to time sequence, and each sub-data comprises the same number of sampling points.
The production state data refers to variable parameters obtained through sensor detection, and comprises data of target variable parameters and reference variable parameters, the target variable parameters are variable parameters which cannot be directly and accurately detected through a sensor in real time and can only be obtained through sensor lag detection, and the reference variable parameters are variable parameters which can be directly and accurately detected through the sensor in real time;
the target data is a target variable parameter which cannot be directly and accurately detected by a sensor in real time and can only be obtained by lagging detection of the sensor, and is also a key parameter in the industrial production process. The target variable parameters generally refer to key parameters such as silicon content of molten iron, butane concentration at the bottom of a well and the like which are difficult to directly measure in the process industry.
Each sensor detects a variable parameter, and the data obtained by different sensors at the same time are different variable parameters.
Step 2, the multi-feature fusion deep neural network mainly comprises a deep convolutional neural network, a deep gating cyclic neural network and a multi-feature fusion layer, and the output ends of the deep convolutional neural network and the deep gating cyclic neural network are connected to the multi-feature fusion layer;
2.1, the time sequence data are variable parameters directly obtained by detection of each sensor at the same time, reference variable parameters in the time sequence data are used as input, target variable parameters obtained by lagging detection of the sensors in the time sequence data are used as output and input into a deep convolutional neural network together for processing, the relation between different variable parameters at the same time is obtained through learning, and the relation between the target variable parameters and the variable parameters of the production state data is obtained;
2.2, setting a time step, carrying out sliding window sampling on a part corresponding to a target variable parameter in the sensor classification data according to the time step, inputting the sampled data into a single-step predicted deep gating cyclic neural network, learning the relation between different moments of the target variable parameter, learning the periodicity and the trend of a time sequence, and obtaining the relation between historical multiple moment data of the target variable parameter and current moment data; the target variable is the sensor measurement parameter measured at the actual lag.
2.3, inputting the output prediction data of the deep convolutional neural network before the last full-connection layer and the output prediction data of the deep gating cyclic neural network before the last full-connection layer into the multi-feature fusion layer together, and fusing by using multi-feature fusion, wherein the method specifically comprises the following steps: adding and connecting the two output prediction data in parallel by using an adding layer to obtain information after merging characteristics, flattening the information into one-dimensional data by flattening the layers, inputting the one-dimensional data into a final full-connection layer, and outputting a prediction value result of a target variable parameter; therefore, high-dimensional feature data with the same number are merged, the features of the data dimension and the features of the time dimension are all mapped into a high-dimensional feature space, and the data are finally input into a fully-connected neural network to output a prediction result of a final target.
The output prediction data comprises the number of windows, the size of the windows and the characteristic dimension, and a plurality of prediction values correspond to a plurality of window numbers.
The output prediction data of the deep convolutional neural network before the last fully connected layer and the output prediction data of the deep gated circular neural network before the last fully connected layer are high-dimensional feature data, and the high-dimensional feature data are obtained by mapping the original data to a higher-dimensional space.
Processing by adopting the steps to obtain two data sets, wherein each data vector of the data set input into the deep convolutional neural network comprises a reference variable parameter except the target variable parameter, and each data vector of the data set input into the deep gated cyclic neural network comprises historical data of a fixed time step before the target variable parameter needing to be predicted; and respectively inputting the data set into a deep convolutional neural network and a deep gated cyclic network for feature extraction. After passing through the two networks, the two parts of training samples are mapped to the same-dimension feature space, the two parts of features are combined, and the high-dimension data is flattened into one dimension through a Flatten layer and is input to a full connection layer for outputting a target variable result.
2.4, inputting the fusion features output by the multi-feature fusion layer into the last full-connection layer, training a deep convolutional neural network and a deep gating cycle network by using a back propagation algorithm, adjusting weight parameters of the deep convolutional neural network and the deep gating cycle network, realizing an approximation process of a predicted value and a true value, and optimizing to obtain the trained deep convolutional neural network and the trained deep gating cycle network;
