CN113139752A - Quality index prediction method and device - Google Patents

Quality index prediction method and device Download PDF

Info

Publication number
CN113139752A
CN113139752A CN202110530591.3A CN202110530591A CN113139752A CN 113139752 A CN113139752 A CN 113139752A CN 202110530591 A CN202110530591 A CN 202110530591A CN 113139752 A CN113139752 A CN 113139752A
Authority
CN
China
Prior art keywords
training
parameter
meta
parameter theta
loss function
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.)
Pending
Application number
CN202110530591.3A
Other languages
Chinese (zh)
Inventor
杜文莉
钟伟民
钱锋
彭鑫
李智
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
East China University of Science and Technology
Original Assignee
East China University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by East China University of Science and Technology filed Critical East China University of Science and Technology
Priority to CN202110530591.3A priority Critical patent/CN113139752A/en
Publication of CN113139752A publication Critical patent/CN113139752A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • 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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Landscapes

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

Abstract

The invention provides a method for predicting a quality index. The prediction method comprises the following steps: inputting the data to be tested into a model independent meta-learning frame with variable zooming step length; and determining a quality index corresponding to the data to be tested according to the first adaptive parameter corrected in the test stage and the first support set corrected in the test stage. The first adaptive parameter is obtained by using a plurality of test samples in the test stage and performing at least one correction iteration operation according to the first support set which is not corrected in the test stage. The first support set which is not modified by the test stage is generated by selecting a plurality of sample data from the plurality of test samples according to the meta-parameter of the second adaptive parameter which is trained by the training stage. The second adaptive parameter is obtained by using a plurality of training samples in the training stage and performing a plurality of training iteration operations on the initial meta-parameter and the corresponding second support set.

