CN110147614A - A kind of engineering safety evaluation method based on the study of diversity of values Stacking multiple-model integration - Google Patents

A kind of engineering safety evaluation method based on the study of diversity of values Stacking multiple-model integration Download PDF

Info

Publication number
CN110147614A
CN110147614A CN201910424901.6A CN201910424901A CN110147614A CN 110147614 A CN110147614 A CN 110147614A CN 201910424901 A CN201910424901 A CN 201910424901A CN 110147614 A CN110147614 A CN 110147614A
Authority
CN
China
Prior art keywords
model
diversity
values
training
sequence
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910424901.6A
Other languages
Chinese (zh)
Other versions
CN110147614B (en
Inventor
肖海斌
毛莺池
齐海
陈豪
程杨堃
杨萍
王龙宝
李然
赵培双
普中勇
杨关发
简云忠
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.)
Hohai University HHU
Huaneng Group Technology Innovation Center Co Ltd
Huaneng Lancang River Hydropower Co Ltd
Original Assignee
Hohai University HHU
Huaneng Group Technology Innovation Center Co Ltd
Huaneng Lancang River Hydropower Co Ltd
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 Hohai University HHU, Huaneng Group Technology Innovation Center Co Ltd, Huaneng Lancang River Hydropower Co Ltd filed Critical Hohai University HHU
Priority to CN201910424901.6A priority Critical patent/CN110147614B/en
Publication of CN110147614A publication Critical patent/CN110147614A/en
Application granted granted Critical
Publication of CN110147614B publication Critical patent/CN110147614B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/13—Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads
    • 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
    • G06Q10/0635—Risk analysis of enterprise or organisation activities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/10—Services
    • G06Q50/26—Government or public services
    • G06Q50/265—Personal security, identity or safety

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Physics & Mathematics (AREA)
  • Human Resources & Organizations (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Geometry (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • Development Economics (AREA)
  • Computer Hardware Design (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • General Business, Economics & Management (AREA)
  • Educational Administration (AREA)
  • Operations Research (AREA)
  • Structural Engineering (AREA)
  • Game Theory and Decision Science (AREA)
  • Health & Medical Sciences (AREA)
  • Quality & Reliability (AREA)
  • Architecture (AREA)
  • Civil Engineering (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Security & Cryptography (AREA)
  • Primary Health Care (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Machine Translation (AREA)

Abstract

The invention discloses a kind of engineering safety evaluation method based on the study of diversity of values Stacking multiple-model integration, steps are as follows: including the structural body feature extraction based on attention mechanism, two stages of model integrated based on diversity of values.Using the thought of divide and conquer, dam is integrally divided into several regional areas (single domain).Result based on structural body region division and each single domain evaluation of running status, using structural body universe Model for Comprehensive, using single domain evaluation result data set as the input of the model, using the autocoder based on attention mechanism, attention weight is dynamically distributed for different single domains；Using a variety of scoring models of Stacking method integration based on diversity of values, realize that carrying out accurate, stable safety comprehensive to the structural body overall situation judges.

