CN112668196B - Mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method - Google Patents
Mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method Download PDFInfo
- Publication number
- CN112668196B CN112668196B CN202110003048.8A CN202110003048A CN112668196B CN 112668196 B CN112668196 B CN 112668196B CN 202110003048 A CN202110003048 A CN 202110003048A CN 112668196 B CN112668196 B CN 112668196B
- Authority
- CN
- China
- Prior art keywords
- model
- data
- soft measurement
- output
- layer
- 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
Images
Classifications
-
- 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
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
Mechanism and data mixingThe driven generating type countermeasure network soft measurement modeling method comprises the following steps: sampling the auxiliary variable to obtain data q, sampling the random noise to obtain data z, inputting the q and z into a generator to be identified, which consists of a mechanism model and a data driving error compensation model, to obtain a generated sample
Will generate a sample
Inputting the real sample x into a discriminator, and back-propagating the obtained loss to obtain the optimal generator parameters; the optimal generator is taken out to serve as a trained soft measurement model, the interpretability of the model is analyzed by using a method for manipulating latent variables, and the prediction performance of the model is analyzed by comparing the model with a single-drive soft measurement model, so that the limitations that dynamic characteristics among data cannot be captured by a static model, modeling errors exist in a mechanism-drive soft measurement model, label-free data are difficult to utilize in data-drive soft measurement and the interpretability is poor are overcome, and compared with the single-drive soft measurement method, the method has the characteristics of accuracy and reliability in prediction.
Description
Technical Field
The invention belongs to the technical field of machine learning, and particularly relates to a mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method.
Background
In the implementation process of the advanced control scheme, besides the conventional process parameters such as temperature, pressure, liquid level and the like are required to be obtained online, some important process variables are required to be detected in real time, but are limited by technology, process conditions or cost, the variables are difficult to be detected directly online through a measuring element, in order to overcome the problem, the soft measurement technology takes an auxiliary variable which is easy to measure in the process as an input, a dominant variable which is required to be measured as an output, and a model which can predict the dominant variable is established, so that the accurate estimation of the key quality variable is realized.
At present, the soft measurement modeling method is mainly divided into two kinds of mechanism-driven soft measurement modeling and data-driven soft measurement modeling. The mechanism-driven soft measurement modeling directly finds out the quantitative relation between the dominant variable and the auxiliary variable, thereby establishing a mathematical model taking a differential equation, an algebraic equation or a state equation as a main expression. The mechanism-driven soft measurement modeling has the defects of high modeling difficulty, sometimes excessive model simplification, poor model portability and the like. In addition, the high nonlinearity, external uncertain disturbance and some system unknown factors in the actual industrial process cannot be generally reflected in the mechanism model, so that modeling errors exist between the mechanism model and the actual process, and the accuracy of the soft measurement detection result is affected. The data-driven soft measurement modeling method does not need to know too much process knowledge, and can mine and establish a mathematical model between the auxiliary variable and the dominant variable only by using the process data. Although the method effectively avoids the defect of mechanism-driven soft measurement, it should be pointed out that the data-driven soft measurement method has poor interpretability, the accuracy is still limited by the data volume of the process, and the utilization of rich information contained in unlabeled samples in the data is usually neglected in the modeling process, so that only the prediction accuracy in a local range can be ensured, and satisfactory online detection performance is difficult to provide.
From the above analysis, it can be seen that a single mechanism-driven or data-driven model has its own advantages and limitations, and therefore, in order to further improve the soft measurement performance, a more efficient modeling method needs to be proposed to combine the advantages of the two models, and at the same time, solve the problem when a single driving model performs soft measurement.
In addition, the data driving part in the existing hybrid driving soft measurement mostly uses an offline model, and in consideration of the fact that an actual industrial process is a dynamic system, industrial process data is always a continuous time sequence, and measurement errors generated by a mechanism model are also dynamically changed, so that the front-back relation of the data cannot be captured if static soft measurement modeling is performed, the model estimation precision is low and the robustness is poor in actual application. Therefore, to build an accurate soft measurement model, it needs to be dynamically modeled.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a mechanism and data hybrid-driven generation type anti-network soft measurement modeling method, which takes a mechanism model and a data driving dynamic error compensation model thereof together as a generator to be identified, and captures dynamic characteristics among data while modeling by using a large number of unlabeled samples, thereby establishing an accurate dynamic soft measurement prediction model with a certain interpretation.