Description

Quality index prediction method and device
Technical Field
The invention belongs to the field of data processing, and particularly discloses a quality index prediction method and a quality index prediction device.
Background
In an industrial production process, there are many quality indexes or key process variables which cannot be directly obtained due to various reasons, and a regression model between the quality indexes and secondary variables which are easy to measure is constructed so as to predict the potential result of the quality index of interest.
Conventional soft-measurement methods include Principal Component Regression (PCR), Partial Least Squares Regression (PLSR), support vector machine (SVR), and the like. Over the last few years, many successful applications of soft measurements have been proposed in the chemical engineering, biochemical engineering, metallurgical and pharmaceutical industries. With the development of technologies such as big data and the like, process data is represented by huge data volume, multiple data sources and high data dimensionality, so that a traditional data-driven method cannot make more accurate prediction due to limited representation and learning capacity. Meanwhile, with the development of deep learning, the advantages of big data can be fully utilized through parallel computing to extract the characterization information of the process data. Therefore, soft-measurements based on deep learning algorithms are receiving increasing attention due to their non-linear extraction capabilities and advantages in large data contexts.
However, in the process of performing soft measurement based on deep learning, variables may change, which causes performance degradation of a model constructed on a training set when applied to a test set, and makes it impossible to predict a quality index of interest more accurately. Taking a continuous catalytic naphtha reforming chemical process (CCR) as an example, fig. 1 is a graph of a prediction result of a neural network model in a training set and a test set in the continuous catalytic naphtha reforming chemical process, and fig. 2 is a scatter diagram of a prediction effect of the neural network model in the training set and the test set in the continuous catalytic naphtha reforming chemical process. Fig. 1 and 2 show that the model trained on the training set can produce valuable predictions on the training set, but the prediction effect on the test set is poor, and the prediction effect on the test set cannot capture the variation trend of the quality index, which is obviously different from that of the reference label.
Instant learning is proposed in the prior art to build local models of samples to reduce the adverse effects of irrelevant samples. However, the original purpose of obtaining a more accurate prediction model through the existing data cannot be achieved well, and the problem of relation change of process variables and quality indexes in the soft measurement process cannot be solved.
Disclosure of Invention
The following presents a simplified summary of one or more aspects in order to provide a basic understanding of such aspects. This summary is not an extensive overview of all contemplated aspects, and is intended to neither identify key or critical elements of all aspects nor delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description that is presented later.
The invention provides a method for predicting a quality index, which is characterized by comprising the following steps of: inputting the data to be tested into a model independent meta-learning frame with variable zooming step length; and in the model independent element learning framework with variable scaling step length, according to the first adaptive parameter a corrected by the testing stagep(θp,gp) And the first support set S modified by the test stagepDetermining a quality index corresponding to the data to be measured, wherein the first adaptive parameter ap(θp,gp) Using a plurality of test samples in the test stage, according to the first support set S without modification in the test stagetPerforming at least one correction iteration operation to obtainpIs the element parameter, g, of the neural network model after the correction iteration operationpA first support set S which is corrected in a test stage for the gradient parameters of the neural network model after correction iterative operationpIs based on the element parameter theta corrected through the test stagepSelecting a plurality of sample data from a plurality of test samples to generate a first support set S without modification during the test phasetIs based on a second adaptive parameter a trained in a training phaset(θt,gt) Of the element parameter thetatSelecting a plurality of sample data from a plurality of test samples to generate, wherein θtIs the element parameter g of the neural network model after t times of training iterative operationtIs the nerveGradient parameter of the network model after t times of training iterative operation, second adaptive parameter at(θt,gt) Using a plurality of training samples in a training stage according to a current meta-parameter theta0～θt-1Corresponding second support set S0～St-1Performing a plurality of training iterations to obtain0～θt-1The element parameters of the neural network model after corresponding training iterative operation are respectively the second support set S0～St-1Are respectively based on the corresponding element parameter theta0～θt-1A plurality of sample data is selected from a plurality of training samples for generation.
In one embodiment, preferably, the training phase comprises the steps of: neural network model f with element parameter thetaθA base model of the model independent meta learning framework as a variable scaling step; for neural network model fθPerforming initialization to determine an initial meta-parameter theta0And a second supporting data set S0～St-1Window size Nt(ii) a Acquiring a plurality of training samples to form a training sample set; according to the initial meta-parameter theta0Selecting N from a set of training samplestTraining samples to generate an initial second support set S0(ii) a Inputting a plurality of training samples into a neural network model fθTo calculate the parameter theta corresponding to the initial parameter0Loss function of
Figure BDA0003067576360000031
According to an initial parameter theta0Loss function
Figure BDA0003067576360000032
Second supporting data set S0And a scaling value alpha of a variable scaling step length, and calculating an element parameter theta after the first training iteration operation1(ii) a And according to the meta-parameter theta1Calculating corresponding loss functions
Figure BDA0003067576360000033
If loss function
Figure BDA0003067576360000034
Greater than or equal to a predetermined loss function threshold
Figure BDA0003067576360000035
According to the meta-parameter theta1And its corresponding second supporting data set S1The element parameter theta after the next training iteration operation is calculated again2And its corresponding loss function
Figure BDA0003067576360000036
And so on until the loss function
Figure BDA0003067576360000037
Less than the loss function threshold
Figure BDA0003067576360000038
Then the element parameter theta after t times of training iterative operationtAnd determining the parameters as the meta parameters trained in the training stage.
In one embodiment, the meta-parameter θ after the first training iteration is preferably calculated1Comprises the following steps: using a loss function
Figure BDA0003067576360000039
Deriving the scaling value alpha, and calculating the local minimum value of the scaling value alpha by using a gradient descent method to serve as a meta-parameter theta1Corresponding optimal scaling value alpha1Wherein the optimum scaling value α1Indicating to initialize a meta-parameter theta0Iteration is a meta-parameter theta1The optimal scaling step size; and according to the initial meta-parameter theta0Loss function
Figure BDA00030675763600000310
Initial second support set S0And an optimum scaling value alpha1Calculating the meta-parameter θ1。
In an embodiment, preferably, the training phase further comprises the steps of: according to the element parameter theta obtained by calculationiFrom training sample setsReselecting NtTraining samples to generate a corresponding second support set SiWherein i is more than or equal to 1 and less than or equal to t-1; according to the meta-parameter thetai-1The element parameter thetaiAnd corresponding optimal scaling value alphaiCalculating a loss function
Figure BDA00030675763600000311
In the second support set SiGradient parameter g ofi(ii) a And according to the meta-parameter thetaiAnd gradient parameter giDetermining adaptive parameter a after i times of training iterative operationi(θi,gi)。
In one embodiment, preferably, the steps and so on include: in response to a loss function