Description

A kind of engineering safety evaluation based on the study of diversity of values Stacking multiple-model integration Method
Technical field
The invention belongs to project security monitoring fields, in particular to a kind of to be based on diversity of values Stacking multiple-model integration The engineering safety evaluation method of study.
Background technique
Spatial dimension shared by structural body is big, deployment measuring point quantity is more, monitoring data amount is huge.In view of its stress condition has There is locality, there is similitude between minor structure, different measuring points spatial position and time series variation rule have correlation.It adopts With the thought of divide and conquer, dam whole (universe) is divided into several regional areas (single domain), structural body overall situation operating status body In now each local single domain, single domain evaluation result is merged, can efficiently utilize the complementarity of multiple single domain features in structural body, To obtain to structural body more comprehensively operational monitoring information.
The stress condition of different single domains has differences in structural body, and the single domain stress as the dam foundation and dam are called in person changes complexity, right Entire dam structure security implication is great.On the contrary, the single domain stress of dam crest and dam abutment is relatively easy, to dam structure security implication It is small.Therefore, different according to single domain three-dimensional effect, the influence degree that evaluation result judges structural body universe is also different.It is another Aspect, during structural body production run, physical features change in single domain, and stress condition has timeliness.Single domain fortune Row state changes over time, and same single domain is also different in influence degree of the different time to structural body universe.Fusion When single domain evaluation result, introduces attention mechanism and extract structural body operation characteristic, different single domain evaluation distribution weights obtain more Reasonable scoring model input.
Existing structural body scoring model includes the support vector machines dam deformation scoring model based on wavelet transformation (Wavelet Support Vector Machine, WSVM), small wave fractal diagnostic model (Wavelet-Fractal Diagnosis Model, WFDM), expert's enabling legislation (Experts Weighting Method, EWM), these models are in difference By parameter adjustment and calibration on data set, learning performance performance has his own strong points.Stacking multi-model based on diversity of values The engineering safety evaluation method of integrated study integrates existing structural body scoring model, is scored with study and measures primary scoring model Learning performance constructs the input of second-level model according to diversity of values, promotes strong primary mold ability to express, Optimized model integration Energy.
Summary of the invention
Goal of the invention: in order to overcome the problems of the prior art, the present invention provides a kind of based on diversity of values Stacking The engineering safety evaluation method of multiple-model integration study, on the basis of structural body region division and each single domain evaluation of running status On, using the autocoder (RED) based on attention mechanism, attention weight is dynamically distributed for different single domains, using being based on The Stacking method (SD-Stacking) of diversity of values integrates a variety of scoring models, realizes and carries out safety to the structural body overall situation Comprehensive Evaluation.
Technical solution: to achieve the above object, the present invention provides a kind of based on diversity of values Stacking multiple-model integration The engineering safety evaluation method of study, includes the following steps:
(1) based on the structural body feature extraction of attention mechanism；
(1.1) attention mechanism is introduced；
(1.2) coding-decoding structure is constructed；
(1.3) RED cataloged procedure；
(1.4) RED decoding process；
(1.5) encoding target function is obtained；
(3) based on the model integrated of diversity of values: integrated expert's enabling legislation EWM, the support vector machines based on wavelet transformation Dam deformation scoring model WSVM, neural network NN choose decision tree DT as second-level model as primary mold；Utilize study Primary mold learning performance is measured in scoring, is exported distribution weight by primary mold of diversity of values, is improved second-level model input.
Further, specific step is as follows for introducing attention mechanism in the step (1.1): using based on RNN's Encoder-Decoder structure (RNN-based Encoder-DEcoder, RED), it introduces attention mechanism and extracts structural body spy Sign distinguishes significance level of the different single domains in the judge of the structural body overall situation.
Further, specific step is as follows for decoding structure for building coding-in the step (1.2): structural body is in the timeSingle domain evaluation result collection isIt is made of, is rewritten as by subscript sequence each single domain evaluation result Source sequence x=(d1,d2,…,dm), it carries out encoding and decoding and obtains target sequence y=(y1,y2,…,yn)；Source sequence and target sequence Length can be unequal, i.e. m ≠ n.