In order to achieve the above purpose, the invention adopts the following technical scheme: the mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method comprises the following steps:
the specific method comprises the following steps of:
step 1.1, inputting a mechanism driving model into data q, inputting the data q and the data z into a vector merging function, and merging in the transverse direction, wherein the obtained vector i is used as the input of a data driving error compensation model;
step 1.2, calculating the output o of the data driving error compensation model, and simultaneously calculating the output y of the mechanism driving model, wherein the output y is added with the output y to obtain a generated sample
Wherein the data driving error compensation model is designed as a depth gating cycle unit (DGRU) which consists of an input layer, a hidden layer and an output layer, wherein the layers of the input layer and the output layer are 1, the output layer is set as a full connection layer, and the hidden layer is composed of gating cycle units (GRUs)The number of hidden layers is denoted by l, and the calculation formula of the output o is as follows:
r t ＝σ(W r i t +U r h t-1 ) (1)
z t ＝σ(W z i t +U z h t-1 ) (2)
wherein the output value of the first hidden layer at time t can be obtained by calculation of formulas (1) - (4)
In which W is r 、U r 、W z 、U z 、W i 、U i For the weight matrix to be trained, i t Input representing the current time step, h t-1 Representing the hidden state of the last time step, +.>
Representing the multiplication of matrix elements. Then, hidden state of time step t +.>
Is updated gate z using the current time step t To h t-1 And->
Performing linear calculation to obtain the value +.>
As the input of the second hidden layer, and so on, finally the value of the first hidden layer at the moment t is obtained>
And transmitting the result to an output layer for calculating the output layer, wherein formulas (5) - (6) are forward propagation calculation processes of the DGRU, GRU (DEG) represents calculation processes of formulas (1) - (4), o t The value of the output unit at time t is represented by V, the weight matrix of the output layer is represented by V, and g is the activation function of the output layer.
The specific method in the step 2 is as follows:
step 2.1, training target calculation formula of hybrid drive soft measurement model based on the generated type countermeasure network is as follows:
wherein in the formula (7), P in the formula penalty Representing penalty samples
Spatial distribution of P r Representing the spatial distribution of the real sample x, P g Representation of the generated samples->
λ represents a penalty factor, D and G represent respectively a discriminator and a generator of the generated countermeasure network, ++>
Self P r And P g In practice, a constrained soft version is implemented, i.e. for random samples +.>
The gradient norm of (2) is penalized, DDM2 represents a data driving model 2, and is designed into a double hidden layer BP neural network with hidden layer unit number of 32-16, and the network structure is determined by a cross verification method.
Step 2.2, the training process uses an RMSprop optimization algorithm, which adaptively adjusts the learning rate on the basis of a small-scale gradient descent algorithm, and is verified to be suitable for training a cyclic neural network and a generative countermeasure network; when the whole model is converged, the optimal generator parameters are obtained, and the generator part is the soft measurement model built offline and used for carrying out online prediction of rotor thermal deformation.
The specific method comprises the following steps:
step 3.1, inputting test data of the data q and the data z into a trained soft measurement model to obtain a prediction result
Assigning q in the same q and z to 0 and inputting the same into a trained soft measurement model to obtain a predicted result +.>
Assigning z in the same q and z to 0 and inputting the same into a trained soft measurement model to obtain a predicted result +.>
Step 3.2, based on the predicted performance evaluation index, observing the obtained predicted result
And->
Compared with +.>
What changes will occur, the speculation system makes decisionsBased on which results with interpretability are obtained.
Compared with the prior art, the invention has the beneficial effects that:
1) The invention relates to a dynamic soft measurement modeling method based on the mechanism and data mixed driving of a generating type countermeasure network, which uses a WGAN (WGAN-GP) with gradient penalty as a basic framework of the model, and uses a mechanism model of a process object and a data driving error compensation model thereof as a generator to be identified, wherein the error compensation model is designed as a DGRU, and the DGRU is used for dynamically modeling the process data, so that the soft measurement model can effectively track the dynamic characteristics of a system and can utilize a large number of unlabeled samples to carry out modeling. After training the model, observing what kind of change occurs to the obtained measurement result by adopting a method of manipulating the latent variables, thereby trying to make reasoning and interpretation on the output of the model and making a conclusion on the reliability of the model.
2) The effectiveness, superiority and interpretability of the novel hybrid driving dynamic soft measurement method are proved by observing and analyzing experimental results by selecting the proposed hybrid driving dynamic soft measurement model to be applied to the thermal deformation of the power station boiler air preheater rotor for measurement.