Figure BDA00030675763600000312
Greater than or equal to the loss function threshold
Figure BDA00030675763600000313
According to the meta-parameter thetaiAnd loss function
Figure BDA00030675763600000314
In the second support set SiGradient parameter g ofiCalculating the element parameter theta after the next training iteration operation by using a gradient descent methodi+1Corresponding optimal scaling value alphai+1(ii) a And according to the meta-parameter thetaiLoss function
Figure BDA00030675763600000315
Second support set SiAnd an optimum scaling value alphai+1Calculating the meta-parameter θi+1。
In an embodiment, preferably, the training phase is further provided with a maximum number of iterations M, and the training phase further includes the following steps: judging whether the current iteration number reaches the maximum iteration number M; and responding to the maximum iteration times M reached by the current iteration times, judging the training completion stage, and performing M times of training iteration operations on the element parameter thetaMIs determined to be trainedMeta-parameter of phase training thetat。
In an embodiment, preferably, the training phase further comprises the steps of: dividing a training sample set according to an input task distribution p (T) to determine a plurality of batches of tasks Tb1～TbBWherein each batch task Tb1～TbBComprises a plurality of training samples according to an initial meta-parameter theta0Selecting N from a set of training samplestTraining samples to generate an initial second support set S0Comprises the following steps: according to the initial meta-parameter theta0From the first batch of tasks Tb1Selecting N from the plurality of training samplestTraining samples to generate an initial second support set S0Inputting a plurality of training samples into the neural network model fθTo calculate the parameter theta corresponding to the initial parameter0Loss function of
Figure BDA0003067576360000041
Comprises the following steps: the first batch of tasks Tb1A plurality of training samples are input into a neural network model fθTo calculate the first tasks Tb1Corresponding to the initial parameter theta0Loss function of
Figure BDA0003067576360000042
Calculating the element parameter theta after the first training iteration operation1Comprises the following steps: according to an initial parameter theta0Loss function
Figure BDA0003067576360000043
Second supporting data set S0And corresponding scaling value alphaTb1Computing the first batch of tasks Tb1The element parameter theta after one iterationTb1(ii) a And according to the meta-parameter thetaTbiLoss function
Figure BDA0003067576360000044
Second supporting data set STbiAnd corresponding scaling value alphaTb(i+1)Calculating the rest tasks T one by oneb2～TbBElement after one iteration operationParameter thetaTb(i+1)And the finally obtained element parameter thetaTbBDetermined as the element parameter theta after the first training iteration operation1Wherein i is more than or equal to 1 and less than or equal to B-1.
In one embodiment, the remaining batches of tasks T are preferably computed one by oneb2～TbBThe element parameter theta after the iterative operationTb(i+1)Comprises the following steps: after each iteration operation, each element parameter theta is calculated respectivelyTb2～θTbBCorresponding loss function
Figure BDA0003067576360000045
A value of (d); in response to any loss function
Figure BDA0003067576360000046
Is less than the previous loss function
Figure BDA0003067576360000047
The current iteration is judged to be effective, and the corresponding element parameter theta is recordedTb(i+1)(ii) a And in response to any loss function
Figure BDA0003067576360000048
Is greater than or equal to the previous loss function
Figure BDA0003067576360000049
The scaling value a of the current iteration is determinedTb(i+1)Too large, will scale the value aTb(i+1)Halving and carrying out the next iteration until the maximum number of iterations is reached or until the local minimum value alpha is converged1。
In one embodiment, preferably, the testing phase comprises the steps of: determining a first supporting data set SpWindow size Np(ii) a Obtaining a plurality of test samples to form a test sample set; according to a second adaptive parameter at(θt,gt) Of the element parameter thetatSelecting N from the test sample setpA test sample to generate a first support set S without modification during the test phaset(ii) a Testing a plurality of test samplesInput neural network model fθTo calculate the corresponding element parameter thetatLoss function of
Figure BDA00030675763600000410
According to the meta-parameter thetatLoss function
Figure BDA00030675763600000411
First support data set StAnd corresponding variable scaling step size scaling value alphatCalculating the element parameter theta after the first correction iteration operationp(ii) a According to the meta-parameter thetapReselecting N from a test sample setpA test sample to generate a first support set S modified by a test phasep(ii) a According to the meta-parameter thetatThe element parameter thetapAnd corresponding optimal scaling value alphapCalculating a loss function
Figure BDA00030675763600000412
In the first support set SpGradient parameter g ofp(ii) a And according to the meta-parameter thetapAnd gradient parameter gpDetermining a first adaptive parameter a corrected in a test stagep(θp,gp)。
In an embodiment, preferably, the training phase and the testing phase further include the following steps: acquiring a plurality of key variable data of a chemical process, wherein the chemical process comprises a continuous catalytic naphtha reforming process, and the key variable data comprise input variable data and output variable data; preprocessing a plurality of key variable data according to a 3 sigma criterion to remove abnormal values and outliers; and dividing the preprocessed multiple key variable data into multiple training samples and multiple testing samples according to a preset proportion.
In one embodiment, the data to be tested is input variable data of a continuous catalytic naphtha reforming process, and the step of determining a quality index corresponding to the data to be tested comprises: in the model independent element learning framework with variable zooming step length, according to the first adaptive parameter a corrected by the testing stagep(θp,gp) And the first support set S modified by the test stagepOutput variable data corresponding to the input variable data is predicted.
The present invention also provides a quality index prediction device, including: a memory; and a processor connected to the memory and configured to implement any of the above quality index prediction methods.
The invention also provides a computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions, when executed by a processor, implement any of the above quality index prediction methods.
Drawings
The above features and advantages of the present disclosure will be better understood upon reading the detailed description of embodiments of the disclosure in conjunction with the following drawings. In the drawings, components are not necessarily drawn to scale, and components having similar relative characteristics or features may have the same or similar reference numerals.
FIG. 1 is a graph of the predicted results of a neural network in a training set and a test set during a continuous catalytic naphtha reforming chemical process;
FIG. 2 is a scatter plot of the predicted effect of neural networks in training and testing sets during continuous catalytic naphtha reforming chemical processes;
FIG. 3 is a flowchart illustrating a prediction method during a training phase and a testing phase according to an embodiment of the present invention;
FIG. 4 is a flow chart illustrating a method for predicting a quality indicator according to an aspect of the present invention;
FIG. 5 is a graph of predicted results for three quality index prediction methods;
FIG. 6 is a graph of error comparisons for three quality index prediction methods; and
fig. 7 is a schematic structural diagram of a quality index prediction apparatus according to another aspect of the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided for illustrative purposes, and other advantages and effects of the present invention will become apparent to those skilled in the art from the present disclosure. While the invention will be described in connection with the preferred embodiments, there is no intent to limit its features to those embodiments. On the contrary, the invention is described in connection with the embodiments for the purpose of covering alternatives or modifications that may be extended based on the claims of the present invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. The invention may be practiced without these particulars. Moreover, some of the specific details have been left out of the description in order to avoid obscuring or obscuring the focus of the present invention.
