CN110147614B - Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning - Google Patents

Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning Download PDF

Info

Publication number
CN110147614B
CN110147614B CN201910424901.6A CN201910424901A CN110147614B CN 110147614 B CN110147614 B CN 110147614B CN 201910424901 A CN201910424901 A CN 201910424901A CN 110147614 B CN110147614 B CN 110147614B
Authority
CN
China
Prior art keywords
model
training
sequence
score
primary
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910424901.6A
Other languages
Chinese (zh)
Other versions
CN110147614A (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

Images

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—Information and communication technology [ICT] specially adapted for implementation of business processes of 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)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Tourism & Hospitality (AREA)
  • Strategic Management (AREA)
  • Geometry (AREA)
  • Economics (AREA)
  • Educational Administration (AREA)
  • Development Economics (AREA)
  • General Business, Economics & Management (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Computer Hardware Design (AREA)
  • Game Theory and Decision Science (AREA)
  • Quality & Reliability (AREA)
  • Mathematical Analysis (AREA)
  • Civil Engineering (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Operations Research (AREA)
  • Architecture (AREA)
  • Structural Engineering (AREA)
  • Evolutionary Computation (AREA)
  • Computer Security & Cryptography (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Machine Translation (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses an engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning, which comprises the following steps: the method comprises structural body feature extraction based on an attention mechanism, and a model based on score difference integrates two stages. The dam is divided into a plurality of local areas (single areas) by adopting the idea of a divide-and-conquer method. Based on the results of structural body region division and each single-domain running state evaluation, a structural body global comprehensive evaluation model is adopted, a single-domain evaluation result data set is used as the input of the model, and attention weights are dynamically distributed to different single domains by using an automatic encoder based on an attention mechanism; and integrating various evaluation models by using a grading difference-based Stacking method to realize accurate and stable comprehensive safety evaluation on the whole structure.

