CN117150931B - Mixed oil length on-line estimation method and system based on mixed single hidden layer neural network - Google Patents
Mixed oil length on-line estimation method and system based on mixed single hidden layer neural network Download PDFInfo
- Publication number
- CN117150931B CN117150931B CN202311411498.6A CN202311411498A CN117150931B CN 117150931 B CN117150931 B CN 117150931B CN 202311411498 A CN202311411498 A CN 202311411498A CN 117150931 B CN117150931 B CN 117150931B
- Authority
- CN
- China
- Prior art keywords
- neural network
- parameters
- mixed
- length
- sample
- 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
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 82
- 238000000034 method Methods 0.000 title claims abstract description 75
- 238000009826 distribution Methods 0.000 claims description 108
- 238000012549 training Methods 0.000 claims description 36
- 230000008569 process Effects 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 11
- 210000002569 neuron Anatomy 0.000 claims description 10
- 239000000203 mixture Substances 0.000 claims description 9
- 230000001174 ascending effect Effects 0.000 claims description 8
- 230000006870 function Effects 0.000 claims description 8
- 239000011159 matrix material Substances 0.000 claims description 8
- 238000007781 pre-processing Methods 0.000 claims description 2
- 238000004519 manufacturing process Methods 0.000 abstract description 5
- 238000005259 measurement Methods 0.000 abstract description 5
- 239000003921 oil Substances 0.000 description 89
- 238000012360 testing method Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000005520 cutting process Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 238000011002 quantification Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000003860 storage 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/08—Fluids
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Geometry (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Investigating Or Analysing Biological Materials (AREA)
- Complex Calculations (AREA)
Abstract
The invention belongs to the field of soft measurement of finished oil pipelines, and provides an online estimation method and an online estimation system for the length of mixed oil based on a mixed single hidden layer neural network, wherein the scheme is as follows: a novel semi-supervised mixed neural network structure is designed, and a principle of divide-and-conquer is adopted, so that the requirement on a shallow neural network is reduced, and a deep neural network with high complexity is avoided; the weighted combination of the plurality of shallow neural networks can obtain better performance than the single shallow neural network or the deep neural network; the regularization coefficient and the number of mixed modes are adaptively determined by adopting a Bayes method, so that the overfitting resistance is effectively improved; the non-label sample information can be fully mined by adopting a semi-supervised learning framework; finally, the method can simultaneously provide the estimated value of the oil mixing length and the confidence interval thereof. By the method, the neural network structure can be effectively determined, the risk of overfitting is reduced, the prediction accuracy is improved, and technical support and guarantee are provided for improving the product quality, optimizing the production operation and scheduling decision.
Description
Technical Field
The invention belongs to the field of soft measurement of finished oil pipelines, and particularly relates to an online estimation method and an online estimation system for the length of mixed oil based on a mixed single hidden layer neural network.
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
In the process of the product oil pipeline, different types of product oil are required to be conveyed according to a certain sequence and batches, and a mixing zone is formed in the contact area of two adjacent batches and different types of oil products. Because the oil quality of the mixing area does not reach the standard, the oil mixing area needs to be cut. If the cutting length is too long, the oil product meeting the requirements can be cut off; too small a cutting length can result in unacceptable oils entering the large storage tank. The reasonable prediction of the oil mixing length can save the cost and improve the economic benefit. Therefore, to achieve optimization, scheduling and decision-making of pipeline transportation, prediction of the oil blend length is required.
The prediction performance of the existing mechanism model excessively depends on expert knowledge, a large number of complex factors and production state variables are not considered, high-value information which can be perceived, such as the oil mixing interface station passing history data, is ignored, and the prediction accuracy is difficult to ensure. The data-driven soft measurement model utilizes historical data to realize online real-time estimation of the oil mixing length. The neural network is a model widely applied in the field of soft measurement, has strong nonlinear approximation capability, can be used for learning nonlinear relations in complex process data, but has obvious multi-modal characteristics due to the fact that the generation of oil mixing areas is very complex and the data difference between different sites is very large, and marked sample data are less, so that a traditional single hidden layer neural network (such as a BP neural network, an extreme learning machine and the like) and a deep network (such as a deep neural network, a stacked self-encoder and the like) are difficult to fully train, cannot fully and accurately capture data characteristics, and obtain good prediction accuracy.
Disclosure of Invention
In order to solve at least one technical problem in the background art, the invention provides the method and the system for on-line estimation of the mixed oil length based on the mixed single hidden layer neural network, which can simultaneously process the problems of nonlinearity, multiple modes, rare labeled samples and the like in the oil mixing process, are beneficial to improving the estimation precision of the mixed oil length and provide guarantee for optimization, scheduling and decision of pipeline transportation.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
the first aspect of the invention provides an oil mixing length on-line estimation method based on a mixed single hidden layer neural network, which comprises the following steps:
selecting an auxiliary variable associated with the length of the mixed oil;
acquiring a labeled sample set simultaneously containing auxiliary variables and the length of mixed oil and a non-labeled sample set only containing the auxiliary variables as training data sets;
training the oil mixing length on-line estimation model based on a training data set to obtain a trained oil mixing length on-line estimation model, wherein the training process comprises the following steps: setting input parameters of an oil mixing length on-line estimation model, initializing the prior and posterior distributed super-parameters of each neural network weight parameter and random variable, and learning the random variable in the step E by adopting a variational reasoning principleUpdating the lower bound of evidence of an observation sample, updating the weight parameters of each neural network in the step M through a gradient ascending method, and alternately performing the step E and the step M until convergence conditions are met, so as to obtain the posterior distribution of the weight parameters of each neural network and random variables in the mixed oil length online estimation model; and obtaining the oil mixing length estimated value and the confidence interval of the sample to be predicted, which only contain auxiliary variables, based on the trained oil mixing length on-line estimation model.