In the description of the present invention, it should be noted that, unless otherwise explicitly specified or limited, the terms "mounted," "connected," and "connected" are to be construed broadly, e.g., as meaning either a fixed connection, a removable connection, or an integral connection; can be mechanically or electrically connected; they may be connected directly or indirectly through intervening media, or they may be interconnected between two elements. The specific meanings of the above terms in the present invention can be understood in specific cases to those skilled in the art.
Additionally, the terms "upper," "lower," "left," "right," "top," "bottom," "horizontal," "vertical" and the like as used in the following description are to be understood as referring to the segment and the associated drawings in the illustrated orientation. The relative terms are used for convenience of description only and do not imply that the described apparatus should be constructed or operated in a particular orientation and therefore should not be construed as limiting the invention.
It will be understood that, although the terms first, second, third, etc. may be used herein to describe various elements, regions, layers and/or sections, these elements, regions, layers and/or sections should not be limited by these terms, but rather are used to distinguish one element, region, layer and/or section from another element, region, layer and/or section. Thus, a first component, region, layer or section discussed below could be termed a second component, region, layer or section without departing from some embodiments of the present invention.
Aiming at the problem of relation change of a quality index and a process variable in the prior art, the invention provides a prediction method of the quality index.
FIG. 3 is a flowchart illustrating a prediction method during a training phase and a testing phase according to an embodiment of the present invention.
Referring to fig. 3, the quality index prediction method provided by the present invention adopts different method steps in the training phase and the prediction phase.
As shown in fig. 3, the training phase includes:
step 301: a plurality of training samples are acquired to form a training sample set.
The training samples here can be obtained by collecting raw data and performing corresponding pre-processing on it. In one embodiment, the raw data is a plurality of key variable data for a chemical process, such as a continuous catalytic naphtha reforming process (CCR). The following table gives the physical meaning of key variable data in the examples of a continuous catalytic naphtha reforming process.
TABLE 1 continuous catalytic naphtha reforming process key variable data
Figure BDA0003067576360000071
Figure BDA0003067576360000081
Figure BDA0003067576360000091
Figure BDA0003067576360000101
In this embodiment, the 84 physical quantities are sampled, and about 10000 samples are selected as a training sample set, and since the raw data contains measurement noise, a corresponding preprocessing operation is required before the subsequent steps are performed. For example, a non-positive value and other outliers of the predictive label are removed through a 3-sigma criterion to obtain preprocessed data, and the preprocessed key variable data are divided into a plurality of training samples and a plurality of test samples according to a preset proportion to respectively form a training sample set and a test sample set.
As shown in fig. 3, after the training sample set is constructed, the quality index prediction method further includes:
step 302: initializing a neural network model fθDetermining an initial meta-parameter θ0And a second support set StWindow size Nt；
Step 303: according to the initial meta-parameter theta0Selecting N from a set of training samplestTraining samples to generate an initial second support set S0(ii) a And
step 304: inputting a plurality of training samples into a neural network model fθTo calculate the parameter theta corresponding to the initial parameter0Loss function of
Figure BDA0003067576360000102
First, a neural network model f having a meta-parameter θ is modeledθA basic model of the model independent meta learning framework as a variable scaling step. We want to learn an initial θ ═ θ0Make the pair of supporting sets SbIs subjected to a small number N of gradient updates to obtain thetaNThe network then targets the set T of tasksbThe above table performed well. Where b is an index of a particular support set task in a collection of support set tasks. This set of N update steps is referred to as an inner loop update process. In the slave support task SbAfter acquiring the data, the updated underlying network parameters may be represented as:
Figure BDA0003067576360000103
where a is the learning rate, where a is,
Figure BDA0003067576360000104
is the adaptation parameter of the neural network after task b has been adapted i times,
Figure BDA0003067576360000105
is the loss function of the support set of b after (i-1) updates (i.e., the previous step), which is also referred to as an inner loop process. The meta-object may be represented as:
Figure BDA0003067576360000111
where B represents the batch size of the task. Thus theta and theta0The relationship of (c) is given by the above formula. Formula based on using this initialization θ in all tasks0The total loss of the initialization is a measure of the quality of the initialization. The meta-objective function is minimized to optimize the initial parameter value θ0Such that this parameter contains cross-task knowledge.
As shown in fig. 3, the quality index prediction method further includes:
step 305: according to an initial parameter theta0Loss function
Figure BDA0003067576360000112
Second supporting data set S0And a scaling value alpha of a variable scaling step length, and calculating an element parameter theta after the first training iteration operation1。
As shown in fig. 1 and fig. 2, it is found in experiments that if the same number of iterations is used in the prediction process as in the training process, the problem of inaccurate prediction result is common. In order to solve the problem, the invention provides a unified adaptation method under different stages, namely, a stage-related adaptation module is used for summarizing the existing meta-learning adaptation method, which mainly represents an approximate strategy, local optimal parameters related to batch tasks are usually achieved through enough iteration in a test stage, and an effective estimation is obtained through an estimation strategy in a training stage.
The assumption that MAML is followed is that the adaptation parameters can be estimated by a gradient descent during the training phase and a variable scaling step is set in the training adaptation method. Therefore, the sub-goal of the training phase is to find the optimal scaling value.
Therefore, we use the loss function
Figure BDA0003067576360000113
Deriving the scaling value alpha, and calculating the local minimum value of the scaling value alpha by using a gradient descent method to serve as a meta-parameter theta1Corresponding optimal scaling value alpha1Wherein the optimum scaling value α1Indicating to initialize a meta-parameter theta0Iteration is a meta-parameter theta1The optimal scaling step size; then according to the initial meta-parameter theta0Loss function
Figure BDA0003067576360000114
Initial second support set S0And an optimum scaling value alpha1Calculating the meta-parameter θ1。
The specific process is as follows: since the optimal scaling value is not a scalar, a method for meta-learning with single-step adaptation by finding the optimal adaptation parameters associated with the support set during the training phase is proposed. The adaptive parameters of the training mode are constrained by the gradient direction, and the adaptive size is a variable scaling value. The optimal scaling value is a free variable λ, estimated from the support data in the inner loop of the training phase. Thus, our training phase adaptation method can be expressed as the following formula:
Figure BDA0003067576360000115
Figure BDA0003067576360000116
Figure BDA0003067576360000117
at(θ0,f,Tb)＝θa,ga
to estimate the optimal scaling value during the training phase, we consider the variable scaling value as an additional parameter λ and in practice, let us note θsFor adaptive parameters derived from the scaling values
Figure BDA0003067576360000121
Then our optimization goals are:
Figure BDA0003067576360000122
we express the adapted parameters as a function of the scaled value:
Figure BDA0003067576360000123
where α is the scaling value. We then sub-target to find the best alpha. To estimate this optimal scaling value, an initial scaling value may be initialized and then a locally optimal solution may be found by gradient descent. The derivative of the loss function with respect to the adaptation step size is expressed in the form:
Figure BDA0003067576360000124
wherein theta is0For the initial parameters of the network, α is the adaptation step scalar, θsAnd (alpha) is an adaptive parameter of the initial parameter under the adaptive step size of the support set.
Therefore, a local minimum value can be obtained by a gradient descent method and is continuously updated, and the value can be estimated more accurately than a scalar quantity which is independent of data. The parameter updating formula is as follows:
Figure BDA0003067576360000125
Figure BDA0003067576360000126
where α is the current scaling value, αlrIs the learning rate of the scaled value, alphanewIs a scaled value, theta, after a gradient update according to alphas(αnew) Is an updated estimate of the adaptive parameter. To speed up the search for the optimal scaling value, we record the value of the adaptation loss function in the inner loop. If the loss function with respect to the adaptation parameter drops, then the adaptation is valid and we record the corresponding adaptation parameter. Otherwise, the scaling step is considered to be too large, the scaling step is halved, and the next iteration is carried out. Until a maximum number of iterations is reached and a local minimum is converged.
As shown in fig. 3, the quality index prediction method further includes:
step 306: according to the meta-parameter theta1Calculating corresponding loss functions
Figure BDA0003067576360000127
If loss function
Figure BDA0003067576360000128
Greater than or equal to a predetermined loss function threshold
Figure BDA0003067576360000129
According to the meta-parameter theta1And its corresponding second supporting data set S1The element parameter theta after the next training iteration operation is calculated again2And its corresponding loss function
Figure BDA0003067576360000131
And so on.
According to the method, the theta can be calculated by repeating the steps1,θ2… obtaining a meta-parameter theta based on the calculationiReselecting N from a training sample settTraining samples to generate a corresponding second support set SiWherein i is more than or equal to 1 and less than or equal to t-1; according to the meta-parameter thetai-1The element parameter thetaiAnd corresponding optimal scaling value alphaiCalculating a loss function
Figure BDA0003067576360000132
In the second support set SiGradient parameter g ofi(ii) a And according to the meta-parameter thetaiAnd gradient parameter giDetermining adaptive parameter alpha after i times of training iterative operationi(θi,gi)。
As shown in fig. 3, the quality index prediction method further includes:
step 307: determining a loss function
Figure BDA0003067576360000133
Whether or not less than a loss function threshold
Figure BDA0003067576360000134
Step 308: the element parameter theta after t times of training iterative operationtAnd determining the parameters as the meta parameters trained in the training stage.
In response to a loss function
Figure BDA0003067576360000135
Greater than or equal to the loss function threshold
Figure BDA0003067576360000136
According to the meta-parameter thetaiAnd loss function
Figure BDA0003067576360000137
In the second support set SiGradient parameter g ofiCalculating the element parameter theta after the next training iteration operation by using a gradient descent methodi+1Corresponding optimal scaling value alphai+1(ii) a And according to the meta-parameter thetaiLoss function
Figure BDA0003067576360000138
Second support set SiAnd an optimum scaling value alphai+1Calculating the meta-parameter θi+1I.e., back to step 306 until the condition is satisfied.
In response to a loss function
Figure BDA0003067576360000139
Less than the loss function threshold
Figure BDA00030675763600001310
Then the element parameter theta after t times of training iterative operationtAnd determining the parameters as the meta parameters trained in the training stage.
In an embodiment, the training phase is further provided with a maximum number of iterations M, and the training phase further comprises the following steps: judging whether the current iteration number reaches the maximum iteration number M; and responding to the maximum iteration times M reached by the current iteration times, judging the training completion stage, and performing M times of training iteration operations on the element parameter thetaMDetermining as the element parameter theta trained by the training staget。
In an embodiment, the tasks may also be batched to perform batch processing. Dividing a training sample set according to an input task distribution p (T) to determine a plurality of batches of tasks Tb1～TbBWherein each batch task Tb1～TbBComprises a plurality of training samples;
according to the initial meta-parameter theta0Selecting N from a set of training samplestTraining samples to generate an initial second support set S0Comprises the following steps: according to the initial meta-parameter theta0From the first batch of tasks Tb1Selecting N from the plurality of training samplestTraining samples to generate an initial second support set S0；
Inputting a plurality of training samples into a neural network model fθTo calculate the parameter theta corresponding to the initial parameter0Loss function of
Figure BDA0003067576360000141
Comprises the following steps: the first batch of tasks Tb1A plurality of training samples are input into a neural network model fθTo calculate the first tasks Tb1Corresponding to the initial parameter theta0Loss function of
Figure BDA0003067576360000142
Calculating the element parameter theta after the first training iteration operation1Comprises the following steps:
according to an initial parameter theta0Loss function
Figure BDA0003067576360000143
Second supporting data set S0And corresponding scaling value alphaTb1Computing the first batch of tasks Tb1The element parameter theta after one iterationTb1(ii) a And
according to the meta-parameter thetaTbiLoss function
Figure BDA0003067576360000144
Second supporting data set STbiAnd corresponding scaling value alphaTb(i+1)Calculating the rest tasks T one by oneb2～TbBThe element parameter theta after one iterationTb(i+1)And the finally obtained element parameter thetaTbBDetermined as the element parameter theta after the first training iteration operation1Wherein i is more than or equal to 1 and less than or equal to B-1.
In one embodiment, task T for each of the remaining batches is computed one by oneb2～TbBThe element parameter theta after the iterative operationTb(i+1)Comprises the following steps: after each iteration operation, each element parameter theta is calculated respectivelyTb2～θTbBCorresponding loss function
Figure BDA0003067576360000145
A value of (d);
in response to any loss function
Figure BDA0003067576360000146
Is less than the previous loss function
Figure BDA0003067576360000147
The current iteration is judged to be effective, and the corresponding element parameter theta is recordedTb(i+1)(ii) a And
in response to any loss function
Figure BDA0003067576360000148
Is greater than or equal to the previous loss function
Figure BDA0003067576360000149
The scaling value a of the current iteration is determinedTb(i+1)Too large, will scale the value aTb(i+1)Halving and carrying out the next iteration until the maximum number of iterations is reached or until the local minimum value alpha is converged1。
The method adopts the optimal scalable step length to replace the fixed step length in the traditional model independent learning training stage, so that only one iteration is used for replacing infinite iterations in the traditional method in the prediction stage, and the effect of simplifying the test process is achieved.
Obtaining the element parameter theta after training in the training stagetAnd then entering a testing phase, as shown in fig. 3, wherein the testing phase comprises the following steps:
step 309: determining a first supporting data set SpWindow size Np(ii) a Obtaining a plurality of test samples to form a test sample set;
step 310: according to a second adaptive parameter at(θt,gt) Of the element parameter thetatSelecting N from the test sample setpA test sample to generate a first support set S without modification during the test phaset(ii) a And inputting a plurality of test samples into the neural network model fθTo calculate the corresponding element parameter thetatLoss function of
Figure BDA00030675763600001410
And according to the meta-parameter thetatLoss function
Figure BDA00030675763600001411
First support data set StAnd corresponding variable scaling step size scaling value alphatCalculating the element parameter theta after the first correction iteration operationp；
Step 311: according to the meta-parameter thetapReselecting N from a test sample setpA test sample to generate a first support set S modified by a test phasep(ii) a According to the meta-parameter thetatThe element parameter thetapAnd corresponding optimal scaling value alphapCalculating a loss function