Description

Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning
Technical Field
The invention belongs to the field of engineering safety monitoring, and particularly relates to an engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning.
Background
The structure occupies a large space range, the number of deployed measuring points is large, and the monitoring data volume is huge. Considering that the stress condition of the strain has locality, the substructures have similarity, and the spatial positions and the time sequence change rules of different measuring points have correlation. The method adopts the concept of a divide-and-conquer method to divide the whole dam (universe) into a plurality of local areas (single areas), the global operation state of the structure is embodied in each local single area, the single area evaluation result is fused, the complementarity of a plurality of single area characteristics in the structure can be effectively utilized, and therefore more comprehensive operation monitoring information of the structure is obtained.
The stress conditions of different single domains in the structure body are different, for example, the single domain stress change of the dam foundation and the dam heel is complex, and the safety of the whole dam structure is greatly influenced. On the contrary, the single-domain stress of the dam crest and the dam abutment is relatively simple, and the influence on the safety of the dam structure is small. Therefore, the evaluation result has different influence on the structure global evaluation according to the single-domain spatial effect. On the other hand, physical characteristics in a single domain are changed during the production and operation of the structure, and the stress condition has timeliness. The operating state of a single domain changes with time, and the influence degree of the same single domain on the whole structure at different times is different. When single-domain evaluation results are fused, an attention mechanism is introduced to extract the operation characteristics of the structural body, weights are distributed in different single-domain evaluations, and more reasonable evaluation model input is obtained.
The existing structural body evaluation Model comprises a Wavelet Support Vector Machine dam deformation evaluation Model (WSVM), a Wavelet Fractal Diagnosis Model (WFDM) and an Expert Weighting Method (EWM) based on Wavelet transformation, wherein the models have long learning performance expression after parameter adjustment and calibration on different data sets. The engineering safety evaluation method of Stacking multi-model ensemble learning based on the grading difference integrates the existing structural body evaluation model, the learning performance of the primary evaluation model is measured by learning grading, the input of the secondary model is constructed according to the grading difference, the strong primary model expression capability is improved, and the model integration performance is optimized.
Disclosure of Invention
The purpose of the invention is as follows: in order to solve the problems in the prior art, the invention provides a grading difference based Stacking multi-model ensemble learning engineering safety evaluation method, which is characterized in that on the basis of structural body region division and evaluation of each single-domain running state, an automatic encoder (RED) based on an attention mechanism is utilized to dynamically distribute attention weights for different single domains, and a grading difference based Stacking method (SD-Stacking) is utilized to integrate multiple evaluation models, so that the overall safety comprehensive evaluation of a structural body is realized.
The technical scheme is as follows: in order to achieve the purpose, the invention provides an engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning, which comprises the following steps:
(1) extracting structural body features based on an attention mechanism;
(1.1) introducing an attention mechanism;
(1.2) constructing an encoding-decoding structure;
(1.3) RED encoding process;
(1.4) RED decoding process;
(1.5) acquiring a coding objective function;
(3) model integration based on score difference: integrating an expert weighting method EWM, a support vector machine dam deformation evaluation model WSVM based on wavelet transformation and a neural network NN as primary models, and selecting a decision tree DT as a secondary model; and measuring the learning performance of the primary model by using the learning score, and distributing weight for the primary model output by using the score difference to improve the input of the secondary model.
Further, the specific steps of introducing the attention mechanism in the step (1.1) are as follows: using an RNN-based Encoder-Decoder structure (RNN-based Encoder-Decoder, RED), an attention mechanism is introduced to extract structural features, and importance degrees of different single domains in global judgment of the structure are distinguished.
Further, the specific steps of constructing the encoding-decoding structure in the step (1.2) are as follows: time of the structure
Figure BDA0002067172610000021
The single domain evaluation result set is
Figure BDA0002067172610000022
Each single-domain evaluation result is composed in the order of superscript, and is rewritten into the source sequence x ═ d1,d2,…,dm) Encoding and decoding to obtain the target sequence y ═ y (y)1,y2,…,yn) (ii) a The source sequence and the target sequence may not be equal in length, i.e., m ≠ n.
Further, the RED encoding process in step (1.3) includes the following specific steps: the Encoder reads each component in the input sequence x cyclically, updating the hidden state at step t:
Figure BDA0002067172610000023
wherein f is a nonlinear activation function; the coding process is to learn the attention weight distribution p (y) of the target sequence under the source sequence1,y2,…,yn|d1,d2,…,dm) (ii) a MiningUsing RNN, it is possible to learn the attention weight p (d) on the source sequence by training to predict the next component in the source sequencet|dt-1,…,d1) (ii) a Combining the attention weights at each step yields the attention distribution of the source sequence:
Figure BDA0002067172610000024
when the encoder loops to the end of the input sequence, the hidden state of the RNN will be updated to the abstract Feature Expression FE, FE corresponding to the entire input sequence x will be the input to the decoder.
Further, the RED decoding process in step (1.4) includes the following specific steps: decoder predicts given hidden state according to characteristic expression FE of input sequence
Figure BDA0002067172610000031
Output sequence component y oftFinally training to obtain an output sequence y; the hidden states at step t are:
Figure BDA0002067172610000032
similar to the encoding process, the output sequence component y is calculatedtThe attention weight of (1) is:
Figure BDA0002067172610000033
wherein g and f in the encoding process are both nonlinear activation functions.