and 2.5, combining the trained deep convolution neural network and the deep gating circulation network with the multi-feature fusion layer to form the trained multi-feature fusion deep neural network, inputting the reference variable parameter data of the condition to be tested into the trained multi-feature fusion deep neural network, and outputting and predicting to obtain the target variable parameter.
The deep convolutional neural network captures data dimension characteristic data of the multivariate time sequence data, and output prediction data of the deep convolutional neural network before the last fully-connected layer is the data dimension characteristic data; and the time dimension characteristic data of the target data is captured by the deep gating circulation network, and the output prediction data of the deep gating circulation neural network before the last full connection layer is the time dimension characteristic data.
The invention relates to periodic signal time sequence data collected by a sensor in industrial production, which comprises important variables in a blast furnace ironmaking process such as blast furnace coal injection quantity, oxygen content, pressure difference in a furnace and the like.
The deep convolution neural network is mainly composed of a plurality of convolution blocks and a maximum pooling layer which are connected in sequence, each convolution block is composed of a plurality of convolution layers which are connected in sequence, time sequence data are normalized firstly to eliminate dimension relations among all subsequences, then all convolution blocks are processed through time domain convolution, convolution calculation is carried out through a learnt convolution kernel, features are obtained through an activation function and are mapped into a feature space, and a calculation formula is as follows:
Ci＝f(Wi*Ci-1+bi)
in the formula, WiRepresenting the sharing weight of the ith volume block, and the characteristic C of the ith-1 volume blocki-1Convolution, i.e. the sign of the convolution operation, biIs the offset vector of the ith convolution block, f () represents the nonlinear activation function; obtaining a characteristic C of an ith volume block by a non-linear activation function relui；
And after time domain convolution processing of each convolution block, the output of the maximum pooling layer is finally the data dimension characteristic data of the target characteristic space.
The deep gated cyclic neural network is formed by three gated cyclic layers which are connected in sequence to form a three-layer gated cyclic network.
The deep gating cyclic neural network comprises three layers of processing modules: the first processing module comprises a first gating circulation layer, and the first gating circulation layer is transmitted to the second processing module through the reset gate and the update gate in sequence; the second processing module comprises a second gate control circulation layer and a second Dropout layer, and the second gate control circulation layer is transmitted to the third processing module through the reset gate and the update gate in sequence; the third processing module comprises a third gating circulation layer, and the third gating circulation layer is transmitted to the mapping layer through the reset gate and the update gate in sequence; the first gated circulation layer, the second gated circulation layer and the third gated circulation layer are mainly composed of sequential gated circulation units. The mapping layer serves as an intermediate output layer to output high-dimensional characteristics of the time dimension of the original data.
In said 2.4, against; the deep convolutional neural network and the deep gated cyclic neural network establish the following loss functions:
Figure BDA0002607758600000051
wherein Loss represents a Loss value between a predicted value and a true value, m represents the number of sampling time points in one sample, and yiThe (i) th real value is represented,
Figure BDA0002607758600000052
representing the ith predicted value;
the predicted value is data output by network prediction, and the true value is a target variable parameter measured by actual lag.
The input data is taken as a sample, the weight derivative value of each convolution layer in each convolution block in the deep convolutional neural network is taken as the gradient of the sample, and the weight derivative value of a gating cycle unit in a first gating cycle layer in the deep gating cyclic neural network is taken as the gradient of the sample; network parameters of the deep convolutional neural network and the deep gated cyclic neural network are updated through the following gradients to adjust weights in the network, an optimizer is an Adam optimizer, and the gradient updating method comprises the following steps:
Figure BDA0002607758600000053
Figure BDA0002607758600000054
Figure BDA0002607758600000055
Figure BDA0002607758600000056
Figure BDA0002607758600000057
where, is the velocity/momentum at time t,
Figure BDA0002607758600000058
is the average gradient of all samples, is the cumulative squared gradient at time t, is the exponential decay rate of the first order moment estimate, is the exponential decay rate of the second order moment estimate,
Figure BDA0002607758600000059
in order to obtain a speed after the deviation correction,
Figure BDA00026077586000000510
the cumulative square gradient after the offset correction is a network parameter at time t, which is a non-zero parameter, particularly, a very small number avoiding the case where the denominator is 0, which is a learning rate, indicates a Hadamard (Hadamard) product.