Further, specific step is as follows for RED cataloged procedure in the step (1.3): encoder Encoder circulation is read Each component in list entries x is taken, updates hidden state in step-length t:Wherein f is nonlinear activation letter Number；Cataloged procedure is attention weight distribution p (y of the learning objective sequence under source sequence1,y2,…,yn|d1,d2,…,dm)； Attention weight p (d on source sequence out can be learnt by next component in training prediction source sequence using RNNt| dt-1,…,d1)；It combines the attention weight under each step-length and obtains the attention distribution of source sequence:When the end of encoder cycles to list entries, the hidden state of RNN will be updated to The corresponding abstract characteristics of entire list entries x express Feature Expression, and FE, FE are by the input as decoder.
Further, specific step is as follows for RED decoding process in the step (1.4): decoder Decoder is according to defeated Enter the feature representation FE of sequence, predicts given hidden stateUnder output sequence component yt, finally training obtains output sequence y； Hidden state at step-length t are as follows:It is similar with cataloged procedure, calculate output sequence component ytAttention Power weight are as follows:Wherein the f in g and cataloged procedure is nonlinear activation letter Number.
Further, specific step is as follows: joint training encoder for acquisition encoding target function in the step (1.5) And decoder, the joint attention weight distribution of output sequence under list entries is given with maximization, then objective function are as follows:Wherein training set length is λ, p (yi|xi) it is target sequence yiIn source sequence xiUnder note Meaning power weight distribution.
Further, specific step is as follows for the model integrated based on diversity of values in the step (2):
(2.1) primary mold scores: for K selected primary mold { model1,model2,…,modelK, it is fixed Adopted its learning performance scoring is { score1,score2,…,scoreK}；If score function are as follows: rate (modelκ)=scoreκ,κ =1,2 ..., K；Comprehensive Evaluation belongs to classification task, then using accuracy rate ACC, F1 as score function；ACC descriptive model is whole Body classification accuracy, i.e., all ratios by correct classification samples；F1 is the probability that accurate rate i.e. some classification is correctly validated With recall rate, that is, correct classification identified ratio harmonic-mean, the stability of model can be measured；ACC and F1 value model Enclosing is [0,1], and it is better that value closer to 1 represents respective performances；
(2.2) based on the weight distribution of scoring: sample data set is divided into Training and Testing two parts；It will Training set Training is averagely allocated to K primary mold and obtains { training1,training2,…,trainingK}；Every A primary mold is trained and tests, and remembers modelκIn trainingκOutput with Testing is respectively AκAnd Bκ；Using first Training set [the A of the output construction second-level model of grade model1；A2；…；AK] and test set mean (B1,B2,…,BK)；Based on primary Model score, the weight of K primary mold of note are { w1,w2,…,wK, definition scoring weighting function are as follows:Weight is the embodiment of different models learning performance on sample data set, thus is Adaptive；The training set of second-level model and test set are rewritten as [w using weight1A1；w2A2；…；wKAK] and mean (w1B1, w2B2,…,wKBK)；
(2.3) second-level model training and Comprehensive Evaluation: second-level model, that is, meta-model is in [w1A1；w2A2；…；wKAK] and mean (w1B1,w2B2,…,wKBK) on be trained and test；The learning performance of different primary molds is distinguished using diversity of values, originally When grade model learning performance is close, respective weights difference is small, limited to second-level model performance boost；And when primary model learning Obvious classification can occur, cause weight difference to become larger, strong model can be obviously improved using diversity of values at this time and inhibit weak mould Type expression, improves second-level model precision.
The utility model has the advantages that compared with the prior art, the present invention has the following advantages:
The present invention integrates the existing support vector machines dam deformation scoring model (Wavelet included based on wavelet transformation Support Vector Machine, WSVM), small wave fractal diagnostic model (Wavelet-Fractal Diagnosis Model, WFDM), the structural bodies scoring model such as expert's enabling legislation (Experts Weighting Method, EWM), commented with study Divide and measure primary scoring model learning performance, the input of second-level model is constructed according to diversity of values, promotes strong primary mold expression Ability, Optimized model integrate performance.
Detailed description of the invention
Fig. 1 is the universe Comprehensive Evaluation mistake of the Stacking multiple-model integration study in specific embodiment based on diversity of values Cheng Tu；