The second aspect of the invention provides an oil mixing length on-line estimation system based on a mixed single hidden layer neural network, comprising:
a variable selection module for selecting an auxiliary variable associated with the length of the mixed oil;
the training data set acquisition module is used for acquiring a labeled sample set which simultaneously comprises an auxiliary variable and a mixed oil length and an unlabeled sample set which only comprises the auxiliary variable as training data sets;
the training module of the oil mixing length online estimation model is used for training the oil mixing length online estimation model based on a training data set to obtain a trained oil mixing length online estimation model, and the training process comprises the following steps: setting input parameters of an oil mixing length on-line estimation model, initializing the prior and posterior distributed super-parameters of each neural network weight parameter and random variable, and learning the random variable in the step E by adopting a variational reasoning principleUpdating the lower bound of evidence of an observation sample, updating the weight parameters of each neural network in the step M through a gradient ascending method, and alternately performing the step E and the step M until convergence conditions are met, so as to obtain the posterior distribution of the weight parameters of each neural network and random variables in the mixed oil length online estimation model; the oil mixing length estimation module is used for obtaining an oil mixing length estimation value based on a sample to be predicted only containing auxiliary variables and a trained oil mixing length online estimation model.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention designs a novel semi-supervised hybrid neural network structure, adopts a principle of divide-and-conquer, establishes a nonlinear relation between input and output by using a shallow neural network in each mode, reduces the requirement on a single shallow neural network, and simultaneously avoids using a deep neural network with higher model complexity; the weighted combination of the plurality of local shallow neural networks can obtain better performance than the single shallow neural network and the deep neural network.
2. According to the invention, the regularization coefficient and the number of mixed modes are adaptively determined by adopting a Bayes method, so that the overfitting resistance is effectively improved under the condition of no need of manual intervention; by adopting the semi-supervised learning framework, a small number of labeled samples can be utilized, and the potential information of a large number of unlabeled samples can be fully mined.
3. The invention not only can provide the estimated value of the oil mixing length, but also can provide the corresponding confidence interval. By the method, the neural network structure can be effectively determined, the risk of overfitting is reduced, and the unlabeled samples are fully utilized, so that the prediction precision is improved, and technical support and guarantee are provided for improving the product quality, optimizing the production operation and making optimal scheduling decisions.
Additional aspects of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
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.
Fig. 1 is a flowchart of an online estimation method of a mixed oil length based on a mixed single hidden layer BP neural network according to an embodiment of the present invention;
fig. 2 is a schematic diagram of an estimation result of a mixed oil length according to an embodiment of the present invention, wherein an abscissa represents a test sample number, an ordinate represents a mixed oil length (scaled to a 0-1 interval), a dotted line represents a true value of the mixed oil length, and a solid line represents an estimated value of the mixed oil length;
fig. 3 is a schematic diagram of an estimation result of a single hidden layer BP neural network on a mixed oil length according to an embodiment of the present invention, where an abscissa represents a test sample number, an ordinate represents a mixed oil length (scaled to a 0-1 interval), a dotted line represents a true value of the mixed oil length, and a solid line represents an estimated value of the mixed oil length;
FIG. 4 is a schematic diagram of an estimation result of an extreme learning machine on a mixed oil length, wherein an abscissa represents a test sample number, an ordinate represents the mixed oil length (scaled to a 0-1 interval), a dotted line represents a true value of the mixed oil length, and a solid line represents an estimated value of the mixed oil length;
fig. 5 is a schematic diagram of an estimation result of a depth neural network on a mixed oil length according to an embodiment of the present invention, wherein an abscissa represents a test sample number, an ordinate represents a mixed oil length (scaled to a 0-1 interval), a dotted line represents a true value of the mixed oil length, and a solid line represents an estimated value of the mixed oil length.
Detailed Description
The invention will be further described with reference to the drawings and examples.
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the invention. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the present invention. As used herein, the singular is also intended to include the plural unless the context clearly indicates otherwise, and furthermore, it is to be understood that the terms "comprises" and/or "comprising" when used in this specification are taken to specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof.
The invention discloses an on-line estimation method of the mixed oil length based on a mixed single hidden layer BP neural network, which adopts a principle of dividing and controlling, establishes a nonlinear relation between input and output by a shallow layer neural network in each mode, reduces the requirement on a single shallow layer neural network, and simultaneously avoids using a deep layer neural network with higher model complexity; the weighted combination of the local shallow neural networks can obtain better performance than that of a single shallow neural network and a deep neural network; the regularization coefficient and the number of mixed modes are adaptively determined by adopting a Bayes method, and the overfitting resistance is effectively improved under the condition of no need of manual intervention; the semi-supervised learning framework is adopted, so that a small amount of labeled samples can be utilized, and meanwhile, the potential information of a large amount of unlabeled samples can be fully mined; finally, the method not only can provide the estimated value of the oil mixing length, but also can provide the corresponding confidence interval. By the method, the neural network structure can be effectively determined, the risk of overfitting is reduced, and the unlabeled samples are fully utilized, so that the prediction precision is improved, and technical support and guarantee are provided for improving the product quality, optimizing the production operation and making optimal scheduling decisions.