When the loss function tends to be stable and does not decrease any more, the performance of the whole network can be tested by using the test set data, and the test set data can be used for predicting the key parameter variables of the target variable parameters in advance in the actual production process.
In the specific implementation, the performance of the multi-feature fusion deep neural network in predicting the process industry key parameters which are difficult to measure directly is tested by using the known multivariate signal time sequence data, the prediction result is counted, and the prediction accuracy is calculated.
Therefore, the method realizes the prediction of the key parameters of the process industry by acquiring the multivariate time sequence signals, training the complete model in an off-line manner and arranging the model in the on-line measuring equipment. The specific operation flow is as follows, firstly, the pair variable time sequence signals under normal production are collected, the process variable signals except the target variable are processed by data normalization to eliminate the dimensional relation among all subsequences, and then the original data are divided into data segments with equal sampling time intervals according to fixed time steps. And respectively inputting the two groups of sample data into a convolutional neural network and a gated cyclic network for training, acquiring hidden state data with fused data dimensions and time dimension characteristics, and then inputting the hidden state data containing the fused characteristics into a full-connection layer to output a result of a specified dimension, namely a target variable prediction result. And the error between the predicted value and the true value is reversely transmitted to the two networks through a back propagation algorithm for optimizing the network parameters of the offline model, so that the prediction progress is improved, and the flow industry key parameter prediction is realized.
The method carries out feature extraction on big data generated by process industry seeds and fuses time dimension and data dimension information to realize the prediction of key parameters of the process industry, and can realize the prediction of key parameters in process industries such as blast furnace iron making, debutanizer industrial refining and the like, including but not limited to the silicon content of molten iron and the concentration of butane at the bottom of a well.
The multivariate time sequence big data are decomposed and are simultaneously input into two different neural networks so as to capture the potential relation among variables in the original data and the time dependence of a target variable; and meanwhile, the high-dimensional features obtained by the deep convolutional neural network and the deep gated cyclic neural network are fused, so that the high accuracy of the prediction result is realized by utilizing a multi-feature fusion technology.
The multi-feature fusion is to combine the mapping results of the deep convolutional neural network and the deep gated cyclic neural network into the same high-dimensional feature space, input the information in the feature space into the last full-connection layer to obtain a predicted value, and reversely propagate errors to adjust the parameters of the two networks, so that the loss function is as minimum as possible.
The process industrial parameters refer to periodic signal time sequence data acquired by a sensor in industrial production, and include but are not limited to important variables in a blast furnace ironmaking process such as blast furnace coal injection quantity, oxygen content, top layer pressure, furnace top temperature, top layer coal gas component content and the like.
The time dependence refers to the degree of influence of the current value of the target predictive variable on the previous system state, and is generally divided into long-term dependence and short-term dependence.
In specific implementation, collected multivariate time sequence data are subjected to default value completion, variables with low correlation degree with target variables are deleted, and time steps are set to divide original data into subdata with the same length according to a specific industrial system.
The invention can give consideration to the linear and nonlinear relation between time sequence information and variables in multivariable data and multi-layer extraction characteristics to avoid dimension cursing problem which is easy to occur when big data is analyzed, and simultaneously keeps and improves the precision of target variable prediction, namely, the invention has good accuracy while realizing data calculation simplification.
Compared with the traditional time sequence prediction method, the method adopts the deep learning of the convolutional neural network and the gating cycle to spontaneously extract a large amount of variable relation and time characteristic information from a given sample, and does not need to rely on excessive flow industrial production information. The invention only provides a training sample set and a testing sample set which are sufficient for the model, the deep learning model can effectively extract the characteristic knowledge contained in the sample, and automatically updates the weight so as to obtain the influence degree of other variables on the key variable, meanwhile, the trend and the periodicity of the key variable sequence are learned from the past information, and finally, the predicted value of the current state of the key variable is obtained through the output layer. And continuously adjusting and optimizing the model parameters through the reverse transmission error so as to continuously improve the prediction accuracy and finally finish the prediction of key variables in the process industry.