Fig. 2 is structural body universe Model for Comprehensive structure chart in specific embodiment；
Fig. 3 is the Encoder-Decoder structure chart based on RNN in specific embodiment；
Fig. 4 is SD-Stacking method Comprehensive Evaluation procedure chart in specific embodiment.
Specific embodiment
Combined with specific embodiments below, the present invention is furture elucidated, it should be understood that these embodiments are merely to illustrate the present invention Rather than limit the scope of the invention, after the present invention has been read, those skilled in the art are to various equivalences of the invention The modification of form falls within the application range as defined in the appended claims.
On the basis of structural body region division and each single domain evaluation of running status, using based on the automatic of attention mechanism Encoder (RED) dynamically distributes attention weight for different single domains, using the Stacking method (SD- based on diversity of values Stacking a variety of scoring models) are integrated, realizes and safety comprehensive judge is carried out to the structural body overall situation.
Based on measuring point space-time characteristic partition structure body, several single domain R={ r are obtainedj| j=1 ..., m }, single domain rjPhysics shape State evaluation result is dj.Remember structural body in the timeSingle domain evaluation result collection isUsual access timeLength before is the single domain evaluation history result set of λ as data sample, i.e.,It utilizes It is as shown in Figure 1 that Stacking method integration model carries out Comprehensive Evaluation process.
Comprehensive Evaluation inputsInput feature vector is extracted using autocoder.Select κ a existing Scoring model model1,model2,…,modelκAs primary mold (level1).It is commented according to primary mold learning performance Divide (Score), the input of analysis score difference (Score Difference) construction second-level model (meta-model, MetaModel). Pass through Stacking method training primary mold and secondary model, inputObtain the timeUniverse Comprehensive Evaluation result Component therein can correspond to different judge states.Such as setting structural body evaluation result includes four states, good, normal, inspection And exception, corresponding Comprehensive Evaluation resultEach component.At this timeIndicate structural body pair Answering four shape probability of states is respectively 26.4%, 42.5%, 22.5% and 8.6%, should be belonged to according to maximum probability decision structure body In normal condition.
Using a kind of structural body universe Model for Comprehensive (Global judgement Model, GUM), model structure is such as Shown in Fig. 2, input by single domain evaluation result data set (Evaluationsof Regions) as the model utilizes one kind Based on RNN Encoder-Decoder structure (RNN-based Encoder-DEcoder, RED), it introduces attention mechanism and mentions Structure body characteristics are taken, significance level of the different single domains in the judge of the structural body overall situation is distinguished.Select several existing judge moulds Type is not all primary scoring model scoring as primary mold, according to learning performance, and it is defeated to construct second-level model according to diversity of values Enter.By Stacking method based on diversity of values (Score-Difference based Stacking Method,SD- Stacking primary scoring model) is integrated, realizes that structural body safety comprehensive is judged.It is specifically divided into two processes: based on attention The structural body feature extraction of mechanism, the model integrated based on diversity of values.Specific implementation step is described as follows:
(1) based on the structural body feature extraction of attention mechanism:
(1.1) it is introduced into attention mechanism: in project security monitoring practice, each single domain space that structural body divides The differences such as structure, stress condition, there is also differences for physical features, meanwhile, structural body stress condition can as time goes by by Gradually change.Section in different times, the influence that single domain evaluates universe are dynamically changeables, are distributed using attention mechanism corresponding Weight makes feature representation more reasonable, improves Comprehensive Evaluation accuracy rate.
(1.2) coding-decoding structure: the present invention proposes a kind of Encoder-Decoder (RED) based on RNN, and structure is such as Shown in Fig. 3, structural body is in the timeSingle domain evaluation result collection isSubscript is pressed by each single domain evaluation result Sequence forms, and is rewritten as source sequence x=(d1,d2,…,dm), it carries out encoding and decoding and obtains target sequence y=(y1,y2,…, yn).Translation between similar different language, source sequence and target sequence length can be unequal, i.e. m ≠ n.