The method and the device analyze the prediction performance of the model by comparing with a single-driven soft measurement model, overcome the defects that the static model cannot capture dynamic characteristics among data, the mechanism-driven soft measurement model has modeling errors, the data-driven soft measurement is difficult to utilize unlabeled data and has poor interpretability, and compared with the single-driven soft measurement method, the method and the device for predicting variables by using the model are more accurate and reliable.
Drawings
FIG. 1 is a block diagram of the mechanism and data hybrid driven dynamic soft measurement modeling of the present invention.
FIG. 2 is a block diagram of a data-driven modeling error dynamic compensation model of the present invention.
FIG. 3 is a block diagram of a gated loop unit in a data-driven modeling error dynamic compensation model of the present invention.
FIG. 4 is a block diagram of an air preheater in an industrial example of a hybrid-driven dynamic soft measurement modeling method of the present invention.
FIG. 5 is a cross-sectional view of an air preheater in an industrial example of a hybrid-driven dynamic soft measurement modeling method of the present invention.
Fig. 6 is a schematic diagram of a method of experimental investigation of the interpretability of a hybrid drive soft measurement model by manipulating hidden variables via the present invention.
FIG. 7 is a graph of a model training process loss function obtained via the present invention.
Fig. 8 is a graph showing the prediction of rotor thermal deformation obtained by the present invention.
Fig. 9 is a graph of rotor thermal deformation prediction error compensation obtained by the present invention.
Fig. 10 is a graph of the rotor thermal deformation prediction error obtained by the present invention.
In the figure: 1-air duct I, 2-air duct II, 3-air preheater rotor, 4-air duct, 5-sector plate, 6-adjusting rod, 7-upper air leakage gap, 8-lower air leakage gap, 9-trapezoid air leakage gap and 10-rotor heat accumulation plate.
Detailed Description
The following will describe embodiments of the present invention in detail by referring to examples, so that the implementation process of how to apply the technical means to solve the technical problems and achieve the technical effects of the present invention can be fully understood and implemented.
The invention discloses a mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method, the integral framework structure of the scheme is shown in figure 1, and the method is implemented specifically according to the following steps:
The method comprises the following specific steps:
step 1.1, inputting a mechanism driving model into data q, inputting q and z into a vector merging function, and merging in the transverse direction, wherein the obtained vector i is used as the input of a data driving error compensation model;
step 1.2, calculating the output o of the data driving error compensation model, and simultaneously calculating the output y of the mechanism driving model, wherein the output y is added with the output y to obtain a generated sample
The data driving error compensation model is designed as a depth gating circulation unit (DGRU), the structure is shown in fig. 2, the DGRU is composed of an input layer, a hidden layer and an output layer, the layers of the input layer and the output layer are 1, the output layer is set as a full connection layer, the hidden layer is composed of gating circulation units, the structure of the gating circulation unit is shown in fig. 3, and the structure of the gating circulation unit mainly comprises two gate structures of an update gate and a reset gate, wherein sigma and tanh represent an activation function. The number of hidden layers of the DGRU is represented by l, and the calculation formula of the output o is as follows:
r t ＝σ(W r i t +U r h t-1 ) (1)
z t ＝σ(W z i t +U z h t-1 ) (2)
wherein the output of the first hidden layer at time t can be obtained by calculation of formulas (1) - (4)Value of
In which W is r 、U r 、W z 、U z 、W i 、U i For the weight matrix to be trained, i t Input representing the current time step, h t-1 Representing the hidden state of the last time step, +.>
Representing the multiplication of matrix elements. Then, hidden state of time step t +.>
Is updated gate z using the current time step t To h t-1 And->
And (5) performing linear calculation. The value +.>
As input of the second hidden layer, and so on, finally the value of the first hidden layer at the moment t is +.>
And transferred to the output layer for calculation of the output layer. Formulas (5) - (6) are forward propagation computation of the DGRU, GRU (. Cndot.) represents computation of formulas (1) - (4), o t The value of the output unit at time t is represented by V, the weight matrix of the output layer is represented by V, and g is the activation function of the output layer.