Example 1
As shown in fig. 1, the embodiment provides an online estimation method for the length of mixed oil based on a mixed single hidden layer neural network, which includes the following steps:
step 1: selection and mixing lengthAssociated auxiliary variable->Wherein->Representing the number of auxiliary variables.
In this embodiment, according to a certain long-distance pipeline, 6 variables with the greatest influence on the length of the mixed oil are selected as auxiliary variables, and are respectively: initial oil mixing amount [ ]) Initial oil blend length (+)>) Inner diameter ()>) Length (+)>) Reynolds number ()>) Equivalent length (+)>) Thus auxiliary variable->I.e. +.>，d=6。
Step 2: collecting historical record data and establishing a model training setThe labeled sample set containing both auxiliary variable and oil blend length is +.>The unlabeled exemplar set containing only auxiliary variables isWherein->And->Representing the number of labeled and unlabeled exemplars, respectively.
In this embodiment, a set 300 of labeled sample sets (denoted as) And a label-free sample set 1278 that contains only auxiliary variables (noted as) As training data set, wherein +.>And->Representing the number of labeled and unlabeled exemplars, respectively.
Step 3: data normalization, data setThe dimensionless treatment is carried out, and the dimensionality removing method comprises the following steps:，/>the method comprises the steps of carrying out a first treatment on the surface of the Wherein,represents the standard deviation of the mth auxiliary variable,indicating the length of the mixed oil->Standard deviation of>Represent the firstnSample numbermSampling values of the auxiliary variables.
Step 4: setting model parameters including the quantity of mixed componentsHidden layer neuron number->Learning rate->Maximum number of iteration steps->Convergence threshold->。
In this embodiment, the number of the first and second electrodes, among others,，/>，/>dynamically adjusting to an initial value of 0.1, and performing current learning 30 times per iterationRate multiplied by 0.1->，/>。
Step 5: initializing shallow neural network weight parametersInitializing model variablesConjugate a priori distribution superparameter->Super parameter of posterior distribution->The method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>Indicate->The mixing ratio of the individual gaussian components; />Andrespectively represent +.>Auxiliary variable +.>A distributed mean vector and an accuracy matrix; />Indicate->Regression between individual mixed modality inputs and outputsCoefficients; />Is->Precision parameters of (a); />Indicate->The accuracy parameters of the noise are measured in the hybrid modes.
The description of the conjugate prior distribution parameter and the posterior distribution parameter is as follows:
and->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />Andrespectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively representA priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />Representation->Posterior distribution parameters of (c).
In this embodiment, the conjugate prior distribution and posterior distribution of the random variable are determined as:
is>And posterior distribution->Are dirichlet distribution, i.e ，/>；Is>And posterior distribution->Are all Gaussian-Weisauter distribution, i.e，；/>Is>And posterior distributionAre all Gaussian distribution, i.e. +.>Wherein->Is thatDimension unit matrix,/->；/>Is>And posterior distributionAre all gamma distributions, i.e.)>，/>；/>Is>And posterior distribution->Are all gamma distributions, i.e.)>，。
In this step, therefore, it is necessary to initialize the parameters of the neural network and the a priori distribution parameters,generating for random initialization->A human giving is required. In this example, the parameters of the a priori distribution are set to，/>，/>Wherein->Is thatdDimension unit matrix,/->。
Step 6: constructing a lower bound of evidence for observed dataIn the model training process, a random variable ++is obtained by maximizing the evidence lower bound>Posterior distribution and model parameters of ∈>Is determined, wherein,/>，/>and->Respectively represent the firstiEach has a label sample and a firstjGaussian component corresponding to each unlabeled sample, < ->，/>And satisfy->，/>The calculation of the lower bound of evidence is:，/>is the posterior distribution of the variables.
In step 6, labeled samples and unlabeled samples，/>，/>And its corresponding hidden variable->The conditional distribution of (2) is: />，，/>，，Wherein-> ， ，/>，/> ，/> ，MThe number of hidden layer neurons is represented,dthe dimensions of the input vector are represented,to activate the function.
Step 7: under the framework of variational reasoning, the lower bound of evidence is maximizedThe method specifically comprises the following steps: learning the random variable in step E->Updating the evidence lower bound of the observation sample, and updating the weight parameter +_in M step through a back propagation algorithm>And (3) alternately performing the step E and the step M until convergence conditions are met. Wherein, step E (prospect Step): expected step, M step (Maximization Step): the maximization step, the convergence condition is: />，/>Is the threshold for convergence.
