CN112001115B - Soft measurement modeling method of semi-supervised dynamic soft measurement network - Google Patents
Soft measurement modeling method of semi-supervised dynamic soft measurement network Download PDFInfo
- Publication number
- CN112001115B CN112001115B CN202010690011.2A CN202010690011A CN112001115B CN 112001115 B CN112001115 B CN 112001115B CN 202010690011 A CN202010690011 A CN 202010690011A CN 112001115 B CN112001115 B CN 112001115B
- Authority
- CN
- China
- Prior art keywords
- soft measurement
- data
- semi
- model
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000005259 measurement Methods 0.000 title claims abstract description 97
- 238000000034 method Methods 0.000 title claims abstract description 72
- 238000012549 training Methods 0.000 claims abstract description 44
- 238000012545 processing Methods 0.000 claims abstract description 13
- 230000009466 transformation Effects 0.000 claims abstract description 5
- 230000008569 process Effects 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 14
- 238000004422 calculation algorithm Methods 0.000 claims description 13
- 230000008030 elimination Effects 0.000 claims description 8
- 238000003379 elimination reaction Methods 0.000 claims description 8
- 238000013528 artificial neural network Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 5
- 230000009467 reduction Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 238000002474 experimental method Methods 0.000 description 11
- 238000000354 decomposition reaction Methods 0.000 description 10
- 238000004519 manufacturing process Methods 0.000 description 9
- 238000012360 testing method Methods 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000000691 measurement method Methods 0.000 description 5
- 230000000295 complement effect Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000003068 static effect Effects 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 230000000052 comparative effect Effects 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 230000006399 behavior Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000007781 pre-processing Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000009776 industrial production Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000002994 raw material Substances 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000011426 transformation method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/02—Preprocessing
- G06F2218/04—Denoising
Abstract
The invention discloses a soft measurement modeling method of a semi-supervised dynamic soft measurement network, which is implemented according to the following steps: denoising and redundancy removing processing is carried out on the training set data based on CEEMD and Isomap methods; serializing and normalizing the training set data; and completing the soft measurement model establishment of the semi-supervised dynamic soft measurement network based on the training set data. The modeling method of the invention removes noise and redundancy of data by using a method combining CEEMD and Isomap, the CEEMD has completeness and no obvious modal aliasing phenomenon in the result, the Isomap has strong nonlinear characteristic transformation capability, the advantages of the CEEMD and the Isomap are combined, thereby effectively removing noise and redundancy in the original data and reducing information loss to the greatest extent, and the data is serialized to introduce historical data for dynamic modeling.
Description
Technical Field
The invention belongs to the technical field of intelligent signal processing and industrial artificial intelligence, and particularly relates to a soft measurement modeling method of a semi-supervised dynamic soft measurement network.
Background
In a complex industrial process, reliable and real-time measurement of key quality variables in the process cannot be realized due to severe production environment, complex working conditions, limited detection technology or cost and the like. To overcome this difficulty, soft measurement techniques have been developed that take as input auxiliary variables that are easily measured during the process, as outputs the dominant variables that are desired to be measured, and build a model that predicts the dominant variables, thereby allowing accurate estimation of the critical quality variables.
Along with development of information technology and wide application of a distributed control system in a complex industrial process, massive industrial process big data are collected and used for establishing a soft measurement model, but external environment disturbance factors and fluctuation of the process include factors such as raw material composition change, random interference in data transmission and storage processes, and the like, so that industrial data often have massive data noise, meanwhile, massive industrial production data inevitably introduce a data redundancy problem, namely the same production condition and collinearity among different auxiliary variables which occur repeatedly in the production process, noise and redundancy contained in the data can seriously influence the accuracy of a data driving soft measurement model, and the data driving soft measurement model needs to be removed in a preprocessing stage. In addition, industrial process data is often a continuous time sequence, if static soft measurement modeling is performed on the industrial process data, the front-back relation of the data cannot be captured, and therefore model estimation accuracy is low and robustness is poor in practical application. Therefore, to build an accurate data-driven soft measurement model, it needs to be dynamically modeled.