(1.3) RED cataloged procedure: encoder (Encoder) circulation reads each component in list entries x, in step-length t Update hidden state:Wherein f is nonlinear activation function.Cataloged procedure is learning objective sequence in source sequence Under attention weight distribution p (y1,y2,…,yn|d1,d2,…,dm).It can be by training prediction source sequence using RNN Next component learns attention weight p (d on source sequence outt|dt-1,…,d1).Combine the attention weight under each step-length Obtain the attention distribution of source sequence:When the end of encoder cycles to list entries, The hidden state of RNN will be updated to the expression of entire list entries x corresponding abstract characteristics (Feature Expression, FE), FE is by the input as decoder.
(1.4) RED decoding process: feature representation FE of the decoder (Decoder) according to list entries predicts given hide StateUnder output sequence component yt, finally training obtains output sequence y.Hidden state at step-length t are as follows:It is similar with cataloged procedure, calculate output sequence component ytAttention weight are as follows:Wherein the f in g and cataloged procedure is nonlinear activation function.
(1.5) encoding target function: joint training encoder and decoder export sequence under given list entries to maximize The joint attention weight distribution of column, then objective function are as follows:Wherein training set length is λ It (corresponds to above), p (yi|xi) it is target sequence yiIn source sequence xiUnder attention weight distribution.It compiles Decoding process can be micro-, the method based on gradient can be used to minimize objective function, to estimate parameter.
The input and output size of RED is variable.When single domain quantity m is very big, it can be carried out at dimensionality reduction by encoding and decoding Reason is so that m > > n.If m is of moderate size, settable output sequence size m ≈ n.Encoding-decoding process is source sequence x=(d1, d2,…,dm) distribution weight, it is translated into more reasonable sequences y=(y1,y2,…,yn), to help to improve global comprehensive It closes and judges training effectiveness (dimensionality reduction) and accuracy rate (attention).
(2) based on the model integrated of diversity of values: it includes that expert assigns power that Model for Comprehensive is commonly used in the practical O&M of dam Method (Expert Weighting Method, EWM), the support vector machines dam deformation scoring model based on wavelet transformation Professional Models such as (Wavelet Support Vector Machine, WSVM) and neural network (Neural Network, NN).As shown in figure 4, the present invention propose it is a kind of based on diversity of values Stacking method (Score-Difference basedStackingMethod, SD-Stacking), EWM, WSVM, NN are integrated as primary mold, are chosen decision tree (DT) and are used as two Grade model, is scored using study and measures primary mold learning performance, is exported distribution weight by primary mold of diversity of values, is improved Second-level model input, improves model integrated performance, realizes that dam universe carries out Comprehensive Evaluation.
(2.1) primary mold scores: for K selected primary mold { model1,model2,…,modelK, it is fixed Adopted its learning performance scoring is { score1,score2,…,scoreK}.If score function are as follows: rate (modelκ)=scoreκ,κ =1,2 ..., K.Comprehensive Evaluation belongs to classification task, then using accuracy rate ACC, F1 as score function.ACC descriptive model is whole Body classification accuracy, i.e., all ratios by correct classification samples.F1 is accurate rate (probability that some classification is correctly validated) With the harmonic-mean of recall rate (the correct identified ratio of classification), the stability of model can be measured.ACC and F1 value model Enclosing is [0,1], and it is better that value closer to 1 represents respective performances.
(2.2) based on the weight distribution of scoring: sample data set is divided into Training and Testing two parts.It will Training set Training is averagely allocated to K primary mold and obtains { training1,training2,…,trainingK}.Every A primary mold is trained and tests, and remembers modelκIn trainingκOutput with Testing is respectively AκAnd Bκ.Using first Training set [the A of the output construction second-level model of grade model1；A2；…；AK] and test set mean (B1,B2,…,BK).Based on primary Model score, the weight of K primary mold of note are { w1,w2,…,wK, definition scoring weighting function are as follows:Weight is the embodiment of different models learning performance on sample data set, thus is Adaptive.The training set of second-level model and test set are rewritten as [w using weight1A1；w2A2；…；wKAK] and mean (w1B1, w2B2,…,wKBK)。
(2.3) second-level model training and Comprehensive Evaluation: second-level model (meta-model) is in [w1A1；w2A2；…；wKAK] and mean (w1B1,w2B2,…,wKBK) on be trained and test.The learning performance of different primary molds, second level are distinguished using diversity of values Model can preferably utilize the feature of primary mold itself.When primary model learning performance is close, respective weights difference is small, It is limited to second-level model performance boost.And when the primary obvious classification of model learning performance appearance, cause weight difference to become larger, at this time Strong model can be obviously improved using diversity of values and inhibits weak model tormulation, improve second-level model precision.