In step E, firstly, the hidden variable is calculatedAnd->Posterior distribution of->And->Under the variational reasoning framework, it is available: />Wherein->Expressed in random variable function->At->Is +.>Represents an item independent of the current random variable, and +.>The definition is as follows:. The probability density function of the posterior distribution is normalized to obtain: />Wherein, the method comprises the steps of, wherein,。
in the same way, the processing method comprises the steps of,posterior distribution of->The posterior distribution of (2) is calculated as follows:wherein->The definition is as follows: />The method comprises the steps of carrying out a first treatment on the surface of the The probability density function of the posterior distribution is normalized to obtain: />Wherein, the method comprises the steps of, wherein,。
the remaining variablesThe posterior distribution calculation of (1) still adopts a variational reasoning method, and for the simplicity and convenience of calculation and expression, the following statistics are defined firstly: />，/>，/>，。
(a) Variable(s)The posterior distribution of (2) is calculated as follows:variable->The posterior distribution form of +.>And->。/>The expectations of (2) are: />During training some->And->The expected value of the mixing coefficient for the less contributing component of the interpretation dataset will tend to be 0 during the optimization process, these components can be removed from the model by means of an automatic correlation determination, so we can choose a relatively large oneKThe excess components can be pruned from the model to automatically determine the amount of the mixed components.
(b) Variable(s)The posterior distribution of (2) is calculated as follows:variable(s)The posterior distribution of (2) is: />And the parameters in the posterior distribution are calculated as follows: /> ，，/>The updating of posterior distribution of (a) simultaneously uses labeled samples +.>And no label sample->Better mining the information of the unlabeled exemplar under the semi-supervised learning framework.
(c) Variable(s)The posterior distribution of (2) is calculated as follows:variable->The posterior distribution form of +.>And the parameters in the posterior distribution are calculated as follows:，/>。
variable(s)The posterior distribution solving process of (1) is equivalent to minimizing the sum of squares error function plus a quadratic regularization term, and +.>，/>The value of (2) can be automatically determined in the iterative process, so that the regularization coefficient can be adaptively determined, thereby improving the generalization new energy of the model.
(d) Variable(s)The posterior distribution of (2) is calculated as follows:variable->The posterior distribution form of +.>And the parameters in the posterior distribution are calculated as follows: />，。
(e) Variable(s)The posterior distribution of (2) is calculated as follows:variable->The posterior distribution form of (a) isAnd the parameters in the posterior distribution are calculated as follows: />，。
Model optimal parameters in M stepsThe updating of (a) adopts a gradient ascent method, and the lower bound of the maximum evidence likelihood is +.>Wherein->Parameter->The relevant items are: />Wherein->。
By chain ruleFor each Gaussian modality +.>Gradient of (C), particularly calculated asThe following steps:wherein->Representation vector->Is the first of (2)mThe number of elements to be added to the composition,，/>for vector->Is->For matrix->Is the first M rows of elements.
It can be seen that in the kth neural network parameter updating process, although all tagged samples participate in the calculation, some samples are responsible for the kth componentNear 0, this portion of the sample does not play a role in the gradient calculation, only the local sample plays a role.
Further calculateFor each Gaussian modality +.>The gradient of (2) is calculated as follows:wherein, the method comprises the steps of, wherein,，/>，/> ，/> ，/>representing Hadamard product, ->A matrix in which the elements representing all positions are 1, and +.> 。
Updating parameters in each mode by gradient ascent methodHas the following form:wherein->Representing the learning rate.
Step 8: and (3) collecting a sample to be predicted which only contains auxiliary variables, carrying out data standardization according to the step (3), and carrying out online estimation on the oil mixing length according to the model parameters trained in the step (7).
The step 8 specifically includes:
collected sample determination to be predicted comprising only auxiliary variablesMeaning asWherein->For the number of samples to be predicted, the samples to be predicted are normalized according to step 3, eliminate dimension and +.>Introducing hidden variable representing modality information>Wherein->，/>And satisfy->。/>Corresponding hidden variable->The posterior distribution of (2) is:wherein, the method comprises the steps of, wherein,，/>representing student t distribution, ->，/>。
Further normalize the probability density function to obtainTo the following form:。
mixed oil length on-line estimation method based on mixed single hidden layer BP neural network, wherein the mixed oil lengthThe probability distribution of (2) is: />Wherein->，/>，/>。
Therefore, the estimated value of the length of the oil blend can be:。
second, the uncertainty of the estimate is determined by the varianceQuantification: />Confidence interval of two standard deviations is denoted +.>The confidence interval plays an important role in the performance evaluation of the soft measurement model and can provide important basis for production operation optimization in engineering practice.
To verify the effectiveness of the present invention, additional sets of tagged samples 1000 were collected from the historical database as test sample sets, and the length of the oil blend was estimated according to step 8, with the average estimation result shown in fig. 2. Meanwhile, fig. 3, fig. 4 and fig. 5 show the single hidden layer BP neural network, the extreme learning machine and the deep neural network, respectively, and the average estimation result of the oil mixing length. In the gaussian mixture model, the number of mixed components is set to 5, and the number of neurons of the neural network in each modality is set to 200; the number of neurons in the single hidden layer BP neural network and the extreme learning machine model is set to 2000; the number of network layers in the deep neural network model is set to 3 layers, and the number of neurons in each layer is 500, 300 and 100 respectively. It can be seen that the single hidden layer BP neural network, the extreme learning machine and the deep neural network can not handle the problems of multiple modes and scarcity of labeled samples, and the estimated value of the oil mixing length provided by the model is obviously deviated from the true value; although the nonlinear processing capacity of the deep neural network model is stronger, the prediction accuracy is improved to a certain extent, the method is still unsatisfactory, and particularly at 800-1000 samples. In contrast, the method provided by the invention provides a mixed oil length estimation value which substantially accords with the true value of the mixed oil length estimation value in all operation areas.