Disclosure of Invention
The invention aims to provide a soft measurement modeling method of a semi-supervised dynamic soft measurement network, which is used for capturing dynamic characteristics among data while removing variable data noise and redundancy so as to establish an accurate dynamic soft measurement prediction model.
In order to solve the technical problems, the invention discloses a soft measurement modeling method of a semi-supervised dynamic soft measurement network, which is implemented according to the following steps:
step 1, denoising and redundancy removing processing is carried out on training set data based on a complementary integrated empirical mode decomposition (Complementary Ensemble Empirical Mode Decomposition, CEEMD) and Isomap method;
step 2, carrying out serialization and normalization processing on the training set data processed in the step 1;
and 3, completing soft measurement model establishment of a semi-supervised dynamic soft measurement network (SSDGRU-MLR) based on the training set data processed in the step 2.
Further, step 1, denoising and redundancy removing processing is performed on training set data based on CEEMD and Isomap methods, specifically comprising the following steps:
step 1.1, an original auxiliary variable training data set X is applied to a CEEMD algorithm to obtain IMFs of each order;
step 1.2, calculating the correlation coefficient index of each IMF and the original variable signal, judging whether the IMF is noise or not based on a set threshold constant, eliminating the IMF judged to be noise, and calculating the correlation coefficient according to the following formula:
ρ (X) in formula (1) v (t),c vi (t)) represents the primary and auxiliary variables X v (t) and ith IMFc vi (t), i=1,..,and->Respectively X v (t) and the ith IMF standard deviation, wherein the value of rho is in the range of 0-1, and the closer to 1, the higher the similarity is;
and 1.3, carrying out nonlinear characteristic transformation on the residual IMF through an Isomap algorithm, and then carrying out data reconstruction by using the new modal function obtained after dimension reduction and the original remainder, thereby finally obtaining an auxiliary variable X' after denoising and redundancy elimination.
Further, in the implementation process of the Isomap algorithm, the calculation method of the geodesic distance in step 1.3 is as follows: the geodesic distance between the sample point and its neighborhood is replaced with the Euclidean distance between them; the sample point and points outside its neighborhood are replaced with the shortest path between them on the manifold.
Further, step 2, the training set data processed in step 1 is processed by serialization and normalization, and the specific steps are as follows:
step 2.1, after the denoising and redundancy removing operation, carrying out serialization operation on auxiliary variable data, and predicting a dominant variable at the t+ts+z time according to auxiliary variable data of a total ts time steps from the t time to the t+ts time, wherein ts is the time window length of input data, and z is the time step of the dominant variable to be predicted, so as to obtain data X 'after the input data X' is serialized;
step 2.2, performing standardization treatment by using a Z-SCORE method, and converting the data into data with a mean value of 0 and a variance of 1, wherein the formula is as follows:
wherein in formula (2), X' is the serialized data, μ and σ represent the mean and variance, respectively, and X in The standardized sequence data used for inputting the neural network is represented, and it is pointed out that only the characteristic data is standardized, and the original value of the tag data is kept unchanged;
further, step 3, the establishment of the SSDGRU-MLR soft measurement model is completed based on the training set data processed in step 2, and the specific steps are as follows:
step 3.1, a semi-supervised dynamic soft measurement network SSDGRU-MLR is a network formed by combining GRU units after multi-layer stacking with MLP in supervised learning, wherein the output of the last layer of the DGRU outputs soft measurement prediction results of the DGRU through a fully connected MLP network, and the MLP is a neural network of a single hidden layer and is used for regression fitting of the final key quality variable;
step 3.2, before training the model, firstly, initializing parameters of the model, and adopting an Xvaier initialization mode to make the node number of the current network layer be n in The number of output nodes is n out The manner of Xvaier initialization is to achieve a uniform distribution as follows:
after initialization is completed, the data X is serialized in The loss function of the whole model training process, which is input into the soft measurement model, can be defined as follows:
wherein in formula (4), y t Representing the label output corresponding to the t-th sequence sample, n-ts +1 representing the number of samples after serialization,prepreparation of samples representing the t-th sequenceMeasuring output; based on an optimization target of a minimized loss function, parameter updating and adjustment are carried out on the whole model through a BPTT algorithm, and batch training of the whole SSDGRU-MLR soft measurement model is finally completed through multiple iterations.