Compared with the prior art, the invention has the following beneficial effects:
1. by normalizing the information of other variables except the target variable, the off-line training time of the network is reduced, and the model is updated more quickly.
2. By the multi-feature fusion technology, the convolution network which is good at processing data dimension information can exert the greatest advantage, the defect of the gated circulation network which is good at processing time dimension information is overcome, and the prediction accuracy of the final network is improved.
3. Through the deep neural network structure, the convolutional neural network can learn the nonlinear relation between other variables and a target variable, the gated cyclic neural network can learn the nonlinear relation between the state at the past moment and the state at the current moment, and finally the capability of the model for processing the nonlinear problem is improved.
4. Through a back propagation algorithm, the weight parameters of the two networks are updated simultaneously by using the errors of the predicted value and the true value, so that the loss value is reduced, and the autonomous learning of the model is realized.
In summary, the invention realizes the combination of data dimension and time dimension information through multi-feature fusion, thereby improving the prediction capability of a neural network, reducing the cost of the process industry key parameter measurement, realizing the process industry key parameter prediction, and providing reliable and effective technical support for the process industry on-line monitoring.
Drawings
FIG. 1 is a schematic diagram of a blast furnace ironmaking structure according to the present invention;
FIG. 2 is a block diagram of a multi-feature fusion deep neural network of the present invention;
FIG. 3 is a diagram of the deep convolutional neural network of the present invention;
FIG. 4 is a diagram of a deep gated recurrent neural network of the present invention;
FIG. 5 is a graph of model training set loss values versus test set loss value variations in accordance with the present invention;
FIG. 6 shows the result of predicting the silicon content in molten iron in a blast furnace according to the present invention.
Detailed Description
The invention is further illustrated by the following figures and examples.
The embodiment and the implementation process of the complete method according to the invention are as follows:
the process industrial data specifically implemented by the method is measured by a blast furnace ironmaking process industrial production line of a Baotong iron and steel plant. Taking the process of predicting the silicon content of the molten iron as an example, the prediction of key parameters of the process industry is described in detail based on the specific processing process of the data.
In specific implementation, the method regularly uses new fault-free data to update the model parameters, so that the failure of the model due to accumulated errors is avoided, and the prediction precision is reduced. Abnormal data caused by artificial factors and sensor failure in the blast furnace ironmaking process can influence the accuracy of model prediction, then the quality of the training data is evaluated, and the abnormal data and irrelevant variables are eliminated.
The process industrial data specifically implemented by the invention is data measured by a blast furnace ironmaking process industrial production line of a Baotong iron and steel plant, but the process industrial key parameter prediction can be carried out by the method as long as continuous variable time sequence data related to the target key parameter can be measured in a process industrial system. The experimental data set consists of the variables of coal injection quantity, air intake quantity, air permeability, air temperature, feeding speed, furnace top pressure difference and oxygen blowing quantity. The quality of the produced molten iron was relatively stable during the test, with silicon content varying between 0.4 and 0.6, and individual cases were outside this range.
As shown in FIG. 1, the industrial process performs iron making from iron ore, in which molten iron contains a small amount of impurities, and in which silicon content is an important parameter for measuring the temperature in a blast furnace.
The industrial process of the process is as follows: iron-containing raw materials (iron ore), fuels (coal powder and coke) and other auxiliary raw materials are loaded into the top of a blast furnace according to a certain proportion, and hot air from a hot blast stove is blown from an air inlet at the lower part of the blast furnace for assisting the combustion of the coke. Carbon reacts with high-temperature air at high temperature to generate CO and H2. As the raw material and the fuel descend in the furnace, the ascending coal gas contacts with the raw material, heat transfer, reduction, melting and decarburization are sequentially carried out to generate molten pig iron, and other components in the raw material are combined with a product after the fuel reacts to form slag. Two byproducts, blast furnace gas and slag, are produced simultaneously in the whole process.
The feeding speed, the coal injection quantity, the blast quantity, the oxygen blowing quantity and the like are controllable variables, the variables such as the blast temperature, the pressure difference in the furnace and the like are measured by a sensor, the values of the controllable variables are stored at equal time intervals, the variable values related to the silicon content of the molten iron in the process industry are sampled at equal time intervals, and all measurement data files are stored in a csv format.