The estimation accuracy of the invention and the single hidden layer BP neural network, the extreme learning machine and the deep neural network is quantized by adopting Root Mean Square Error (RMSE) and a decision coefficient (R2), and is defined as follows:
，/>the method comprises the steps of carrying out a first treatment on the surface of the Wherein (1)>And->Respectively represent the firsteTrue oil blend length and estimate for each test sample, +.>The average of the true values of all test samples is represented. The RMSE of the method and the single hidden layer BP neural network, the extreme learning machine and the deep neural network provided by the invention are respectively 0.0150, 0.0437, 0.0365 and 0.0318, and the R2 is respectively 0.9874, 0.8933, 0.9256 and 0.9433. Compared with a single hidden layer BP neural network, an extreme learning machine and a deep neural network, the estimation accuracy of the mixed oil length is obviously improved, and the estimation errors are reduced by about 65.68%, 58.9% and 52.83% respectively from the RMSE index.
Example two
The embodiment provides an oil mixing length on-line estimation system based on a mixed single hidden layer neural network, which comprises the following steps:
a variable selection module for selecting an auxiliary variable associated with the length of the mixed oil;
the training data set acquisition module is used for acquiring a labeled sample set which simultaneously comprises an auxiliary variable and a mixed oil length and an unlabeled sample set which only comprises the auxiliary variable as training data sets;
the training module of the oil mixing length online estimation model is used for training the oil mixing length online estimation model based on a training data set to obtain a trained oil mixing length online estimation model, and the training process comprises the following steps: setting input parameters of an oil mixing length on-line estimation model, initializing the prior and posterior distributed super-parameters of each neural network weight parameter and random variable, and learning the random variable in the step E by adopting a variational reasoning principleUpdating the lower evidence bound of the observation sample, updating the weight parameters of each neural network in the step M through a gradient ascending method, and alternately performing the step E and the step M until convergence conditions are met, so as to obtain the posterior distribution of the weight parameters and random variables of each neural network in the mixed oil length online estimation model; and the oil mixing length estimation module is used for obtaining an oil mixing length estimation value and a confidence interval of a sample to be predicted, which only comprise auxiliary variables, based on the trained oil mixing length on-line estimation model.
Further, in the training module of the online estimation model of the oil mixing length, the input parameters of the online estimation model of the oil mixing length include: quantity of the mixed componentsHidden layer neuron number->Learning rate->Maximum number of iteration steps->And Convergence threshold->The method comprises the steps of carrying out a first treatment on the surface of the Initializing weight parameter of shallow neural network +.>The method comprises the steps of carrying out a first treatment on the surface of the The super-parameters of the a priori and posterior distributions of the random variables include: initializing model variables +.>Is a conjugate prior distribution parameter of (b)And posterior distribution parameters->Wherein->Indicate->The mixing ratio of the individual gaussian components; />And->Respectively represent +.>An average value vector and an accuracy matrix of auxiliary variable distribution in each mixed mode; />Indicate->Regression coefficients between the inputs and outputs of the mixed modes; />Is->Precision parameters of (a);is->Measuring accuracy parameters of noise in the mixed modes; />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a);and->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively representA priori distribution parameters and posterior distribution parameters of (a); />Representation->Posterior distribution parameters of (c).
Further, the system also comprises a preprocessing module, specifically: and normalizing the training data set, and performing dimensionless treatment on a labeled sample set containing both the auxiliary variable and the mixed oil length and an unlabeled sample set containing only the auxiliary variable.