Compared with the prior art, the invention can obtain the following technical effects:
1) The invention discloses a soft measurement modeling method of a semi-supervised dynamic soft measurement network, which uses a CEEMD and Isomap combined method to remove noise and redundancy of data. Semi-supervised depth-gated loop perceptron network (Semi-Supervised Deep Gated Recurrent Units-aided MLP, SSDGRU-MLR) composed of DGRU and MLP is used for Semi-supervised dynamic modeling of the preprocessed serialized data, the SSDGRU-MLR can utilize a large number of unlabeled samples in the process, the depth structure of the SSDGRU-MLR is also beneficial to extracting high-level representation in variables, and GRU units in the structure can capture dynamic characteristics of the data and spread the dynamic characteristics over time, so that modeling effect is improved.
2) In the training process of SSDGRU-MLR, a dropout technology is used to avoid the generation of over fitting, and a callback function is designed to ensure the smooth training of the model. According to the soft measurement prediction experiment, the analysis of the experimental result can prove that compared with the traditional soft measurement method, the model is used for predicting the variable more accurately, and the effectiveness and the superiority of the method are proved in the comparison experiment based on the air preheater industrial example.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention and do not constitute a limitation on the invention. In the drawings:
FIG. 1 is a flow chart of a soft measurement modeling method of a semi-supervised dynamic soft measurement network of the present invention;
FIG. 2 is a schematic diagram of an SSDGRU-MLR model of a soft measurement modeling method of a semi-supervised dynamic soft measurement network according to the present invention;
FIG. 3 is a cross-sectional view of an air preheater in an industrial example of a soft measurement modeling method via a semi-supervised dynamic soft measurement network of the present invention;
FIG. 4 is a graph of a loss function obtained during training by a soft measurement modeling method of a semi-supervised dynamic soft measurement network of the present invention;
FIG. 5 is a graph of the decomposition result of the CEEMD of the inlet temperature variable obtained by a soft measurement modeling method of a semi-supervised dynamic soft measurement network according to the present invention;
FIG. 6 is a graph of rotor thermal deformation prediction results obtained by a soft measurement modeling method of a semi-supervised dynamic soft measurement network of the present invention;
FIG. 7 is a graph of soft measurement prediction error analysis obtained by a soft measurement modeling method of a semi-supervised dynamic soft measurement network of the present invention.
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 soft measurement modeling method of a semi-supervised dynamic soft measurement network, which is implemented according to the following steps as shown in figure 1:
step 1, denoising and redundancy removing processing is carried out on training set data based on a complementary integrated empirical mode decomposition (Complementary Ensemble Empirical Mode Decomposition, CEEMD) and Isomap method;
the method comprises the following specific steps:
step 1.1, an original auxiliary variable training data set X is applied to a CEEMD algorithm to obtain IMFs of each order;
step 1.2, calculating the correlation coefficient index of each IMF and the original variable signal, judging whether the IMF is noise or not based on a set threshold constant, eliminating the IMF judged to be noise, and calculating the correlation coefficient according to the following formula:
ρ (X) in formula (1) v (t),c vi (t)) represents the primary and auxiliary variables X v (t) and ith IMFc vi (t), i=1,..,and->Respectively X v (t) and the ith IMF standard deviation, wherein the value of rho is in the range of 0-1, and the closer to 1, the higher the similarity is;
step 1.3, carrying out nonlinear characteristic transformation on the residual IMF through an Isomap algorithm, and then carrying out data reconstruction by using a new mode function obtained after dimension reduction and an original remainder, so as to finally obtain an auxiliary variable X' after denoising and redundancy elimination;
in the implementation process of the Isomap algorithm, the calculation method of the geodesic distance is as follows: the geodesic distance between the sample point and its neighborhood is replaced with the Euclidean distance between them; the sample point and the points outside its neighborhood are replaced with the shortest path between them on the manifold;
step 2, carrying out serialization and normalization processing on the training set data processed in the step 1;
the method comprises the following specific steps:
step 2.1, after denoising and redundancy removing operation, carrying out serialization operation on auxiliary variable data, and predicting a dominant variable at the t+ts+z time according to auxiliary variable data of a total ts time step from the t time to the t+ts time, wherein ts is the time window length of input data, z is the time step of the dominant variable to be predicted, (two parameters of ts and z are required to be combined with industrial background setting) to obtain data X 'afterserialization of the input data X';