Figure BDA0003067576360000151
In the first support set SpGradient parameter g ofp(ii) a And
step 312: according to the meta-parameter thetapAnd gradient parameter gqDetermining a first adaptive parameter a corrected in a test stagep(θp,gp)。
The method of the adaptation module related to the above stage can construct the relation between the adaptation parameters and the initial parameters through the constraint, so that the gradient update of the initial parameters by the adapted parameters is effective.
Fig. 4 is a flow chart illustrating a method for predicting a quality indicator according to an aspect of the invention.
As shown in fig. 4, the quality index prediction method provided by the present invention includes:
step 401: inputting the data to be tested into a model independent meta-learning frame with variable zooming step length; and
step 402: in the model independent element learning framework with variable zooming step length, according to the first adaptive parameter a corrected by the testing stagep(θp,gp) And the first support set S modified by the test stagepAnd determining a quality index corresponding to the data to be measured.
Wherein the first adaptive parameter ap(θp,gp) Using a plurality of test samples in the test stage, according to the first support set S without modification in the test stagetPerforming at least one correction iteration operation to obtainpIs the element parameter, g, of the neural network model after the correction iteration operationpThe gradient parameters of the neural network model after the correction iteration operation,
first support set S modified by test stagepIs based on the element parameter theta corrected through the test stagepSelecting a plurality of sample data from a plurality of test samples to generate,
first support set S without test phase modificationtIs based on a second adaptive parameter a trained in a training phaset(θt,gt) Of the element parameter thetatSelecting a plurality of sample data from a plurality of test samples to generate, wherein θtIs the element parameter g of the neural network model after t times of training iterative operationtIs the gradient parameter of the neural network model after t times of training iterative operations,
second adaptive parameter at(θt,gt) Using a plurality of training samples in a training stage according to a current meta-parameter theta0～θt-1Corresponding second support set S0～St-1Performing a plurality of training iterations to obtain0～θt-1The element parameters of the neural network model after corresponding training iterative operation are respectively the second support set S0～St-1Are respectively based on the corresponding element parameter theta0～θt-1A plurality of sample data is selected from a plurality of training samples for generation.
The step of determining a quality indicator corresponding to the data to be measured comprises: in the model independent element learning framework with variable zooming step length, according to the first adaptive parameter a corrected by the testing stagep(θp,gp) And the first support set S modified by the test stagepOutput variable data corresponding to the input variable data is predicted.
In the above example of a continuous catalytic naphtha reforming process (CCR), the quality index or output variable data is RON Barrel (octane bucket number).
While, for purposes of simplicity of explanation, the methodologies are shown and described as a series of acts, it is to be understood and appreciated that the methodologies are not limited by the order of acts, as some acts may, in accordance with one or more embodiments, occur in different orders and/or concurrently with other acts from that shown and described herein or not shown and described herein, as would be understood by one skilled in the art.
FIG. 5 is a graph of predicted results for three quality index prediction methods; fig. 6 is a diagram of error comparison of three quality index prediction methods.
As shown in FIG. 6, RMSE and R can be used2To measure the accuracy of the method
Figure BDA0003067576360000161
Figure BDA0003067576360000162
More accurate calculation data are shown in tables 2 and 3 below:
table 2: prediction results on test set
RMSE-T R2-T RMSE-P R2-P
MAML 0.5740 -0.6701 0.1783 0.8774
Reptile 0.1913 0.8498 0.1769 0.8777
MAML-OASV 0.2727 0.8269 0.1388 0.9200
Table 3: prediction results on a test set by a training phase and prediction phase adaptation method for meta learning
PLS NN MAML Reptile MAML-OASV
RMSE 0.5855 0.7060 0.1787 0.1769 0.1388
R2 -3.9751 -3.7765 0.8774 0.8777 0.9200
As can be seen from fig. 5 and 6 and tables 2 and 3, compared with other model-independent meta-learning methods, such as MAML and replay, the model-independent meta-learning method based on the optimal scalable step size proposed by the present invention MAML-OASV better solves the problem of variation of the variable relationship in the industrial process, and achieves the most effective prediction effect.
Fig. 7 is a schematic structural diagram of a quality index prediction apparatus according to another aspect of the present invention.
As shown in fig. 7, the present invention further provides a quality index prediction apparatus 700, which includes a memory 701 and a controller 702 connected thereto, where the controller 702 is configured to implement the steps of any one of the quality index prediction methods described above.
Although the controller 702 of the above-described embodiment may be implemented by a combination of software and hardware. It is understood that the controller 702 may be implemented in software or hardware. For a hardware implementation, the controller 702 may be implemented in one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), digital signal processing devices (DAPDs), Programmable Logic Devices (PLDs), Field Programmable Gate Arrays (FPGAs), processors, controllers, micro-controllers, microprocessors, other electronic devices designed to perform the functions described herein, or a selected combination thereof. For a software implementation, the controller 702 may be implemented by separate software modules running on a common chip, such as program modules (processes) and function modules (functions), each of which performs one or more of the functions and operations described herein.
The present invention also provides an embodiment of a computer readable medium having stored thereon computer instructions which, when executed by a processor, perform the steps of any of the quality indicator prediction methods described above.
Those of skill in the art would understand that information, signals, and data may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits (bits), symbols, and chips that may be referenced throughout the above description may be represented by voltages, currents, electromagnetic waves, magnetic fields or particles, optical fields or particles, or any combination thereof.
Those of skill would further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The various illustrative logical modules, and circuits described in connection with the embodiments disclosed herein may be implemented or performed with a general purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration.
The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. A software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium may be integral to the processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In the alternative, the processor and the storage medium may reside as discrete components in a user terminal.
In one or more exemplary embodiments, the functions described may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functions may be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media includes both computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. A storage media may be any available media that can be accessed by a computer. By way of example, and not limitation, such computer-readable media can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a web site, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, Digital Subscriber Line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of medium. Disk (disk) and disc (disc), as used herein, includes Compact Disc (CD), laser disc, optical disc, Digital Versatile Disc (DVD), floppy disk and blu-ray disc where disks (disks) usually reproduce data magnetically, while discs (discs) reproduce data optically with lasers. Combinations of the above should also be included within the scope of computer-readable media.
The previous description of the disclosure is provided to enable any person skilled in the art to make or use the disclosure. Various modifications to the disclosure will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other variations without departing from the spirit or scope of the disclosure. Thus, the disclosure is not intended to be limited to the examples and designs described herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (13)