Based on the above, the present embodiment utilizes 7 types of blast furnace state sample data related to the silicon content of molten iron in common, and the variable number and specific variable name are shown in table 1.
TABLE 1
Figure BDA0002607758600000081
Figure BDA0002607758600000091
Wherein represents a target variable parameter, uiRepresenting other variable parameters than the target variable. As shown in fig. 2, in the implementation process, multivariate data are subjected to data segmentation, and original data are divided into two data sets, so that the input requirements of two deep neural networks can be conveniently adapted, and the process for predicting the process industrial key variables comprises the following steps:
step 1, acquiring time series signal data which has high correlation with target variable parameters in a process industrial system by using a sensor through interval sampling as a training sample;
and 2, processing the time series signal data to obtain training samples which can be input into a deep convolutional neural network and a deep gating cyclic network. Normalizing the process variable signal time sequence data except the target variable to eliminate the dimensional relation among the subsequences; dividing the signal time series data of the target variable into sub-data with the same length according to time sequence, wherein each sub-data comprises the same number of sampling points, as shown in FIG. 2;
step 3, processing the multivariate signal time sequence data of the training samples by adopting the step 2 to obtain two samples, wherein each sample of the training samples input into the deep convolutional neural network comprises other variable values except the key parameters, and the structure of the deep convolutional neural network is shown in figure 3; inputting training samples of a deep gated cyclic neural network, wherein each sample comprises historical data of a fixed time step before a target key parameter to be predicted, the structure of the deep gated cyclic neural network is shown in figure 4, and two training samples are respectively input into a deep convolutional neural network and the deep gated cyclic network for feature extraction; and finally, fusing information output by the two neural networks to the same feature space through a coordinate layer, merging the two parts of features, flattening the high-dimensional data into one dimension through a Flatten layer, inputting the one dimension into a full connection layer, and outputting a target variable result.
And 4, after the characteristics of the deep convolutional neural network and the deep gated cyclic neural network are combined, inputting the hidden state into an output layer of the fully connected network to output a result, simultaneously adjusting parameters in the convolutional neural network and the gated cyclic neural network according to a loss value returned in the reverse transfer process, judging the quality of the network training degree according to a result obtained by the loss function, and controlling the training to stop. The model training results for the data of the blast furnace ironmaking process are shown in fig. 5 and 6. The obtained model is used for an intelligent online monitoring device, the process industry key parameters are predicted by utilizing the information measured in real time, and an operator can adjust the process industry system in advance.
For comparison, model evaluation is performed on the same data set by respectively adopting a traditional Deep Belief Network (DBN), a long-and-short-term memory neural network (LSTM), a single Convolutional Neural Network (CNN) and a gated cyclic neural network (GRU), the difference between a predicted value and a real value of the model is evaluated by Mean Square Error (MSE), and a decision coefficient R is used2To evaluate the fit of the model. The results of the model performance evaluation are shown in table 2.
TABLE 2
Figure BDA0002607758600000101
From table 2, it can be seen that the remaining methods, either the degree of fit or the mean square error, are less good than the method proposed by the present invention,and the method proposed by the invention determines the coefficient R2Much higher than other methods. Therefore, the reliability and the practicability of the calculation method provided by the invention are verified. Target variable parameter prediction of the DBN network and the process industry is also commonly used, but the method provided by the invention is improved by 34.6% in the data relative to the DBN network. LSTM networks are more complex than GRU networks, but do not perform as well as GRU networks in process industry problems. When the GRU network or the CNN network is used for training and testing independently, due to the defects of the network structures of the GRU network and the CNN network, the GRU network or the CNN network cannot sense the data dimension characteristic and the time dimension characteristic simultaneously, the obtained fitting result is not ideal, and after the concatemate layer structure is added, the data dimension characteristic and the time dimension characteristic can be captured simultaneously, so that the performance of the whole network model is improved, and the more complex process industrial condition can be processed conveniently.
The implementation shows that the method provided by the invention can be used for predicting the key parameters in the process industry based on the flow industry key parameter prediction method of the multi-feature fusion deep neural network, and has higher accuracy. This demonstrates the feasibility of online health monitoring and key variable prediction in such process industries.