step 2.2, performing standardization treatment by using a Z-SCORE method, and converting the data into data with a mean value of 0 and a variance of 1, wherein the formula is as follows:
wherein in formula (2), X' is the serialized data, μ and σ represent the mean and variance, respectively, and X in The standardized sequence data used for inputting the neural network is represented, and it is pointed out that only the characteristic data is standardized, and the original value of the tag data is kept unchanged;
step 3, completing soft measurement model establishment of a semi-supervised dynamic soft measurement network (SSDGRU-MLR) based on the training set data processed in the step 2;
the method comprises the following specific steps:
step 3.1, a semi-supervised dynamic soft measurement network SSDGRU-MLR is a network formed by combining GRU units with MLP in supervised learning after multi-layer stacking, the overall structure is shown in figure 2, and the output of the last layer of the DGRU outputs soft measurement prediction results through a fully connected MLP network. Wherein MLP is a neural network of a single hidden layer, used for regression fitting of the last key quality variable;
step 3.2, before training the model, it is first necessary to initialize the parameters of the model, including the parameters W inside each GRU unit r 、U r 、W z 、U z Weights and offsets W between layers of the network, W, U 1 、b 1 、W j 、b j 、W l 、b l And the like, adopting an Xvaier initialization mode to enable the node number of the current network layer to be n in The number of output nodes is n out The manner of Xvaier initialization is to achieve a uniform distribution as follows:
after initialization, the serialized data is input into a soft measurement model, and is obtained through DGRU forward propagationOutputting h by hidden layer corresponding to last GRU layer at last time step, and collecting dataInputting the key dominant variable into an MLP network, and obtaining a predicted value about the key dominant variable through MLP forward propagation>Wherein h is t Representing hidden layer output obtained by DGRU in t-th sequence sample input model, y t Representing the label output corresponding to the t-th sequence sample, and n-ts+1 represents the number of samples after serialization. />Representing the predicted output of the t-th sequence sample. The prediction error can be obtained by the label value and the predicted value, so the loss function of the whole model training process can be defined as follows:
wherein y is t Representing the label output corresponding to the t-th sequence sample, and n-ts+1 represents the number of samples after serialization.Representing the predicted output of the t-th sequence sample. Based on an optimization target of a minimized loss function, parameter updating and adjustment are carried out on the whole model through a BPTT algorithm, and batch training of the whole SSDGRU-MLR soft measurement model is finally completed through multiple iterations.
The following experiment shows that the soft measurement modeling method of the semi-supervised dynamic soft measurement 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, the test set data is input into the soft measurement model established by the invention, the deformation prediction results obtained by adopting other soft measurement methods and the method are compared, the effectiveness and superiority of the modeling method of the invention are analyzed,
the method comprises the following specific steps:
(1) Fig. 3 shows a cross-sectional view of an air preheater rotor, and based on an industrial example of 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 8571, the number of samples of the verification set is 2143, and the number of samples of the test set is 1000. Training of the model is completed according to the steps 1-3, and in the training process, the models are ensured to be effectively trained by using the tricks: carrying out disorder processing on data; using the same dropout mask for each time step; setting a callback function in a program to monitor the change of verification loss;
(2) The validity and superiority of the prediction capability of the soft measurement model are tested by the input test set, 10 soft measurement models are established for comparison test experiments, and the 10 soft measurement models comprise a traditional method MLP and support vector regression SVR in the field of machine learning, a DBN, DLSTM, DGRU method of deep learning and a corresponding deep learning method of denoising and redundancy elimination treatment. The decomposition effects of EMD and CEEMD on auxiliary variables were also compared among these 10 comparative experiments. Using signal-to-noise ratio SNR, overall orthogonality index I OT And the root mean square error RMSE measures the denoising and redundancy removing results numerically; using mean absolute error (MeanAbsolute Error, MAE), mean square error (Mean Square Error, MSE) and measured coefficient R 2 As an evaluation index of the prediction performance of the dynamic soft measurement model.