1. A method for predicting a quality index, comprising the steps of:
inputting the data to be tested into a model independent meta-learning frame with variable zooming step length; and
in the model independent element learning framework with variable zooming step length, according to the first adaptive parameter a corrected by the testing stagep(θp，gp) And a first support set S modified by the test stagepDetermining a quality indicator corresponding to the data to be tested, wherein,
the first adaptive parameter ap(θp，gp) Using a plurality of test samples in the test phase according to a first support set S which is not modified in the test phasetPerforming at least one correction iteration operation to obtain the value of thetapThe g is the element parameter of the neural network model after the correction iteration operationpAs the neural networkThe gradient parameters of the model after the correction iteration operation,
the first support set S modified by the test stagepIs based on the element parameter theta corrected by the test stagepSelecting a plurality of sample data from the plurality of test samples to generate,
the first support set S which is not modified in the test stagetIs based on a second adaptive parameter a trained in a training phaset(θt，gt) Of the element parameter thetatSelecting a plurality of sample data from the plurality of test samples to generate, wherein θtThe g is the element parameter of the neural network model after t times of training iterative operationtGradient parameters of the neural network model after t times of training iterative operations,
the second adaptive parameter at(θt，gt) Using a plurality of training samples in the training stage according to the current meta-parameter theta0～θt-1Corresponding second support set S0～St-1Performing a plurality of times of the training iteration operation to obtain, wherein the theta0～θt-1Each second support set S is the element parameter of the neural network model after the corresponding training iterative operation0～St-1Are respectively based on the corresponding element parameter theta0～θt-1Selecting a plurality of sample data from the plurality of training samples to generate.
2. The prediction method of claim 1, wherein the training phase comprises the steps of:
the neural network model f with element parameter thetaθA base model as a model-independent meta-learning framework for the variable scaling step;
for the neural network model fθPerforming initialization to determine an initial meta-parameter theta0And said second supporting data set S0～St-1Window size Nt；
Obtaining the plurality of training samples to form a training sample set;
according to the initial meta-parameter theta0Selecting N from the training sample settTraining samples to generate an initial second support set S0；
Inputting the plurality of training samples into the neural network model fθTo calculate the initial parameter theta0Loss function of
Figure FDA0003067576350000021
According to the initial parameter theta0The loss function
Figure FDA0003067576350000022
The second supporting data set S0And a scaling value alpha of a variable scaling step length, and calculating an element parameter theta after the first training iteration operation1(ii) a And
according to the meta-parameter theta1Calculating corresponding loss functions
Figure FDA0003067576350000023
If the loss function
Figure FDA0003067576350000024
Greater than or equal to a predetermined loss function threshold
Figure FDA0003067576350000025
According to said meta-parameter θ1And its corresponding second supporting data set S1The element parameter theta after the next training iteration operation is calculated again2And its corresponding loss function
Figure FDA0003067576350000026
And so on until the loss function
Figure FDA0003067576350000027
Less than the loss function threshold
Figure FDA0003067576350000028
Then the element parameter theta after the training iterative operation for t times is usedtAnd determining the meta-parameters as the meta-parameters trained in the training phase.
3. The prediction method of claim 2, wherein the computing of the meta-parameter θ after the first training iteration is performed1Comprises the following steps:
using said loss function
Figure FDA0003067576350000029
Deriving the scaling value alpha, and calculating a local minimum value of the scaling value alpha by using a gradient descent method to serve as the element parameter theta1Corresponding optimal scaling value alpha1Wherein the optimal scaling value α1Indicating to use the initial meta-parameter θ0The iteration is the meta-parameter theta1The optimal scaling step size; and
according to the initial meta-parameter theta0The loss function
Figure FDA00030675763500000210
The initial second support set S0And the optimal scaling value alpha1Calculating said meta-parameter θ1。
4. The prediction method of claim 3, wherein the training phase further comprises the steps of:
according to the element parameter theta obtained by calculationiReselecting N from the training sample settTraining samples to generate a corresponding second support set SiWherein i is more than or equal to 1 and less than or equal to t-1;
according to the meta-parameter thetai-1The meta-parameter θiAnd corresponding optimal scaling value alphaiCalculating a loss function
Figure FDA00030675763500000211
In the second support set SiGradient parameter g ofi(ii) a And
according to the meta-parameter thetaiAnd the gradient parameter giDetermining the adaptive parameter a after i times of the training iteration operationi(θi，gi)。
5. The prediction method of claim 4, wherein the step of repeating the steps comprises:
in response to a loss function
Figure FDA0003067576350000031
Greater than or equal to the loss function threshold
Figure FDA0003067576350000032
According to the meta-parameter thetaiAnd the loss function
Figure FDA0003067576350000033
In the second support set SiGradient parameter g ofiCalculating the element parameter theta after the next training iteration operation by using the gradient descent methodi+1Corresponding optimal scaling value alphai+1(ii) a And
according to the meta-parameter thetaiThe loss function
Figure FDA0003067576350000034
The second support set SiAnd the optimal scaling value alphai+1Calculating said meta-parameter θi+1。
6. The prediction method of claim 2, wherein the training phase is further provided with a maximum number of iterations M, the training phase further comprising the steps of:
judging whether the current iteration number reaches the maximum iteration number M; and
responding to the current iteration times reaching the maximum iteration times M, judging to finish the training stage, and performing the training iteration operations for M times to obtain an element parameter thetaMDetermining the element parameter theta as trained by the training staget。
7. The prediction method of claim 2, wherein the training phase further comprises the steps of: dividing the training sample set according to the input task distribution p (T) to determine a plurality of batches of tasks Tb1～TbBWherein each of the batch tasks Tb1～TbBA plurality of said training samples are included,
the parameter theta according to the initial element0Selecting N from the training sample settTraining samples to generate an initial second support set S0Comprises the following steps: according to the initial meta-parameter theta0From the first batch of tasks Tb1Selecting N from a plurality of said training samplestTraining samples to generate an initial second support set S0，
The inputting of the plurality of training samples into the neural network model fθTo calculate the initial parameter theta0Loss function of
Figure FDA0003067576350000035
Comprises the following steps: the first batch of tasks Tb1A plurality of the training samples are input into the neural network model fθTo calculate said first plurality of tasks Tb1Corresponding to the initial parameter theta0Loss function of
Figure FDA0003067576350000036
The element parameter theta after the first training iteration operation is calculated1Comprises the following steps:
according to the initial parameter theta0The loss function
Figure FDA0003067576350000037
The second supporting data set S0And corresponding scaling value alphaTb1Computing the first batch of tasks Tb1The element parameter theta after one iterationTb1(ii) a And
according to the meta-parameter thetaTbiLoss function
Figure FDA0003067576350000041
Second supporting data set STbiAnd corresponding scaling value alphaTb(i+1)Calculating the rest of the batch tasks T one by oneb2～TbBThe element parameter theta after one iteration operationTb(i+1)And the finally obtained element parameter thetaTbBIs determined as the element parameter theta after the first training iteration operation1Wherein i is more than or equal to 1 and less than or equal to B-1.
8. The prediction method of claim 7, wherein said calculating one by one is for each remaining of said plurality of tasks Tb2～TbBThe element parameter theta after the iterative operationTb(i+1)Comprises the following steps:
after each iteration operation is carried out, each element parameter theta is calculated respectivelyTb2～θTbBCorresponding loss function
Figure FDA0003067576350000042
A value of (d);
in response to any loss function
Figure FDA0003067576350000043
Is less than the previous loss function
Figure FDA0003067576350000044
The current iteration is judged to be effective, and the corresponding element parameter theta is recordedTb(i+1)(ii) a And
in response to any loss function
Figure FDA0003067576350000045
Is greater than or equal to the previous loss function
Figure FDA0003067576350000046
The scaling value a of the current iteration is determinedTb(i+1)Too large, the scaling value a is setTb(i+1)Halving and carrying out the next iteration until the maximum number of iterations is reached or until the local minimum value alpha is converged1。
9. The prediction method of claim 1, wherein the testing phase comprises the steps of:
determining the first supporting data set SpWindow size Np；
Obtaining the plurality of test samples to form a test sample set;
according to the second adaptive parameter at(θt，gt) Of the element parameter thetatSelecting N from said set of test samplespA test sample to generate the first support set S without the test phase modificationt；
Inputting the plurality of test samples into the neural network model fθTo calculate the corresponding meta-parameter thetatLoss function of
Figure FDA0003067576350000047
According to the meta-parameter thetatThe loss function
Figure FDA0003067576350000048
The first support data set StAnd corresponding variable scaling step size scaling value alphatCalculating the element parameter theta after the first correction iteration operationp；
According to the meta-parameter thetapReselecting N from the test sample setpAn assayA sample to generate the first support set S modified by the test stagep；
According to the meta-parameter thetatThe meta-parameter θpAnd corresponding optimal scaling value alphapCalculating a loss function
Figure FDA0003067576350000051
In the first support set SpGradient parameter g ofp(ii) a And
according to the meta-parameter thetapAnd the gradient parameter gpDetermining a first adaptive parameter a modified in the test phasep(θp，gp)。
10. The prediction method according to claim 2 or 9, wherein the training phase and the testing phase further comprise the steps of:
acquiring a plurality of key variable data of a chemical process, wherein the chemical process comprises a continuous catalytic naphtha reforming process, and the key variable data comprise input variable data and output variable data;
preprocessing the key variable data according to a 3 sigma criterion to remove abnormal values and outliers; and
and dividing the preprocessed multiple key variable data into the multiple training samples and the multiple testing samples according to a preset proportion.
11. The prediction method of claim 10, wherein the data to be tested is input variable data for the continuous catalytic naphtha reforming process, and the step of determining a quality indicator corresponding to the data to be tested comprises:
in the model independent element learning framework with variable zooming step length, according to the first adaptive parameter a corrected by the testing stagep(θp，gp) And a first support set S modified by the test stagepPredicting output variables corresponding to the input variable dataVolume data.
12. An apparatus for predicting a quality index, comprising:
a memory; and
a processor connected to the memory and configured to implement the method of predicting a quality indicator as claimed in any one of claims 1 to 11.
13. A computer-readable storage medium having stored thereon computer instructions, which, when executed by a processor, implement a method of predicting a quality indicator according to any one of claims 1 to 11.
CN202110530591.3A 2021-05-15 2021-05-15 Quality index prediction method and device Pending CN113139752A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110530591.3A CN113139752A (en) 2021-05-15 2021-05-15 Quality index prediction method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110530591.3A CN113139752A (en) 2021-05-15 2021-05-15 Quality index prediction method and device