the method comprises the following specific steps:
step 2.1, training target calculation formula of the soft measurement model based on GAN is as follows:
wherein in the formula (7), P in the formula penalty Representing penalty samples
Spatial distribution of P r Representing the spatial distribution of the real sample x, P g Representation of the generated samples->
λ represents a penalty factor, D and G represent respectively a discriminator and a generator of the generated countermeasure network, ++>
Self P r And P g In practice, a constrained soft version is implemented, i.e. for random samples +.>
Punishment is done for the gradient norms of (c). DDM2 is a data driving model 2, and is designed into a double hidden layer BP neural network with hidden layer unit number of 32-16, and the network structure is determined by a cross verification method;
step 2.2, the training process uses an RMSprop optimization algorithm, which adaptively adjusts the learning rate on the basis of a small-scale gradient descent algorithm, has been verified to be suitable for training of a cyclic neural network and a generative type countermeasure network, and when the whole model converges, the optimal generator parameters are obtained, and the generator part is a soft measurement model which is built offline and can be used for online prediction of rotor thermal deformation.
the method comprises the following specific steps:
step 3.1, recipeThe schematic diagram of the method is shown in FIG. 6, and the test data of q and z are input into a trained soft measurement model to obtain a prediction result
Assigning q in the same q and z to 0 and inputting the same into a trained soft measurement model to obtain a predicted result +.>
Assigning z in the same q and z to 0 and inputting the value into a trained soft measurement model to obtain a prediction result
/>
Step 3.2, based on the predicted performance evaluation index, observing the obtained predicted result
And->
Compared with +.>
What kind of change occurs, so that the basis for making decisions by the system is presumed, and an interpretable result is obtained.
The following experiment shows that the dynamic soft measurement modeling method based on the mechanism and the data hybrid drive of the generation type countermeasure network is effective and feasible, and has certain advantages:
based on the industrial example of the thermal deformation soft measurement of the air preheater rotor, inputting test set data into the soft measurement model established by the invention, and analyzing the prediction performance of the model by comparing the test set data with the single-drive soft measurement model;
the method comprises the following specific steps:
(1) Fig. 4 and 5 show a structural diagram and a sectional view of the air preheater rotor 3, and the working principle is that the heat contained in the flue gas is utilized to heat the air to be fed into the hearth for supporting combustion, so that the energy utilization rate can be effectively improved. The rotating rotor alternately passes through a flue gas zone (flue 4) and an air zone (flue 1 and flue 2), the flue gas passes through to heat the rotor heat accumulation plate 10, and when the rotor rotates to an air flow zone, the rotor heat accumulation plate 10 is cooled by air and simultaneously heats the air. And (5) completing a heat exchange process every time the rotor rotates for one circle. In the cold state, the upper air leakage gap 7, the lower air leakage gap 8 and the trapezoidal air leakage gap 9 shown in fig. 5 are all tiny in variation, the air leakage amount is kept in an acceptable range, and the variation is small. In a hot state, the rotor is heated unevenly as a whole, the rotor can generate thermal deformation under the action of complex internal stress, at the moment, the air leakage gap at the lower part of the rotor and the trapezoidal air leakage gap at the circumference of the rotor become smaller than those in a cold state, the air leakage gap at the upper part of the rotor becomes larger than those in the cold state, and the air leakage quantity is increased. A large amount of hot air leakage causes huge energy waste and economic loss, and even forces the unit to run under load, so that the air leakage of the air preheater, particularly the upper air leakage gap 7, needs to be controlled. The height of the movable sector plate 5 is adjusted in a certain range by an adjusting lever 6 connected to the upper portion of the air preheater rotor 3 based on the rotor thermal deformation value measured from the gap measuring sensor in general to track and compensate the deformation gap, thereby controlling the air leakage. However, due to the harsh operating conditions of the air preheater, the gap measurement sensor is easily damaged, thereby causing the gap control system to stop automatically adjusting. And once the air preheater is damaged, the air preheater can be replaced only when the air preheater is waiting for the shutdown maintenance, so that a mathematical model is required to be established by using a soft measurement method to approach the thermal deformation rule of the air preheater, and the detection value obtained by soft measurement is used as the basis of system adjustment instead of the actual measurement value. Therefore, the air leakage of the air preheater can be continuously and effectively controlled in the period from the damage of the measuring probe to the next furnace shutdown, so that the automatic input rate of the system is ensured. Based on the industrial example of the thermal deformation soft measurement of the air preheater rotor, the serialized real-time temperature data of the preheater is divided into a training set, a verification set and a test set, wherein the number of samples of the training set is 22000, and the number of samples of the test set is 1000. And (3) training the model according to the steps 1-3. Wherein the GRU units of each hidden layer of the DGRU are set to 128-64-64-32, and each unit of the arbiter network hidden layer units are set to 32-16, these structures being selected by cross-validation. In the training process, random noise is uniformly distributed and sampled from [ -1, 1), the dimension of the random noise is set to 9, the penalty coefficient lambda is set to 10, the batch size m is set to 64, the learning rate is set to 0.001, the moving average parameter is set to a default value of 0.9, and the time frame parameter L is set to 12 according to sensitivity analysis.