Table 1 is the test results of comparative experiments using 10 soft measurement models, which verify the necessity and superiority of introducing deep learning into soft measurements to develop semi-supervised modeling.
For 6 models in total, each deep learning model is provided with two groups of experiments which are preprocessed and not preprocessed, as can be seen from table 1, no matter which model of DBN, DLSTM and DGRU is adopted, the model which is subjected to denoising and redundancy elimination realizes higher prediction precision than the corresponding model which is not preprocessed, and the result also shows that the denoising and redundancy elimination operation on auxiliary variable data is necessary, and the preprocessing method provided by the invention is effective, so that the data quality is improved effectively. Meanwhile, the DBN in the model No. 3 and the model No. 4 is established based on the industrial process static assumption, compared with a static DBN model, the estimation precision of the dynamic DLSTM and DGRU model is greatly improved, and the MAE value and the MSE value are obviously smaller, so that the dynamic characteristics in the data are acquired by virtue of the feedback structures of the DLSTM and the DGRU model, and the prediction performance of the soft measurement model is improved;
in addition, the realization of good dynamic performance needs effective training of a dynamic model, in 6 models, an optimal prediction result is realized by a model No. 8, and a graph of a loss function during training, which is drawn by the soft measurement modeling method of the semi-supervised dynamic soft measurement network, is shown in fig. 4, wherein a triangle curve represents a training set loss, a Y-shaped curve represents a verification set loss, and two loss curves are stable after the continuous descending trend along with the increase of the number of iteration rounds, and the curves are closely adjacent and fit well, so that the designed algorithm and the applied trick help the dynamic model realize effective training, and the effectiveness of dynamic soft measurement provided by the invention is also ensured for the good training of the model. Compared with widely used DLSTM, the SSDGRU-MLR model provided by the invention is worthy of further popularization in the field of soft measurement modeling;
FIG. 5 is a graph of the decomposition results of the inlet temperature variable CEEMD obtained by the soft measurement modeling method of the semi-supervised dynamic soft measurement network of the present invention, illustrating the decomposition of the inlet temperature variable for the first behavior of the inlet temperature variable signal, the IMFs obtained by CEEMD decomposition for the remaining behaviors, the IMFs obtained by CEEMD having no significant modal aliasing, thus enabling accurate expression of noise signals, while the RMSE index for model 8 and model 10 is an order of magnitude smaller than model 9 and the SNR is an order of magnitude greater than model 9, which illustrates the good decomposition results based on CEEMD, resulting in better denoising performance for model 8 and model 10,then to I OT The index is analyzed, although the PCA and the Isomap method are suitable for dimension reduction processing, the linear PCA method emphasizes the orthogonality of principal components, and faces to IMF data with strong nonlinearity, the Isomap method keeps the geodesic distance unchanged in the characteristic transformation process, so that important information of original variables is reserved more, high-quality reconstruction signals with less information loss are obtained, and the nonlinear characteristic transformation method is more effective in helping to realize accurate prediction of dominant variables. Through the analysis, the effectiveness and the superiority of the noise reduction and redundancy elimination method combining CEEMD with Isomap are further verified.
Fig. 6 is a graph of rotor thermal deformation prediction results obtained by the soft measurement modeling method of the semi-supervised dynamic soft measurement network, 1000 actual continuous test samples are selected, and meanwhile, model 1, model 3, model 4 and model 8 in table 1 are used for comparison experiments, and as can be known from fig. 6, the rotor deformation prediction value of model 8 better tracks and adapts to the change of the true value compared with other 3 static models, and the prediction error is smaller, so that the most accurate prediction is realized. The model 8 not only applies the serialized data modeling, but also further captures the dynamic characteristics among the data by adopting the SSDGRU-MLR model, thereby remarkably improving the prediction precision, and the dynamic performance is the strongest in the four models.