Further, in the training module of the online estimation model of the oil mixing length, the step E of learning the random variableUpdating the evidence lower bound of the observation sample, and updating the weight parameters of each neural network in the step M by a gradient ascent method, wherein the method comprises the following steps: in step E, firstly the hidden variable +.>And->Posterior distribution of->And->Obtaining a probability density function of posterior distribution; the posterior distribution calculation of the rest variables still adopts a variational reasoning method to obtain posterior distribution corresponding to each variable; in the M-step model, a gradient ascending method is adopted to maximize the lower bound of evidence likelihood +.>The method comprises the steps of carrying out a first treatment on the surface of the By chain ruleFor each Gaussian modality +.>Gradient calculation of (2); further ask for->For each Gaussian modeIs used to update the parameters in each mode by gradient ascent method>。
The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (6)
1. The mixed oil length on-line estimation method based on the mixed single hidden layer neural network is characterized by comprising the following steps of:
selecting an auxiliary variable associated with the length of the mixed oil;
acquiring a labeled sample set simultaneously containing auxiliary variables and the length of mixed oil and a non-labeled sample set only containing the auxiliary variables as training data sets;
training the oil mixing length on-line estimation model based on a training data set to obtain a trained oil mixing length on-line estimation model, wherein the training process comprises the following steps: the input parameters of the on-line estimation model for setting the oil mixing length comprise the quantity of mixed componentsHidden layer neuron number->Learning rate->Maximum number of iteration steps->And Convergence threshold->Initializing weight parameters of a shallow neural networkAnd model variable->Conjugate prior distribution superparameter of (c)Super parameter of posterior distribution->Learning the random variable +.>Updating the lower bound of evidence of an observation sample, updating the weight parameters of each neural network in the step M through a gradient ascending method, and alternately performing the step E and the step M until convergence conditions are met, so as to obtain the posterior distribution of the weight parameters of each neural network and random variables in the mixed oil length online estimation model;
wherein, step E: expected step, M: the step of maximizing the degree of freedom of the device,indicate->The mixing ratio of the individual gaussian components; />And->Respectively represent +.>An average value vector and an accuracy matrix of auxiliary variable distribution in each mixed mode; />Indicate->Regression coefficients between the inputs and outputs of the mixed modes; />Is->Precision parameters of (a); />Is->Measuring accuracy parameters of noise in the mixed modes; />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a);representation->Posterior distribution parameters of (a);
wherein, the learning of the random variable in step EUpdating the evidence lower bound of the observation sample, and updating the weight parameters of each neural network in the step M by a gradient ascent method, wherein the method comprises the following steps:
first calculate the mark in step ESign sample、/>And no label sample->Corresponding hidden variable->And->Posterior distribution of->And->Obtaining a probability density function of posterior distribution; other variables->The posterior distribution calculation of (1) still adopts a variational reasoning method to obtain posterior distribution corresponding to respective variables; wherein the random variableThe method comprises the steps of carrying out a first treatment on the surface of the Labeled sample and unlabeled sample->、/>、，/>Represent the firstiEach with label input sample,/->Represent the firstiEach with label output sample,/>Represent the firstjLabel-free input samples,>and->Representing the number of labeled and unlabeled samples, respectively,/->Represent the firstiComponent information of individual labeled samples, +.>Represent the firstjComponent information of each unlabeled exemplar and corresponding hidden variable，/>，/>，/>，/>；
In the M-step model, a gradient ascending method is adopted to maximize the lower bound of evidence likelihood；
By chain ruleFor each Gaussian modality +.>Gradient calculation of (2);
further calculateFor each Gaussian modality +.>Is used to update the parameters in each mode by gradient ascent method>，/> ，MThe number of hidden layer neurons is represented,drepresenting the dimension of the input vector;
and obtaining the oil mixing length estimated value and the confidence interval of the sample to be predicted, which only contain auxiliary variables, based on the trained oil mixing length on-line estimation model.
2. The method for estimating the length of a mixed oil based on a mixed single hidden layer neural network according to claim 1, further comprising, after acquiring the training data set: and normalizing the training data set, and performing dimensionless treatment on a labeled sample set containing both the auxiliary variable and the mixed oil length and an unlabeled sample set containing only the auxiliary variable.
3. The method for online estimation of the length of a mixed oil based on a mixed single hidden layer neural network according to claim 1, wherein the mixed oil isWhen the length on-line estimation model is trained, a labeled sample and an unlabeled sample are obtained、/>、And its corresponding hidden variable->，/>，，/>，/>The conditional distribution of (2) is:
，
，
，，，
wherein,represent the firstiEach with label input sample,/->Represent the firstiEach with label output sample,/>Represent the firstjLabel-free input samples,>and->Representing the number of labeled and unlabeled samples, respectively,/->Represent the firstiComponent information of individual labeled samples, +.>Represent the firstjComponent information of individual unlabeled exemplars, +.>Indicating the mixing ratio of->The gaussian distribution is represented by the formula,is the firstkMean vector of the individual gaussian components, +.>To show the firstkPrecision matrix of individual gaussian components, +.>Is the firstkExtended form of the output vector of hidden layer in the individual neural network,/->Is the firstkRegression coefficients between hidden layer and output layer in the individual neural networks,is the firstkThe accuracy parameters of the noise are measured in the individual components.