Claims (6)

1. A target data prediction method for an industrial production process of a multi-feature fusion deep neural network is characterized by comprising the following steps: the method specifically comprises the following steps:
step 1, sampling an industrial production process under normal working conditions at different times at equal time intervals to obtain production state data containing target variable parameters, and preprocessing the production state data to obtain a data set;
and 2, inputting the data set into a multi-feature fusion deep neural network for processing to obtain a prediction result of the target variable parameter.
2. The method for predicting the target data of the industrial production process of the multi-feature fusion deep neural network according to claim 1, wherein the method comprises the following steps: the method specifically comprises the following steps:
step 1, sampling an industrial production process under normal working conditions at different times at equal time intervals to obtain production state data, wherein the production state data are multivariable parameters obtained through detection of different sensors, dividing the collected production state data according to time and according to the sensors, and inputting time sequence data and sensor classification data into two data sets of a multi-feature fusion deep neural network comprising two neural networks as follow-up data; the production state data refers to variable parameters obtained through sensor detection, and comprises data of target variable parameters and reference variable parameters, the target variable parameters are variable parameters which cannot be directly and accurately detected through a sensor in real time and can only be obtained through sensor lag detection, and the reference variable parameters are variable parameters which can be directly and accurately detected through the sensor in real time;
step 2, the multi-feature fusion deep neural network mainly comprises a deep convolutional neural network, a deep gating cyclic neural network and a multi-feature fusion layer, and the output ends of the deep convolutional neural network and the deep gating cyclic neural network are connected to the multi-feature fusion layer;
2.1, the time sequence data are variable parameters directly obtained by detection of each sensor at the same time, reference variable parameters in the time sequence data are used as input, target variable parameters in the time sequence data are used as output and are input into a deep convolutional neural network together for processing, the relation between different variable parameters at the same time is obtained through learning, and the relation between the target variable parameters and the variable parameters of the production state data is obtained;
2.2, setting a time step, carrying out sliding window sampling on a part corresponding to a target variable parameter in the sensor classification data according to the time step, inputting the sampled data into a single-step predicted deep gating cyclic neural network, and learning the relation between different moments of the target variable parameter to obtain the relation between historical multiple moment data of the target variable parameter and current moment data;
2.3, inputting the output prediction data of the deep convolutional neural network before the last full-connection layer and the output prediction data of the deep gating cyclic neural network before the last full-connection layer into the multi-feature fusion layer together, and fusing by using multi-feature fusion, wherein the method specifically comprises the following steps: adding and connecting the two output prediction data in parallel by using an adding layer to obtain information with combined characteristics, flattening the information into one-dimensional data by flattening the data, and inputting the one-dimensional data into a final full-connection layer;
2.4, inputting the fusion features output by the multi-feature fusion layer into the last full-connection layer, training a deep convolutional neural network and a deep gating cycle network by using a back propagation algorithm, adjusting the weight parameters of the deep convolutional neural network and the deep gating cycle network, and optimizing to obtain the trained deep convolutional neural network and the trained deep gating cycle network;
and 2.5, combining the trained deep convolution neural network and the deep gating circulation network with the multi-feature fusion layer to form the trained multi-feature fusion deep neural network, inputting the reference variable parameter data of the condition to be tested into the trained multi-feature fusion deep neural network, and outputting and predicting to obtain the target variable parameter.
3. The method for predicting the target data of the industrial production process of the multi-feature fusion deep neural network as claimed in claim 2, wherein the method comprises the following steps: the deep convolutional neural network captures data dimension characteristic data of the multivariate time sequence data, and output prediction data of the deep convolutional neural network before the last fully-connected layer is the data dimension characteristic data; and the time dimension characteristic data of the target data is captured by the deep gating circulation network, and the output prediction data of the deep gating circulation neural network before the last full connection layer is the time dimension characteristic data.
4. The method for predicting the target data of the industrial production process of the multi-feature fusion deep neural network as claimed in claim 2, wherein the method comprises the following steps: the deep convolutional neural network is mainly composed of a plurality of convolutional blocks and a maximum pooling layer which are connected in sequence, each convolutional block is composed of a plurality of convolutional layers which are connected in sequence, time sequence data are normalized firstly and then processed by each convolutional block of time domain convolution, features are obtained by utilizing an activation function and are mapped into a feature space, and the calculation formula is as follows:
Ci＝f(Wi*Ci-1+bi)
in the formula, WiRepresenting the sharing weight of the ith volume block, and the characteristic C of the ith-1 volume blocki-1Convolution, i.e. the sign of the convolution operation, biIs the offset vector of the ith convolution block, f () represents the nonlinear activation function; obtaining a characteristic C of an ith volume block by a non-linear activation function relui；
And after time domain convolution processing of each convolution block, the output of the maximum pooling layer is finally the data dimension characteristic data of the target characteristic space.
5. The method for predicting the target data of the industrial production process of the multi-feature fusion deep neural network as claimed in claim 2, wherein the method comprises the following steps: the deep gated cyclic neural network is formed by three gated cyclic layers which are connected in sequence to form a three-layer gated cyclic network.
6. The method for predicting the target data of the industrial production process of the multi-feature fusion deep neural network according to claim 5, wherein the method comprises the following steps: the deep gating cyclic neural network comprises three layers of processing modules: the first processing module comprises a first gating circulation layer, and the first gating circulation layer is transmitted to the second processing module through the reset gate and the update gate in sequence; the second processing module comprises a second gate control circulation layer and a second Dropout layer, and the second gate control circulation layer is transmitted to the third processing module through the reset gate and the update gate in sequence; the third processing module comprises a third gating circulation layer, and the third gating circulation layer is transmitted to the mapping layer through the reset gate and the update gate in sequence; the first gated circulation layer, the second gated circulation layer and the third gated circulation layer are mainly composed of sequential gated circulation units.
CN202010744152.8A 2020-07-29 2020-07-29 Industrial production process target data prediction method of multi-feature fusion depth neural network Active CN112001527B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010744152.8A CN112001527B (en) 2020-07-29 2020-07-29 Industrial production process target data prediction method of multi-feature fusion depth neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010744152.8A CN112001527B (en) 2020-07-29 2020-07-29 Industrial production process target data prediction method of multi-feature fusion depth neural network