In order to evaluate the dynamic prediction performance of the model deeply, the rotor thermal type variable prediction error plot of the model 8 is also analyzed, fig. 7 is a prediction error analysis chart obtained by the soft measurement modeling method of the semi-supervised dynamic soft measurement network according to the present invention, where fig. 7 (a) is a prediction error chart at each time, fig. 7 (b) is a frequency histogram of the prediction error (and a kernel density estimation curve is drawn), and fig. 7 (c) is a time delay scatter plot of the prediction error. As can be seen from the observation of FIG. 7, the prediction error drift of each point is small, the kernel density curve is approximately zero-mean, the shape is bell-shaped, the prediction error of the model is approximately normal distribution, and the dynamic soft measurement result is true and reliable. The point distribution does not show a correlation on the whole, so that useful information of each auxiliary variable and dynamic characteristics among data are effectively utilized by the proposed model, and no redundant useful information can be used for prediction in errors.
By observing fig. 4-7 and table 1-table 2 and combining the above analysis, it can be clearly seen that the soft measurement modeling method of the semi-supervised dynamic soft measurement network of the present invention is effective and feasible and has certain advantages.
The invention relates to a soft measurement modeling method of a semi-supervised dynamic soft measurement network based on CEEMD, isomap and DGRU, which uses a method of combining CEEMD and Isomap to remove noise and redundancy of data. Semi-supervised depth-gated loop perceptron network (Semi-Supervised Deep Gated Recurrent Units-aided MLP, SSDGRU-MLR) composed of DGRU and MLP is used for Semi-supervised dynamic modeling of the preprocessed serialized data, the SSDGRU-MLR can utilize a large number of unlabeled samples in the process, the depth structure of the SSDGRU-MLR is also beneficial to extracting high-level representation in variables, and GRU units in the structure can capture dynamic characteristics of the data and spread the dynamic characteristics over time, so that modeling effect is improved. In the training process of SSDGRU-MLR, a dropout technology is used to avoid the generation of over fitting, and a callback function is designed to ensure the smooth training of the model. According to the soft measurement prediction experiment, the analysis of the experimental result can prove that compared with the traditional soft measurement method, the novel semi-supervised dynamic soft measurement method in the patent is used for predicting the variable more accurately, and the effectiveness and the superiority of the method are proved in the comparison experiment based on the air preheater industrial example.
Table 1 shows the test results of comparative experiments using 10 soft measurement models in the examples
Table 2 shows the results of the denoising and redundancy elimination 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 (3)
1. A soft measurement modeling method of a semi-supervised dynamic soft measurement network is characterized by comprising the following steps:
step 1, denoising and redundancy removing processing is carried out on training set data based on CEEMD and Isomap methods;
step 2, carrying out serialization and normalization processing on the training set data processed in the step 1;
step 3, completing soft measurement model establishment of the semi-supervised dynamic soft measurement network based on the training set data processed in the step 2;
in the step 1, denoising and redundancy removing processing are performed on training set data based on CEEMD and Isomap methods, and the specific steps are as follows:
step 1.1, an original auxiliary variable training data set X is applied to a CEEMD algorithm to obtain IMFs of each order;
step 1.2, calculating the correlation coefficient index of each IMF and the original variable signal, judging whether the IMF is noise or not based on a set threshold constant, eliminating the IMF judged to be noise, and calculating the correlation coefficient according to the following formula:
(1)In cov (X) v (t),c vi (t)) represents the primary and auxiliary variables X v (t) and ith IMFc vi (t), i=1,..,and->Respectively X v (t) and the ith IMF standard deviation, wherein the value of rho is in the range of 0-1, and the closer to 1, the higher the similarity is;
step 1.3, carrying out nonlinear characteristic transformation on the residual IMF through an Isomap algorithm, and then carrying out data reconstruction by using a new mode function obtained after dimension reduction and an original remainder, so as to finally obtain an auxiliary variable X' after denoising and redundancy elimination;
in the step 3, the establishment of the soft measurement model of the semi-supervised dynamic soft measurement network SSDGRU-MLR is completed based on the training set data processed in the step 2, and the specific steps are as follows:
step 3.1, a semi-supervised dynamic soft measurement network SSDGRU-MLR is a network formed by combining GRU units after multi-layer stacking with MLP in supervised learning, wherein the output of the last layer of the DGRU outputs soft measurement prediction results of the DGRU through a fully connected MLP network, and the MLP is a neural network of a single hidden layer and is used for regression fitting of the final key quality variable;
step 3.2, before training the model, firstly, initializing parameters of the model, and adopting an Xvaier initialization mode to make the node number of the current network layer be n in The number of output nodes is n out The manner of Xvaier initialization is to achieve a uniform distribution as follows:
after initialization is completed, the data X is serialized in The loss function of the whole model training process is defined as follows, and is input into a soft measurement model:
wherein in formula (4), y t Representing the label output corresponding to the t-th sequence sample, n-ts+1 representing the number of samples after serialization, y t predict A prediction output representing a t-th sequence sample; based on an optimization target of a minimized loss function, parameter updating and adjustment are carried out on the whole model through a back propagation BPTT algorithm, and batch training of the whole SSDGRU-MLR soft measurement model is finally completed through multiple iterations.