4. The method for online estimation of a mixed oil length based on a mixed single hidden layer neural network according to claim 1, wherein the samples to be predicted including only auxiliary variables are defined asWherein->For the number of samples to be predicted, normalizing the samples to be predicted, eliminating the dimension, and +.>Introducing hidden variable representing modality information>；
The hidden variableThe posterior distribution of (2) is calculated as:
，
wherein,，/>representing studentstDistribution of->，/>，dRepresenting the number of auxiliary variables, wherein +.>，/>And satisfy->。
5. The mixed oil length on-line estimation system based on the mixed single hidden layer neural network is characterized by comprising the following components:
a variable selection module for selecting an auxiliary variable associated with the length of the mixed oil;
the training data set acquisition module is used for acquiring a labeled sample set which simultaneously comprises an auxiliary variable and a mixed oil length and an unlabeled sample set which only comprises the auxiliary variable as training data sets;
the training module of the oil mixing length online estimation model is used for training the oil mixing length online estimation model based on a training data set to obtain a trained oil mixing length online estimation model, and the training process comprises the following steps: setting the quantity of mixed components of input parameters of an on-line estimation model of the length of mixed oilHidden layer neuron number->Learning rate->Maximum number of iteration steps->And Convergence threshold->Initializing weight parameters of a shallow neural network>And model variable->Conjugate a priori distribution superparameter->Posterior distribution superparameterLearning the random variable +.>Updating the lower bound of evidence of an observation sample, updating the weight parameters of each neural network in the step M through a gradient ascending method, and alternately performing the step E and the step M until convergence conditions are met, so as to obtain the posterior distribution of the weight parameters of each neural network and random variables in the mixed oil length online estimation model;
wherein, step E: expected step, M: the step of maximizing the degree of freedom of the device,indicate->The mixing ratio of the individual gaussian components; />And->Respectively represent +.>An average value vector and an accuracy matrix of auxiliary variable distribution in each mixed mode; />Indicate->Regression coefficients between the inputs and outputs of the mixed modes; />Is->Precision parameters of (a); />Is->Measuring accuracy parameters of noise in the mixed modes; />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a); />And->Respectively indicate->A priori distribution parameters and posterior distribution parameters of (a);representation->Posterior distribution parameters of (a);
wherein, the learning of the random variable in step EUpdating the evidence lower bound of the observation sample, and updating the weight parameters of each neural network in the step M by a gradient ascent method, wherein the method comprises the following steps:
first calculate the labeled sample in step E、/>And no label sample->Corresponding hidden variable->And->Posterior distribution of->And->Obtaining a probability density function of posterior distribution; other variables->The posterior distribution calculation of (1) still adopts a variational reasoning method to obtain posterior distribution corresponding to respective variables; wherein the random variableLabeled sample and unlabeled sample->、/>、，/>Represent the firstiEach with label input sample,/->Represent the firstiEach with label output sample,/>Represent the firstjLabel-free input samples,>and->Representing the number of labeled and unlabeled samples, respectively,/->Represent the firstiComponent information of individual labeled samples, +.>Represent the firstjComponent information of each unlabeled exemplar and corresponding hidden variable，/>，/>，/>，/>；
In the M-step model, a gradient ascending method is adopted to maximize the lower bound of evidence likelihood；
By chain ruleFor each Gaussian modality +.>Gradient calculation of (2);
further calculateFor each Gaussian modality +.>Is used to update the parameters in each mode by gradient ascent method>，/> ，MThe number of hidden layer neurons is represented,drepresenting the dimension of the input vector;
the oil mixing length estimation module is used for obtaining an oil mixing length estimation value based on a sample to be predicted only containing auxiliary variables and a trained oil mixing length online estimation model.
6. The hybrid single hidden layer neural network-based oil blend length online estimation system of claim 5, further comprising a preprocessing module, specifically: and normalizing the training data set, and performing dimensionless treatment on a labeled sample set containing both the auxiliary variable and the mixed oil length and an unlabeled sample set containing only the auxiliary variable.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311411498.6A CN117150931B (en) | 2023-10-30 | 2023-10-30 | Mixed oil length on-line estimation method and system based on mixed single hidden layer neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311411498.6A CN117150931B (en) | 2023-10-30 | 2023-10-30 | Mixed oil length on-line estimation method and system based on mixed single hidden layer neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117150931A CN117150931A (en) | 2023-12-01 |
| CN117150931B true CN117150931B (en) | 2024-01-30 |
Family
ID=88884681
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311411498.6A Active CN117150931B (en) | 2023-10-30 | 2023-10-30 | Mixed oil length on-line estimation method and system based on mixed single hidden layer neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117150931B (en) |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108664706A (en) * | 2018-04-16 | 2018-10-16 | 浙江大学 | A kind of synthetic ammonia process primary reformer oxygen content On-line Estimation method based on semi-supervised Bayes's gauss hybrid models |
| CN109991396A (en) * | 2019-05-06 | 2019-07-09 | 西南石油大学 | A kind of calculation method for the processed oil pipeline mixture spread considering dead leg influence |
| EP3913547A1 (en) * | 2020-05-19 | 2021-11-24 | Robert Bosch GmbH | Modelling input-output relation of computer-controlled entity |
| CN113780958A (en) * | 2021-09-22 | 2021-12-10 | 国家石油天然气管网集团有限公司 | Oil product sequential conveying method, and oil mixing interface tracking method and device |