(2) The paper builds up a total of 3 soft measurement models for comparative testing experiments, and in addition to the Hybrid Drive Model (HDM) presented herein, a Data Driven Model (DDM) and a mechanism driven soft measurement model (MDM) for testing. Wherein the DDM is built using a Deep Belief Network (DBN) and the structure of the DBN is set to 156-128-64-32-1 through cross-validation. MDM is built based on formulas (8) - (9), and the specific calculation formula is as follows:
y＝0.006×ΔTR 2 /H(8)
wherein in the formulae (8) to (9), Y is the maximum thermal deformation amount, mm; delta T is the temperature difference between the cold end and the hot end, and the temperature is lower than the temperature; r is the radius of the rotor, m; h is the rotor height, m; t (T) gi The temperature is the flue gas side inlet temperature; t (T) ao Is the air side outlet temperature; t (T) go Is the flue gas side outlet temperature; t (T) ai Is the air side inlet temperature.
The average absolute error MAE and the correlation coefficient R index are adopted to evaluate the prediction performance of the dynamic soft measurement model: the smaller the MAE, the closer the predicted value is to the corresponding real value, and the more accurate the model is predicted. The value range of R is between [ -1,1], and the closer R is to 1, the higher the similarity between the predicted value and the true value in the changing process is, and the dynamic performance of the model is good.
Fig. 7 is a graph of change of training loss values drawn after 50 rounds of training the model, in which fig. 7 (a) shows the change of the loss value of the discriminant during the training process, fig. 7 (b) shows the change of the loss value of the generator during the training process, and it can be seen that the change trends of the two curves are opposite, and the two curves are always in a state of mutual opposition in the whole training process, because the loss value of the WGAN discriminant is considered to be able to indicate the training process, the smaller the value thereof is, the more effective the WGAN training is, and the observation 7 (a) shows that the loss value of the discriminant approaches to 0 along with the training process, which indicates that the whole GAN-based hybrid driving model is better trained, and the model can stabilize the training to achieve convergence.
Table 1 is the test results of a comparative experiment using 3 soft measurement models in which the MAE value of the HDM is minimum and the R value is maximum and close to 1, thus the proposed HDM achieves better predictive performance than a single drive soft measurement for the task of rotor thermal deformation measurement. In order to observe the predicted results more clearly and intuitively, the predicted results of the three models on 100 test samples are selected, a predicted result curve is drawn, and one section of the predicted curve is selected for amplification, as shown in fig. 8. From an observation of the image, compared with MDM and DDM, the predicted curve of HDM has no excessive burrs or intense shaking, and the predicted value of HDM to rotor thermal deformation is more accurate.
FIG. 9 is a graph of rotor thermal deformation prediction error compensation obtained by a hybrid drive dynamic soft measurement modeling method of the present invention, which is plotted as an error compensation image for further analysis of the dynamic performance of the model. The solid triangle curve in the figure represents the prediction error generated when the MDM model is used for prediction, the open triangle curve represents the output value of the DDM1 part of the MDM model when the HDM model is used for prediction, namely the compensation value of the DGRU on the MDM prediction error, so that the error compensation value can track the change of the real error curve in time, the capability of capturing dynamic information of the DGRU network in the model is reflected, and the R value of the HDM is as high as 0.9984 in combination with the capability of capturing dynamic information shown in the table 1, which shows that the built hybrid driving soft measurement model has better dynamic performance.
FIG. 10 is a graph of the predicted error of rotor thermal deformation obtained by a hybrid drive dynamic soft measurement modeling method of the present invention, wherein the error between the predicted value and the true value of the rotor thermal deformation is calculated due to the optimal prediction performance achieved by HDM, and the error change curve is drawn as shown in the solid triangle curve in FIG. 10. It is believed that the accuracy of measurement of air preheater rotor heat distortion at (-0.2,0.3) has negligible impact on the air leakage gap control. It can be seen that the error generated by the HDM method on the test set individually exceeds the measurement accuracy requirement, but the exceeding value is not large, and the hybrid driving soft measurement method is used as an emergency measurement means in a fault state, and the error is in an acceptable range.