2. The soft measurement modeling method of a semi-supervised dynamic soft measurement network according to claim 1, wherein in the implementation process of the Isomap algorithm, the calculation method of the geodesic distance is as follows: the geodesic distance between the sample point and its neighborhood is replaced with the Euclidean distance between them; the sample point and points outside its neighborhood are replaced with the shortest path between them on the manifold.
3. The soft measurement modeling method of a semi-supervised dynamic soft measurement network according to claim 1, wherein step 2, the training set data processed in step 1 is serialized and normalized, and the specific steps are as follows:
step 2.1, after the denoising and redundancy removing operation, carrying out serialization operation on auxiliary variable data, and predicting a dominant variable at the t+ts+z time according to auxiliary variable data of a total ts time steps from the t time to the t+ts time, wherein ts is the time window length of input data, and z is the time step of the dominant variable to be predicted, so as to obtain data X 'after the input data X' is serialized;
step 2.2, performing standardization treatment by using a Z-SCORE method, and converting the data into data with a mean value of 0 and a variance of 1, wherein the formula is as follows:
wherein the method comprises the steps ofIn formula (2), X' is the serialized data, μ and σ represent the mean and variance, respectively, and X in The normalized sequence data used for inputting the neural network is shown, and it should be noted that only the normalization processing is performed on the feature data, and the original value of the tag data is kept unchanged.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010690011.2A CN112001115B (en) | 2020-07-17 | 2020-07-17 | Soft measurement modeling method of semi-supervised dynamic soft measurement network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010690011.2A CN112001115B (en) | 2020-07-17 | 2020-07-17 | Soft measurement modeling method of semi-supervised dynamic soft measurement network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN112001115A CN112001115A (en) | 2020-11-27 |
CN112001115B true CN112001115B (en) | 2024-04-02 |
Family
ID=73467052
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010690011.2A Active CN112001115B (en) | 2020-07-17 | 2020-07-17 | Soft measurement modeling method of semi-supervised dynamic soft measurement network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN112001115B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112668196B (en) * | 2021-01-04 | 2023-06-09 | 西安理工大学 | Mechanism and data hybrid-driven generation type countermeasure network soft measurement modeling method |
CN114841000A (en) * | 2022-05-09 | 2022-08-02 | 杭州师范大学 | Soft measurement modeling method based on modal common feature separation |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105354170A (en) * | 2015-10-26 | 2016-02-24 | 河海大学 | CEEMD and singular value decomposition based recognition method |
CN107451102A (en) * | 2017-07-28 | 2017-12-08 | 江南大学 | A kind of semi-supervised Gaussian process for improving self-training algorithm returns soft-measuring modeling method |
CN108197743A (en) * | 2017-12-31 | 2018-06-22 | 北京化工大学 | A kind of prediction model flexible measurement method based on deep learning |
CN108628164A (en) * | 2018-03-30 | 2018-10-09 | 浙江大学 | A kind of semi-supervised flexible measurement method of industrial process based on Recognition with Recurrent Neural Network model |
CN108760316A (en) * | 2018-08-16 | 2018-11-06 | 苏州大学 | Information fusion method is joined in the change of variation mode decomposition |
WO2019179340A1 (en) * | 2018-03-19 | 2019-09-26 | 河北工业大学 | Eemd- and msb-based failure feature extraction method for rolling-element bearing |
CN110852527A (en) * | 2019-11-20 | 2020-02-28 | 成都理工大学 | Reservoir physical property parameter prediction method combining deep learning |
US10678196B1 (en) * | 2020-01-27 | 2020-06-09 | King Abdulaziz University | Soft sensing of a nonlinear and multimode processes based on semi-supervised weighted Gaussian regression |
-
2020