Claims (7)

1. a kind of engineering safety evaluation method based on the study of diversity of values Stacking multiple-model integration, which is characterized in that packet Include following steps:
(1) based on the structural body feature extraction of attention mechanism；
(1.1) attention mechanism is introduced；
(1.2) coding-decoding structure is constructed；
(1.3) RED cataloged procedure；
(1.4) RED decoding process；
(1.5) encoding target function is obtained；
(2) based on the model integrated of diversity of values: integrated expert's enabling legislation EWM, the support vector machines dam based on wavelet transformation Scoring model WSVM, neural network NN are deformed as primary mold, chooses decision tree DT as second-level model；It is scored using study Primary mold learning performance is measured, distribution weight is exported by primary mold of diversity of values, improves second-level model input.
2. a kind of engineering safety evaluation based on the study of diversity of values Stacking multiple-model integration according to claim 1 Method, which is characterized in that attention mechanism is introduced in the step (1.1), and specific step is as follows: using based on RNN's Encoder-Decoder structure introduces attention mechanism and extracts structure body characteristics, distinguishes different single domains and judges in the structural body overall situation In significance level.
3. a kind of engineering safety evaluation based on the study of diversity of values Stacking multiple-model integration according to claim 1 Method, which is characterized in that specific step is as follows for decoding structure for building coding-in the step (1.2): structural body is in the time Single domain evaluation result collection isIt is made of each single domain evaluation result by subscript sequence, is rewritten as source Sequence x=(d1,d2,…,dm), it carries out encoding and decoding and obtains target sequence y=(y1,y2,…,yn)；Source sequence and target sequence are long Degree can be unequal, i.e. m ≠ n.
4. a kind of engineering safety evaluation based on the study of diversity of values Stacking multiple-model integration according to claim 1 Method, which is characterized in that specific step is as follows for RED cataloged procedure in the step (1.3): encoder Encoder circulation is read Each component in list entries x is taken, updates hidden state in step-length t:Wherein f is nonlinear activation letter Number；Cataloged procedure is attention weight distribution p (y of the learning objective sequence under source sequence1,y2,…,yn|d1,d2,…,dm)； Attention weight p (d on source sequence out can be learnt by next component in training prediction source sequence using RNNt| dt-1,…,d1)；It combines the attention weight under each step-length and obtains the attention distribution of source sequence:When the end of encoder cycles to list entries, the hidden state of RNN will be updated to The corresponding abstract characteristics of entire list entries x express FE, and FE is by the input as decoder.
5. a kind of engineering safety evaluation based on the study of diversity of values Stacking multiple-model integration according to claim 1 Method, which is characterized in that specific step is as follows for RED decoding process in the step (1.4): decoder Decoder is according to defeated Enter the feature representation FE of sequence, predicts given hidden stateUnder output sequence component yt, finally training obtains output sequence y； Hidden state at step-length t are as follows:It is similar with cataloged procedure, calculate output sequence component ytAttention Power weight are as follows:Wherein the f in g and cataloged procedure is nonlinear activation letter Number.
6. a kind of engineering safety evaluation based on the study of diversity of values Stacking multiple-model integration according to claim 1 Method, which is characterized in that encoding target function is obtained in the step (1.5), and specific step is as follows: joint training encoder And decoder, the joint attention weight distribution of output sequence under list entries is given with maximization, then objective function are as follows:Wherein training set length is λ, p (yi|xi) it is target sequence yiIn source sequence xiUnder note Meaning power weight distribution.
7. a kind of engineering safety evaluation based on the study of diversity of values Stacking multiple-model integration according to claim 1 Method, which is characterized in that specific step is as follows for the model integrated based on diversity of values in the step (2):
(2.1) primary mold scores: for K selected primary mold { model1,model2,…,modelK, define it Learning performance scoring is { score1,score2,…,scoreK}；If score function are as follows: rate (modelκ)=scoreκ, κ=1, 2,…,K；Comprehensive Evaluation belongs to classification task, then using accuracy rate ACC, F1 as score function；ACC descriptive model integrally divides Class accuracy rate, i.e., all ratios by correct classification samples；F1 is the probability and call together that accurate rate i.e. some classification is correctly validated Rate, that is, correct classification identified ratio harmonic-mean is returned, the stability of model can be measured；ACC and F1 value range is equal For [0,1], it is better that value closer to 1 represents respective performances；
(2.2) based on the weight distribution of scoring: sample data set is divided into Training and Testing two parts；It will train Collection Training is averagely allocated to K primary mold and obtains { training1,training2,…,trainingK}；Each first Grade model is trained and tests, and remembers modelκIn trainingκOutput with Testing is respectively AκAnd Bκ；Utilize primary mould Training set [the A of the output construction second-level model of type1；A2；…；AK] and test set mean (B1,B2,…,BK)；Based on primary mold Scoring, the weight of K primary mold of note are { w1,w2,…,wK, definition scoring weighting function are as follows:Weight is the embodiment of different models learning performance on sample data set, thus is Adaptive；The training set of second-level model and test set are rewritten as [w using weight1A1；w2A2；…；wKAK] and mean (w1B1, w2B2,…,wKBK)；
(2.3) second-level model training and Comprehensive Evaluation: second-level model, that is, meta-model is in [w1A1；w2A2；…；wKAK] and mean (w1B1,w2B2,…,wKBK) on be trained and test；The learning performance of different primary molds is distinguished using diversity of values, originally When grade model learning performance is close, respective weights difference is small, limited to second-level model performance boost；And when primary model learning Obvious classification can occur, cause weight difference to become larger, strong model can be obviously improved using diversity of values at this time and inhibit weak mould Type expression, improves second-level model precision.
CN201910424901.6A 2019-05-21 2019-05-21 Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning Active CN110147614B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910424901.6A CN110147614B (en) 2019-05-21 2019-05-21 Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910424901.6A CN110147614B (en) 2019-05-21 2019-05-21 Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning

Publications (2)

Publication Number Publication Date
CN110147614A true CN110147614A (en) 2019-08-20
CN110147614B CN110147614B (en) 2022-02-08

Family

ID=67592394

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910424901.6A Active CN110147614B (en) 2019-05-21 2019-05-21 Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning

Country Status (1)

Country Link
CN (1) CN110147614B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116502145A (en) * 2023-03-15 2023-07-28 西南交通大学 Method, device and medium for grading stack learning surrounding rock based on improved ER rule

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040133531A1 (en) * 2003-01-06 2004-07-08 Dingding Chen Neural network training data selection using memory reduced cluster analysis for field model development
CN104200263A (en) * 2014-07-23 2014-12-10 浙江工业大学 Power distribution network route planning method based on tabu differential evolution and GIS (Geographic Information System)
CN106532720A (en) * 2016-12-20 2017-03-22 国网辽宁省电力有限公司沈阳供电公司 Dynamic partition fault recovery method of power distribution network containing distributed power supply
CN108122068A (en) * 2017-12-05 2018-06-05 中国电力科学研究院有限公司 A kind of power distribution network risk-averse retailer method and system

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20040133531A1 (en) * 2003-01-06 2004-07-08 Dingding Chen Neural network training data selection using memory reduced cluster analysis for field model development
CN104200263A (en) * 2014-07-23 2014-12-10 浙江工业大学 Power distribution network route planning method based on tabu differential evolution and GIS (Geographic Information System)
CN106532720A (en) * 2016-12-20 2017-03-22 国网辽宁省电力有限公司沈阳供电公司 Dynamic partition fault recovery method of power distribution network containing distributed power supply
CN108122068A (en) * 2017-12-05 2018-06-05 中国电力科学研究院有限公司 A kind of power distribution network risk-averse retailer method and system

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116502145A (en) * 2023-03-15 2023-07-28 西南交通大学 Method, device and medium for grading stack learning surrounding rock based on improved ER rule
CN116502145B (en) * 2023-03-15 2024-01-30 西南交通大学 Method, device and medium for grading stack learning surrounding rock based on improved ER rule

Also Published As

Publication number Publication date
CN110147614B (en) 2022-02-08

Similar Documents

Publication Publication Date Title
CN112149316B (en) Aero-engine residual life prediction method based on improved CNN model
CN108985335B (en) Integrated learning prediction method for irradiation swelling of nuclear reactor cladding material
CN106971240A (en) The short-term load forecasting method that a kind of variables choice is returned with Gaussian process
CN109816095A (en) Based on the network flow prediction method for improving gating cycle neural network
CN102411687B (en) Deep learning detection method of unknown malicious codes
CN105160249B (en) A kind of method for detecting virus based on improved Artificial neural network ensemble
CN110210495B (en) XGboost soft measurement modeling method based on parallel LSTM self-encoder dynamic feature extraction
CN106980877A (en) A kind of Prediction of Blasting Vibration method based on particle cluster algorithm Support Vector Machines Optimized
CN108921279A (en) Reservoir day enters water prediction technique
CN106778846A (en) A kind of method for forecasting based on SVMs
Arciszewski et al. A methodology of design knowledge acquisition for use in learning expert systems
CN106778838A (en) A kind of method for predicting air quality
CN105808689A (en) Drainage system entity semantic similarity measurement method based on artificial neural network
CN112614021B (en) Tunnel surrounding rock geological information prediction method based on built tunnel information intelligent identification
CN108062566A (en) A kind of intelligent integrated flexible measurement method based on the potential feature extraction of multinuclear
CN115130375A (en) Rock burst intensity prediction method
CN112686372A (en) Product performance prediction method based on depth residual GRU neural network
CN113239211A (en) Reinforced learning knowledge graph reasoning method based on course learning
CN110147614A (en) A kind of engineering safety evaluation method based on the study of diversity of values Stacking multiple-model integration
CN111123232A (en) Radar individual identification system with task adaptability
Richards et al. Optimisation of competition indices using simulated annealing and artificial neural networks
CN115565669B (en) Cancer survival analysis method based on GAN and multitask learning
CN117034767A (en) Ceramic roller kiln temperature prediction method based on KPCA-GWO-GRU
CN110288080A (en) A kind of neural network model and learning method based on SELU and FW mechanism
CN114239397A (en) Soft measurement modeling method based on dynamic feature extraction and local weighted deep 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
GR01 Patent grant
GR01 Patent grant