Table 2 is an experimental result of an experimental study on the interpretability of the HDM, and the interpretation of the HDM is studied according to the scheme shown in fig. 6, and it is known from the table that the MAE value is increased and the R value is decreased in both cases, which indicates that the prediction performance of the model is reduced to different extents, and the auxiliary variable and the random noise are taken as the input of the generator in the modeling, which helps to improve the prediction performance of the model to a greater extent, and the effectiveness of the error compensation mechanism in the HDM is represented, that is, the output of the model is obtained after the MDM is compensated, so that the reliability of the prediction result of the model is further ensured. In addition, the MAE value of HDM1 is greater than that of HDM2, i.e., the predictive performance of HDM2 is worse, which suggests that temperature variation contributes less to the improvement in model predictive performance relative to random noise. The experimental results are consistent with the actual cognition, because MDM shown in the formulas (8) - (9) already determines the functional relation between rotor thermal deformation and each temperature variable through strict mathematical derivation, and then the same temperature variable is input into DDM1 for information mining, so that compensation of the prediction error of the mechanism model is necessarily limited. In fact, in addition to the temperature variables, the variables such as smoke and air flow rate also have a certain influence on the thermal deformation, but these factors are not considered in the MDM, so this modeling error is compensated by the introduced random noise, and the experimental result is observed that if the random noise input is assigned as a 0 vector, the MAE of the model predicted value is greatly increased from 0.1891 to 0.2792, and the r value is also reduced by 0.0033, which fully explains the necessity of introducing the random noise to compensate the prediction error. Compared with HDM1, HDM has only 4-dimensional input (four dimensions correspond to four temperature auxiliary variables respectively), MAE is reduced by 0.0245, and the introduction of temperature variables still helps to improve the model prediction performance, which is considered to be two reasons: 1. an error compensation part (DGRU) in the model effectively learns the time sequence relation of temperature variable data, and compensates the unavoidable prediction error of establishing MDM by using a static function; 2. the relation between rotor thermal deformation and temperature in the actual process is not necessarily a linear relation as shown by MDM, and the DGRU realizes effective fitting on the nonlinear relation between the rotor thermal deformation and temperature by virtue of the depth structure of the DGRU.
By observing fig. 7-10 and table 1-table 2 and combining the above analysis, it can be clearly seen that the dynamic soft measurement modeling method based on the mechanism and data hybrid driving of the generated countermeasure network is effective and feasible, has a certain interpretation, and has a certain superiority compared with the single driving soft measurement method.
The invention relates to a dynamic soft measurement modeling method based on the mechanism and data mixed driving of a generating type countermeasure network, which takes a mechanism model of a process object and a data driving error compensation model thereof as a generator to be identified, wherein the error compensation model is designed as a DGRU, and the DGRU is used for carrying out dynamic modeling on process data containing a large amount of non-tag data, so that the soft measurement model can effectively track the dynamic characteristics of a system and effectively mine information contained in the non-tag data. According to the soft measurement prediction experiment, the experimental result is analyzed, so that the prediction performance is good, and the method has certain superiority compared with a soft measurement method driven by a single mechanism or data; the designed data driving model DGRU can effectively compensate the prediction error of the mechanism model, and the DGRU can capture the time sequence relation among variable data and utilize unlabeled data to carry out training, so that dynamic soft measurement modeling is realized; experiments are carried out on the interpretive performance of HDMs, and the built HDMs are high in reliability and have a certain interpretive performance; the variable is predicted more accurately by using the novel hybrid drive soft measurement method in the patent, and the effectiveness, the interpretability and the superiority of the method are proved in a comparison experiment based on an air preheater industrial example.
Table 1 shows the test results of comparative experiments using 3 soft measurement models in the examples
Table 2 shows the test results of the experiment for model interpretability in the examples
While the foregoing description illustrates and describes several preferred embodiments of the invention, it is to be understood that the invention is not limited to the forms disclosed herein, but is not to be construed as limited to other embodiments, and is capable of use in various other combinations, modifications and environments and is capable of changes or modifications within the spirit of the invention described herein, either as a result of the foregoing teachings or as a result of the knowledge or skill of the relevant art. And that modifications and variations which do not depart from the spirit and scope of the invention are intended to be within the scope of the appended claims.