- 2020-07-17 CN CN202010690011.2A patent/CN112001115B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105354170A (en) * | 2015-10-26 | 2016-02-24 | 河海大学 | CEEMD and singular value decomposition based recognition method |
CN107451102A (en) * | 2017-07-28 | 2017-12-08 | 江南大学 | A kind of semi-supervised Gaussian process for improving self-training algorithm returns soft-measuring modeling method |
CN108197743A (en) * | 2017-12-31 | 2018-06-22 | 北京化工大学 | A kind of prediction model flexible measurement method based on deep learning |
WO2019179340A1 (en) * | 2018-03-19 | 2019-09-26 | 河北工业大学 | Eemd- and msb-based failure feature extraction method for rolling-element bearing |
CN108628164A (en) * | 2018-03-30 | 2018-10-09 | 浙江大学 | A kind of semi-supervised flexible measurement method of industrial process based on Recognition with Recurrent Neural Network model |
CN108760316A (en) * | 2018-08-16 | 2018-11-06 | 苏州大学 | Information fusion method is joined in the change of variation mode decomposition |
CN110852527A (en) * | 2019-11-20 | 2020-02-28 | 成都理工大学 | Reservoir physical property parameter prediction method combining deep learning |
US10678196B1 (en) * | 2020-01-27 | 2020-06-09 | King Abdulaziz University | Soft sensing of a nonlinear and multimode processes based on semi-supervised weighted Gaussian regression |
Also Published As
Publication number | Publication date |
---|---|
CN112001115A (en) | 2020-11-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zhang et al. | LSTM-based analysis of industrial IoT equipment | |
CN107657088B (en) | Rolling bearing fault diagnosis method based on MCKD algorithm and support vector machine | |
CN109145516B (en) | Analog circuit fault identification method based on improved extreme learning machine | |
CN111539132B (en) | Dynamic load time domain identification method based on convolutional neural network | |
CN111562108A (en) | Rolling bearing intelligent fault diagnosis method based on CNN and FCMC | |
CN112001115B (en) | Soft measurement modeling method of semi-supervised dynamic soft measurement network | |
CN114564982B (en) | Automatic identification method for radar signal modulation type | |
CN111931983B (en) | Precipitation prediction method and system | |
CN108399434B (en) | Analysis and prediction method of high-dimensional time series data based on feature extraction | |
CN113837000A (en) | Small sample fault diagnosis method based on task sequencing meta-learning | |
CN112504682A (en) | Chassis engine fault diagnosis method and system based on particle swarm optimization algorithm | |
Huang et al. | A novel Bayesian deep dual network with unsupervised domain adaptation for transfer fault prognosis across different machines | |
CN114298134A (en) | Wind power prediction method and device and electronic equipment | |
Lv et al. | Deep transfer network with multi-kernel dynamic distribution adaptation for cross-machine fault diagnosis | |
CN114819054B (en) | Power electronic system state monitoring method based on physical information neural network | |
Sun et al. | Data-driven fault diagnosis method based on second-order time-reassigned multisynchrosqueezing transform and evenly mini-batch training | |
CN111859241B (en) | Unsupervised sound source orientation method based on sound transfer function learning | |
CN110222390B (en) | Gear crack identification method based on wavelet neural network | |
CN116595465A (en) | High-dimensional sparse data outlier detection method and system based on self-encoder and data enhancement | |
CN116226739A (en) | Map convolution network industrial process fault diagnosis method based on space-time fusion | |
CN111160464B (en) | Industrial high-order dynamic process soft measurement method based on multi-hidden-layer weighted dynamic model | |
CN114563671A (en) | High-voltage cable partial discharge diagnosis method based on CNN-LSTM-Attention neural network | |
CN111797979A (en) | Vibration transmission system based on LSTM model | |
Pan et al. | Specific radar emitter identification using 1D-CBAM-ResNet | |
CN109521176B (en) | Virtual water quality monitoring method based on improved deep extreme learning machine |
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 |