Publications (2)

Publication Number Publication Date
CN112001527A true CN112001527A (en) 2020-11-27
CN112001527B CN112001527B (en) 2024-01-30

Family

ID=73463963

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010744152.8A Active CN112001527B (en) 2020-07-29 2020-07-29 Industrial production process target data prediction method of multi-feature fusion depth neural network

Country Status (1)

Country Link
CN (1) CN112001527B (en)

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112985096A (en) * 2021-03-05 2021-06-18 广州东兆信息科技有限公司 Kiln ceramic quality monitoring system and method based on Internet of things cloud platform
CN113009906A (en) * 2021-03-04 2021-06-22 青岛弯弓信息技术有限公司 Big data prediction analysis method and system based on industrial Internet
CN113360851A (en) * 2021-06-22 2021-09-07 北京邮电大学 Industrial flow line production state detection method based on Gap-loss function
CN113642234A (en) * 2021-08-09 2021-11-12 贵州电网有限责任公司 Power grid icing prediction method based on multi-source characteristic time convolution deep learning
CN114015825A (en) * 2021-11-09 2022-02-08 上海交通大学 Method for monitoring abnormal state of blast furnace heat load based on attention mechanism
CN114216349A (en) * 2021-12-06 2022-03-22 浙江大学 Sintering end point forecasting method based on coding and decoding network
CN114332474A (en) * 2021-11-25 2022-04-12 中国计量大学 Method for predicting maintenance time of key part of aircraft engine
CN114519610A (en) * 2022-02-16 2022-05-20 支付宝(杭州)信息技术有限公司 Information prediction method and device
CN114638117A (en) * 2022-03-29 2022-06-17 施昆宏 Data processing method and system based on artificial intelligence and cloud platform
CN115457367A (en) * 2022-09-22 2022-12-09 淮阴工学院 Lightweight target detection method and system based on Light-Net
CN115713987A (en) * 2022-11-17 2023-02-24 广州瑞博新材料技术研究有限公司 Polycaprolactone test data analysis method and system
CN116612836A (en) * 2023-07-19 2023-08-18 福建德尔科技股份有限公司 Tail gas amount prediction method and system for trifluoromethane production
CN117035559A (en) * 2023-10-08 2023-11-10 国网江苏省电力有限公司电力科学研究院 Electrical equipment multi-parameter transmitter simulation installation evaluation method and system

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018156314A1 (en) * 2017-02-24 2018-08-30 Qualcomm Incorporated Method and apparatus for multi-dimensional sequence prediction
CN108985376A (en) * 2018-07-17 2018-12-11 东北大学 It is a kind of based on convolution-Recognition with Recurrent Neural Network rotary kiln sequence operating mode's switch method
CN111222798A (en) * 2020-01-13 2020-06-02 湖南师范大学 Soft measurement method for key indexes of complex industrial process
US20200175015A1 (en) * 2018-11-29 2020-06-04 Koninklijke Philips N.V. Crf-based span prediction for fine machine learning comprehension
CN111325112A (en) * 2020-01-31 2020-06-23 贵州大学 Cutter wear state monitoring method based on depth gate control circulation unit neural network

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2018156314A1 (en) * 2017-02-24 2018-08-30 Qualcomm Incorporated Method and apparatus for multi-dimensional sequence prediction
CN108985376A (en) * 2018-07-17 2018-12-11 东北大学 It is a kind of based on convolution-Recognition with Recurrent Neural Network rotary kiln sequence operating mode's switch method
US20200175015A1 (en) * 2018-11-29 2020-06-04 Koninklijke Philips N.V. Crf-based span prediction for fine machine learning comprehension
CN111222798A (en) * 2020-01-13 2020-06-02 湖南师范大学 Soft measurement method for key indexes of complex industrial process
CN111325112A (en) * 2020-01-31 2020-06-23 贵州大学 Cutter wear state monitoring method based on depth gate control circulation unit neural network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
QIANGQIANG NIU等: "A High-performance Web Attack Detection Method based on CNN-GRU Model", 《2020 IEEE 4TH INFORMATION TECHNOLOGY, NETWORKING, ELECTRONIC AND AUTOMATION CONTROL CONFERENCE (ITNEC 2020)》, pages 804 - 808 *
袁小峰: "深度学习在流程工业过程数据建模中的应用", 《智能科学与技术学报》, pages 107 - 115 *