Claims (2)
1. The mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method is characterized by comprising the following steps of:
step 1, sampling data q from auxiliary variables, sampling data z from random noise, inputting q and z into a generator to be identified, which consists of a mechanism model and a data driving error compensation model, and obtaining a generated sample
The specific method comprises the following steps of:
step 1.1, inputting a mechanism driving model into data q, inputting the data q and the data z into a vector merging function, and merging in the transverse direction, wherein the obtained vector i is used as the input of a data driving error compensation model;
step 1.2, calculating the output o of the data-driven error compensation model, while calculating the output o of the data-driven error compensation modelThe output y, o and y of the principle driving model are added to obtain a generated sample
The data driving error compensation model is designed into a depth gating circulation unit, the depth gating circulation unit is composed of an input layer, a hidden layer and an output layer, the number of layers of the input layer and the output layer is 1, the output layer is set to be a full connection layer, the hidden layer is composed of gating circulation units, the number of layers of the hidden layer is represented by l, and then the calculation formula of the output o is as follows:
r t ＝σ(W r i t +U r h t-1 )(1)
z t ＝σ(W z i t +U z h t-1 )(2)
wherein the output value of the first hidden layer at time t can be obtained by calculation of formulas (1) - (4)
In which W is r 、U r 、W z 、U z 、W i 、U i For the weight matrix to be trained, i t Input representing the current time step, h t-1 Representing the hidden state of the last time step,
representing the multiplication of matrix elements, then the hidden state h of time step t t 1 Is updated gate z using the current time step t To h t-1 And->
Performing linear calculation to obtain the value +.>
As the input of the second hidden layer, and so on, finally the value of the first hidden layer at the moment t is obtained>
And pass it to the output layer for output layer computation, formulas (5) - (6) are forward propagation computation of DGRU, GRU (·) represents computation of formulas (1) - (4), o t The value of the output unit at the time t is represented by V, the weight matrix of the output layer is represented by g, the activation function of the output layer is represented by sigma, and h t Is the hidden state of the current time step t, r t Reset gate for current time step t, z t Update gate for current time step t +.>
The candidate hidden state is the candidate hidden state of the current time step t;
step 2, generating a sample
Inputting the real sample x into a discriminator, and back-propagating the obtained loss to obtain the optimal generator parameters;
the specific method in the step 2 is as follows:
step 2.1, training target calculation formula of hybrid drive soft measurement model based on the generated type countermeasure network is as follows:
wherein in the formula (7), P in the formula penalty Representing penalty samples
Spatial distribution of P r Representing the spatial distribution of the real sample x, P g Representation of the generated samples->
λ represents a penalty factor, D and G represent respectively a discriminator and a generator of the generated countermeasure network, ++>
Self P r And P g In practice, a constrained soft version is implemented, i.e. for random samples +.>
The gradient norm of (2) is punished, DDM2 represents a data driving model 2, a double hidden layer BP neural network with the hidden layer unit number of 32-16 is designed, and the network structure is determined by a cross verification method;
step 2.2, the training process uses an RMSprop optimization algorithm, which adaptively adjusts the learning rate on the basis of a small-scale gradient descent algorithm, and is verified to be suitable for training a cyclic neural network and a generative countermeasure network; when the whole model is converged, the optimal generator parameters are obtained, and the generator part is the soft measurement model built offline and is used for carrying out online prediction on rotor thermal deformation;
and 3, taking out the optimal generator as a trained soft measurement model, and analyzing the interpretability of the model by using a method for manipulating the latent variables.