| CN113780584A (en) * | 2021-09-28 | 2021-12-10 | 京东科技信息技术有限公司 | Label prediction method, apparatus, storage medium and program product |
| CN114065807A (en) * | 2021-10-28 | 2022-02-18 | 河北工程大学 | Monthly runoff prediction method based on variational modal decomposition and Bayesian neural network |
| CN114169459A (en) * | 2021-12-13 | 2022-03-11 | 广东石油化工学院 | Robust soft measurement method based on semi-supervised Bayesian regularization hybrid Student's t model |
| CN115455827A (en) * | 2022-09-14 | 2022-12-09 | 中国石油大学(北京) | Method and system for calculating oil mixing amount of finished oil pipeline based on real-time data |
| CN116307303A (en) * | 2023-05-24 | 2023-06-23 | 中国石油大学(华东) | Mechanism-data dual-drive oil mixing length prediction method and system for finished oil pipeline |
| CN116362153A (en) * | 2023-03-16 | 2023-06-30 | 西南石油大学 | CFD-based calculation method for sequential conveying mixed oil length of reducer pipe finished oil |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| GB201906234D0 (en) * | 2019-05-03 | 2019-06-19 | Microsoft Technology Licensing Llc | An interpretable neural network |
-
2023
- 2023-10-30 CN CN202311411498.6A patent/CN117150931B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108664706A (en) * | 2018-04-16 | 2018-10-16 | 浙江大学 | A kind of synthetic ammonia process primary reformer oxygen content On-line Estimation method based on semi-supervised Bayes's gauss hybrid models |
| CN109991396A (en) * | 2019-05-06 | 2019-07-09 | 西南石油大学 | A kind of calculation method for the processed oil pipeline mixture spread considering dead leg influence |
| EP3913547A1 (en) * | 2020-05-19 | 2021-11-24 | Robert Bosch GmbH | Modelling input-output relation of computer-controlled entity |
| CN113780958A (en) * | 2021-09-22 | 2021-12-10 | 国家石油天然气管网集团有限公司 | Oil product sequential conveying method, and oil mixing interface tracking method and device |
| CN113780584A (en) * | 2021-09-28 | 2021-12-10 | 京东科技信息技术有限公司 | Label prediction method, apparatus, storage medium and program product |
| CN114065807A (en) * | 2021-10-28 | 2022-02-18 | 河北工程大学 | Monthly runoff prediction method based on variational modal decomposition and Bayesian neural network |
| CN114169459A (en) * | 2021-12-13 | 2022-03-11 | 广东石油化工学院 | Robust soft measurement method based on semi-supervised Bayesian regularization hybrid Student's t model |
| CN115455827A (en) * | 2022-09-14 | 2022-12-09 | 中国石油大学(北京) | Method and system for calculating oil mixing amount of finished oil pipeline based on real-time data |
| CN116362153A (en) * | 2023-03-16 | 2023-06-30 | 西南石油大学 | CFD-based calculation method for sequential conveying mixed oil length of reducer pipe finished oil |
| CN116307303A (en) * | 2023-05-24 | 2023-06-23 | 中国石油大学(华东) | Mechanism-data dual-drive oil mixing length prediction method and system for finished oil pipeline |
Non-Patent Citations (6)
| Title |
|---|
| A novel predictive model of mixed oil length of products pipeline driven by traditional model and data;Lei Chen，et al;《Journal of Petroleum Science and Engineering》;第205卷;全文 * |
| Data-Driven Predictive Model Based on Locally Weighted Bayesian Gaussian Regression;Weiming Shao，et al;《2019 IEEE 8th Data Control and Learning Systems Conference》;全文 * |
| Knowledge-informed Variational Bayesian Gaussian mixture regression model for predicting mixed oil length;Ziyun Yuan，et al;《Energy》;全文 * |
| Physics-based Bayesian linear regression model for predicting length of mixed oil;Ziyun Yuan，et al;《Geoenergy Science and Engineering 》;第223卷;全文 * |
| Semisupervised Bayesian Gaussian Mixture Models for Non-Gaussian Soft Sensor;Weiming Shao，et al;《IEEE TRANSACTIONS ON CYBERNETICS》;第51卷(第7期);全文 * |
| 融合机制与高斯混合回归算法的成品油管道顺序输送混油长度预测模型;袁子云，等;《中国石油大学学报( 自然科学版)》;第47卷(第2期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117150931A (en) | 2023-12-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109060001B (en) | Multi-working-condition process soft measurement modeling method based on feature transfer learning | |
| CN111814871B (en) | Image classification method based on reliable weight optimal transmission | |
| CN108900346B (en) | Wireless network flow prediction method based on LSTM network | |
| CN104574220A (en) | Power customer credit assessment method based on least square support vector machine | |
| CN112967088A (en) | Marketing activity prediction model structure and prediction method based on knowledge distillation | |
| CN106022954A (en) | Multiple BP neural network load prediction method based on grey correlation degree | |
| CN115131618A (en) | Semi-supervised image classification method based on causal reasoning | |
| CN108596204B (en) | Improved SCDAE-based semi-supervised modulation mode classification model method | |
| CN111191823A (en) | Production logistics prediction method based on deep learning | |
| CN111079926A (en) | Equipment fault diagnosis method with self-adaptive learning rate based on deep learning | |
| CN112650933A (en) | High-order aggregation-based graph convolution and multi-head attention mechanism conversation recommendation method | |
| CN116187835A (en) | Data-driven-based method and system for estimating theoretical line loss interval of transformer area | |
| CN114596726B (en) | Parking berth prediction method based on interpretable space-time attention mechanism | |
| CN113780420B (en) | GRU-GCN-based method for predicting concentration of dissolved gas in transformer oil | |
| CN113988415A (en) | Medium-and-long-term power load prediction method | |
| CN114266321A (en) | Weak supervision fuzzy clustering algorithm based on unconstrained prior information mode | |
| CN117150931B (en) | Mixed oil length on-line estimation method and system based on mixed single hidden layer neural network | |
| CN117636183A (en) | Small sample remote sensing image classification method based on self-supervision pre-training | |
| CN116521863A (en) | Tag anti-noise text classification method based on semi-supervised learning | |
| CN115641153A (en) | Vehicle price evaluation method based on deep neural network | |
| Gao et al. | Lazy Estimation of Variable Importance for Large Neural Networks | |
| CN115310559A (en) | Target identification method based on Bayes-Transformer neural network model | |
| CN113139464A (en) | Power grid fault detection method | |
| Chan et al. | Reducing Cost of Process Modeling through Multi-source Data Transfer Learning | |
| Srimani et al. | Regression model for Edu-data in technical education system: A linear approach |
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 |