Cited By (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113009906A (en) * 2021-03-04 2021-06-22 青岛弯弓信息技术有限公司 Big data prediction analysis method and system based on industrial Internet
CN112985096A (en) * 2021-03-05 2021-06-18 广州东兆信息科技有限公司 Kiln ceramic quality monitoring system and method based on Internet of things cloud platform
CN112985096B (en) * 2021-03-05 2021-09-07 广州东兆信息科技有限公司 Kiln ceramic quality monitoring system and method based on Internet of things cloud platform
CN113360851A (en) * 2021-06-22 2021-09-07 北京邮电大学 Industrial flow line production state detection method based on Gap-loss function
CN113360851B (en) * 2021-06-22 2023-03-03 北京邮电大学 Industrial flow line production state detection method based on Gap-loss function
CN113642234A (en) * 2021-08-09 2021-11-12 贵州电网有限责任公司 Power grid icing prediction method based on multi-source characteristic time convolution deep learning
CN114015825A (en) * 2021-11-09 2022-02-08 上海交通大学 Method for monitoring abnormal state of blast furnace heat load based on attention mechanism
CN114015825B (en) * 2021-11-09 2022-12-06 上海交通大学 Method for monitoring abnormal state of blast furnace heat load based on attention mechanism
CN114332474A (en) * 2021-11-25 2022-04-12 中国计量大学 Method for predicting maintenance time of key part of aircraft engine
CN114216349B (en) * 2021-12-06 2022-09-16 浙江大学 Sintering end point forecasting method based on coding and decoding network
CN114216349A (en) * 2021-12-06 2022-03-22 浙江大学 Sintering end point forecasting method based on coding and decoding network
CN114519610A (en) * 2022-02-16 2022-05-20 支付宝(杭州)信息技术有限公司 Information prediction method and device
CN114638117A (en) * 2022-03-29 2022-06-17 施昆宏 Data processing method and system based on artificial intelligence and cloud platform
CN115457367A (en) * 2022-09-22 2022-12-09 淮阴工学院 Lightweight target detection method and system based on Light-Net
CN115457367B (en) * 2022-09-22 2023-09-26 淮阴工学院 Light-Net-based lightweight target detection method and system
CN115713987A (en) * 2022-11-17 2023-02-24 广州瑞博新材料技术研究有限公司 Polycaprolactone test data analysis method and system
CN115713987B (en) * 2022-11-17 2023-06-13 广州瑞博新材料技术研究有限公司 Polycaprolactone test data analysis method and system
CN116612836A (en) * 2023-07-19 2023-08-18 福建德尔科技股份有限公司 Tail gas amount prediction method and system for trifluoromethane production
CN116612836B (en) * 2023-07-19 2023-10-03 福建德尔科技股份有限公司 Tail gas amount prediction method and system for trifluoromethane production
CN117035559A (en) * 2023-10-08 2023-11-10 国网江苏省电力有限公司电力科学研究院 Electrical equipment multi-parameter transmitter simulation installation evaluation method and system
CN117035559B (en) * 2023-10-08 2024-01-23 国网江苏省电力有限公司电力科学研究院 Electrical equipment multi-parameter transmitter simulation installation evaluation method and system

Also Published As

Publication number Publication date
CN112001527B (en) 2024-01-30

Similar Documents

Publication Publication Date Title
CN112001527B (en) Industrial production process target data prediction method of multi-feature fusion depth neural network
CN109884892B (en) Process industrial system prediction model based on cross correlation time-lag grey correlation analysis
CN110066895B (en) Stacking-based blast furnace molten iron quality interval prediction method
CN109935280B (en) Blast furnace molten iron quality prediction system and method based on ensemble learning
CN107526927B (en) Blast furnace molten iron quality online robust soft measurement method
WO2021093362A1 (en) Precision compensation method for discharge caustic alkali concentration measurement device during evaporation process
CN106156434A (en) Sliding window time difference Gaussian process regression modeling method based on the low and deep structure of local time
CN108897354B (en) Aluminum smelting process hearth temperature prediction method based on deep belief network
Yan et al. DSTED: A denoising spatial–temporal encoder–decoder framework for multistep prediction of burn-through point in sintering process
CN114015825B (en) Method for monitoring abnormal state of blast furnace heat load based on attention mechanism
CN111860839A (en) Shore bridge fault monitoring method based on multi-signal fusion and Adam optimization algorithm
CN114282443A (en) Residual service life prediction method based on MLP-LSTM supervised joint model
CN112363462A (en) Static-dynamic cooperative sensing complex industrial process running state evaluation method
CN115759409A (en) Water gate deformation prediction method for optimizing LSTM (least Square TM) model by multi-time mode attention mechanism
CN111983140B (en) Carbon monoxide measuring system and method for dry quenching production
Liu et al. Temporal hypergraph attention network for silicon content prediction in blast furnace
CN113517037A (en) Method and system for predicting sintering ore FeO by fusing data and knowledge
CN116244596A (en) Industrial time sequence data anomaly detection method based on TCN and attention mechanism
CN108709426B (en) Sintering machine air leakage fault online diagnosis method based on frequency spectrum characteristic bilateral detection method
CN111798023A (en) Method for predicting comprehensive coke ratio in steelmaking sintering production
CN116432699A (en) Method for predicting silicon content of blast furnace molten iron based on time sequence chart packing network
CN116305985A (en) Local intelligent ventilation method based on multi-sensor data fusion
CN113836813B (en) Blast furnace tuyere water leakage detection method based on data analysis
CN113849479A (en) Comprehensive energy supply station oil tank leakage detection method based on instant learning and self-adaptive threshold
Trofimov et al. Generalized structure of an intelligent control system for technological objects and experience of its use

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