Publications (1)

Publication Number Publication Date
CN113139752A true CN113139752A (en) 2021-07-20

Family

ID=76817842

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110530591.3A Pending CN113139752A (en) 2021-05-15 2021-05-15 Quality index prediction method and device

Country Status (1)

Country Link
CN (1) CN113139752A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116644377A (en) * 2023-05-16 2023-08-25 华东理工大学 Soft measurement model construction method based on one-dimensional convolution-stacking self-encoder

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116644377A (en) * 2023-05-16 2023-08-25 华东理工大学 Soft measurement model construction method based on one-dimensional convolution-stacking self-encoder

Similar Documents

Publication Publication Date Title
JP5440394B2 (en) Evaluation prediction apparatus, evaluation prediction method, and program
CN109376267B (en) Method and apparatus for generating a model
US20210081798A1 (en) Neural network method and apparatus
CN112101630A (en) Multi-target optimization method for injection molding process parameters of thin-wall plastic part
CN109063416B (en) Gene expression prediction technique based on LSTM Recognition with Recurrent Neural Network
US20190286791A1 (en) Creation of new chemical compounds having desired properties using accumulated chemical data to construct a new chemical structure for synthesis
CN106709588B (en) Prediction model construction method and device and real-time prediction method and device
JP2012058972A (en) Evaluation prediction device, evaluation prediction method, and program
JP2017049674A (en) Prediction performance curve estimation program, prediction performance curve estimation device and prediction performance curve estimation method
CN110309608B (en) Blast furnace molten iron silicon content forecasting method aiming at time lag uncertain information
CN107480386B (en) Test excitation optimization method based on response aliasing measurement and genetic algorithm
CN109684851A (en) Evaluation of Software Quality, device, equipment and computer storage medium
TW202123098A (en) Method and electronic device for selecting neural network hyperparameters
Guerrero-Pena et al. Multi-objective evolutionary algorithm with prediction in the objective space
CN113139752A (en) Quality index prediction method and device
CN115809624A (en) Automatic analysis design method for microstrip line transmission line of integrated circuit
CN115812210A (en) Method and apparatus for enhancing performance of machine learning classification tasks
CN110991494A (en) Method for constructing prediction model based on improved moth optimization algorithm
CN115856646B (en) Lithium ion battery early life prediction method based on three-dimensional voltage characteristics
CN116612814A (en) Regression model-based batch detection method, device, equipment and medium for gene sample pollution
JP6233432B2 (en) Method and apparatus for selecting mixed model
CN113357138B (en) Method and device for predicting remaining service life of hydraulic pump and terminal equipment
CN109345274B (en) Neighbor user selection method based on BP neural network scoring prediction error
JP2023061477A (en) Program, data processing method, and data processing device
CN116596061B (en) Power amplifier harmonic prediction method based on transfer learning

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