2. The method for modeling mechanism and data hybrid driven generation type countermeasure network soft measurement according to claim 1, wherein the step 3 comprises the following specific steps:
step 3.1, inputting test data of the data q and the data z into a trained soft measurement model to obtain a prediction result
Assigning q in the same q and z to 0 and inputting the same into a trained soft measurement model to obtain a predicted result +.>
Assigning z in the same q and z to 0 and inputting the same into a trained soft measurement model to obtain a predicted result +.>
Step 3.2, based on the predicted performance evaluation index, observing the obtained predicted result
And->
Compared with +.>
What kind of change occurs, the basis for decision making by the system is presumed, and an interpretable result is obtained. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110003048.8A CN112668196B (en) | 2021-01-04 | 2021-01-04 | Mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110003048.8A CN112668196B (en) | 2021-01-04 | 2021-01-04 | Mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112668196A CN112668196A (en) | 2021-04-16 |
| CN112668196B true CN112668196B (en) | 2023-06-09 |
Family
ID=75412683
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110003048.8A Active CN112668196B (en) | 2021-01-04 | 2021-01-04 | Mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112668196B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113239565B (en) * | 2021-05-27 | 2022-07-12 | 中南大学 | Soft measurement method and device for product quality of fluidized bed roaster |
| CN113627064B (en) * | 2021-09-03 | 2023-11-21 | 广东工业大学 | Roller kiln firing zone temperature prediction method based on mechanism and data mixed driving |
| CN116226702B (en) * | 2022-09-09 | 2024-04-26 | 武汉中数医疗科技有限公司 | Thyroid sampling data identification method based on bioelectrical impedance |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019071384A1 (en) * | 2017-10-09 | 2019-04-18 | Bl Technologies, Inc. | Intelligent systems and methods for process and asset health diagnosis, anomoly detection and control in wastewater treatment plants or drinking water plants |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7313550B2 (en) * | 2002-03-27 | 2007-12-25 | Council Of Scientific & Industrial Research | Performance of artificial neural network models in the presence of instrumental noise and measurement errors |
| CN109002686B (en) * | 2018-04-26 | 2022-04-08 | 浙江工业大学 | Multi-grade chemical process soft measurement modeling method capable of automatically generating samples |
| US11048974B2 (en) * | 2019-05-06 | 2021-06-29 | Agora Lab, Inc. | Effective structure keeping for generative adversarial networks for single image super resolution |
| CN111753881B (en) * | 2020-05-28 | 2024-03-29 | 浙江工业大学 | Concept sensitivity-based quantitative recognition defending method against attacks |
| CN112001115B (en) * | 2020-07-17 | 2024-04-02 | 西安理工大学 | Soft measurement modeling method of semi-supervised dynamic soft measurement network |
-
2021
- 2021-01-04 CN CN202110003048.8A patent/CN112668196B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019071384A1 (en) * | 2017-10-09 | 2019-04-18 | Bl Technologies, Inc. | Intelligent systems and methods for process and asset health diagnosis, anomoly detection and control in wastewater treatment plants or drinking water plants |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112668196A (en) | 2021-04-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112668196B (en) | Mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method | |
| CN108985376B (en) | Rotary kiln sequence working condition identification method based on convolution-cyclic neural network | |
| Han et al. | An optimized long short-term memory network based fault diagnosis model for chemical processes | |
| CN112085277B (en) | SCR denitration system prediction model optimization method based on machine learning | |
| Lian et al. | Soft sensor based on DBN-IPSO-SVR approach for rotor thermal deformation prediction of rotary air-preheater | |
| Li et al. | Dynamic time features expanding and extracting method for prediction model of sintering process quality index | |
| CN110782067B (en) | Sintering end point fluctuation range prediction method based on fuzzy information granulation | |
| CN110189800B (en) | Furnace oxygen content soft measurement modeling method based on multi-granularity cascade cyclic neural network | |
| CN107038307A (en) | Mechanism predicts integrated modelling approach with the Roller Conveying Kiln for Temperature that data are combined | |
| Guo et al. | A hybrid mechanism-and data-driven soft sensor based on the generative adversarial network and gated recurrent unit | |
| CN115700330A (en) | Boiler reheater temperature deviation prediction method based on time-space fusion deep neural network | |
| Liu et al. | Temporal hypergraph attention network for silicon content prediction in blast furnace | |
| Tian et al. | Coke oven flue temperature control based on improved implicit generalized predictive control | |
| CN116757078A (en) | Method and system for measuring flow velocity of pulverized coal based on acting force | |
| CN114036758B (en) | Boiler combustion state dynamic display method based on numerical simulation and machine learning | |
| CN116429269A (en) | Infrared intelligent analysis system for ethylene cracking furnace tube | |
| CN115688865A (en) | Industrial soft measurement method for long and short term memory network for flue gas of desulfurization process | |
| Zhang et al. | Information Complementary Fusion Stacked Autoencoders for Soft Sensor Applications in Multimode Industrial Processes | |
| Karri et al. | Artificial neural networks and neuro-fuzzy inference systems as virtual sensors for hydrogen safety prediction | |
| CN113111911A (en) | Defect depth detection method based on principal component analysis and gate control circulation unit network | |
| Zhang et al. | Energy consumption diagnosis in the iron and steel industry via the Kalman filtering algorithm with a data-driven model | |
| Liu et al. | Modeling of hydraulic turbine systems based on a Bayesian-Gaussian neural network driven by sliding window data | |
| CN113223634A (en) | Blast furnace molten iron silicon content prediction method based on two-dimensional self-attention enhanced GRU model | |
| Shi et al. | A detection method for porosity of gas turbine blade coating based on gray gradient space histogram entropy | |
| CN111291020A (en) | Dynamic process soft measurement modeling method based on local weighted linear dynamic system |
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 |