Further, the specific steps of obtaining the encoding objective function in the step (1.5) are as follows: jointly training the encoder and decoder to maximize the joint attention weight distribution of the output sequence given the input sequence, the objective function is then:
Figure BDA0002067172610000034
wherein the training set length is λ, p (y)i|xi) Is the target sequence yiIn the source sequence xiAttention weight distribution of the following.
Further, the specific steps of the model integration based on the score difference in the step (2) are as follows:
(2.1) primary model scoring: for the selected K elementary models { model { number1,model2,…,modelKDefine its learning performance score as { score1,score2,…,scoreK}; let the scoring function be: rate (model)κ)＝scoreκK ═ 1,2, …, K; if the comprehensive evaluation belongs to the classification task, the accuracy rates ACC and F1 are used as scoring functions; the ACC describes the overall classification accuracy of the model, namely the proportion of all correctly classified samples; f1 is the harmonic mean value of the precision rate, namely the probability that a certain category is correctly identified, and the recall rate, namely the proportion that the correct category is identified, and can measure the stability of the model; ACC and F1 both range from [0,1 ]]The closer the value is to 1, the better the corresponding performance is;
(2.2) score-based weight assignment: dividing the sample data set into two parts of Training and Testing; averagely distributing Training set Training to K primary models to obtain { Training1,training2,…,trainingK}; training and testing were performed on each of the primary models, memory modelκIn trailingκAnd the outputs of Testing are A respectivelyκAnd Bκ(ii) a Constructing a training set of secondary models using the outputs of the primary models [ A ]1；A2；…；AK]And test set mean (B)1,B2,…,BK) (ii) a Based on the primary model scores, the weights of K primary models are recorded as { w }1,w2,…,wKDefine a scoring weight function as:
Figure BDA0002067172610000035
the weight is the embodiment of the learning performance of different models on the sample data set, so the method is self-adaptive; rewriting the training set and test set of the secondary model to [ w ] using weights1A1；w2A2；…；wKAK]And mean (w)1B1,w2B2,…,wKBK)；
(2.3) training and comprehensive judgment of a secondary model: the second-level model, i.e., the meta-model, is in [ w ]1A1；w2A2；…；wKAK]And mean (w)1B1,w2B2,…,wKBK) Training and testing; the learning performances of different primary models are distinguished by using the grading difference, and when the learning performances of the primary models are close to each other, the corresponding weight difference is small, so that the improvement on the performance of the secondary model is limited; and when the learning performance of the primary model is obviously graded, the weight difference is increased, and at the moment, the strong model can be obviously promoted by utilizing the grading difference to inhibit the expression of the weak model, so that the precision of the secondary model is improved.
Has the advantages that: compared with the prior art, the invention has the following advantages:
the Method integrates the existing structural body evaluation models such as a Wavelet Support Vector Machine (WSVM) deformation evaluation Model, a Wavelet Fractal Diagnosis Model (WFDM), an Expert Weighting Method (EWM) and the like based on Wavelet transformation, so as to measure the learning performance of the primary evaluation Model by learning scores, construct the input of a secondary Model according to the score difference, improve the expression capability of the strong primary Model and optimize the integration performance of the Model.
Drawings
FIG. 1 is a diagram illustrating a global comprehensive evaluation process of Stacking multi-model ensemble learning based on score difference in an embodiment;
FIG. 2 is a structural universe comprehensive evaluation model structure diagram in an embodiment;
FIG. 3 is a diagram of an RNN-based Encode-Decoder architecture in an exemplary embodiment;
FIG. 4 is a diagram of the comprehensive evaluation process of the SD-Stacking method in the embodiment.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
On the basis of structural body region division and each single-domain running state evaluation, attention weights are dynamically allocated to different single domains by using an attention mechanism-based automatic encoder (RED), and a grading difference-based grading method (SD-grading) is adopted to integrate multiple evaluation models, so that the overall safety comprehensive evaluation of the structural body is realized.
Dividing the structure based on the space-time characteristics of the measuring points to obtain a plurality of single domains R ═ { R ═ Rj1, …, m, single field rjThe physical State evaluation result was dj. Time of structure body
Figure BDA0002067172610000041
The single domain evaluation result set is
Figure BDA0002067172610000042
Time is usually selected
Figure BDA0002067172610000043
Previous sets of single domain evaluation history results of length λ as data samples, i.e.
Figure BDA0002067172610000044
The comprehensive evaluation process by using the Stacking method to integrate the model is shown in figure 1.
The comprehensive judgment input is
Figure BDA0002067172610000045
Input features are extracted using an auto-encoder. Select kappa existing assessment model models1,model2,…,modelκAs a primary model (level 1). The primary model was scored for learning performance (Score) and the Score Difference (Score Difference) was analyzed to construct the input to a secondary model (MetaModel). Training the primary model and the secondary model by a Stacking method, inputting
Figure BDA0002067172610000046
Time of arrival
Figure BDA0002067172610000047
Global comprehensive evaluation result
Figure BDA0002067172610000051
The components may correspond to different evaluation states. For example, the structure body evaluation result includes four states, good, normal, inspection and abnormal, corresponding to the comprehensive evaluation result
Figure BDA0002067172610000052
Each component of (a). At this time
Figure BDA0002067172610000053
The probabilities of the structure corresponding to the four states are respectively 26.4%, 42.5%, 22.5% and 8.6%, and the structure is judged to be in a normal state according to the maximum probability.
A Global evaluation Model (GUM) with a structure as shown in FIG. 2 is adopted, a single-domain evaluation result data set (Evaluationsof Regions) is used as the input of the GUM, and an RNN-based Encoder-Decoder structure is usedRNN-based Encoder-Decoder, RED), an attention mechanism is introduced to extract structural features, and importance degrees of different single domains in global judgment of the structure are distinguished. Selecting a plurality of existing evaluation models as primary models, scoring the primary evaluation models according to different learning performances, and constructing secondary model input according to scoring differences. By means of a scoring method based on score differences (Score-Difference based StackingMethod, SD-Stacking) integrates the primary evaluation model, and realizes comprehensive evaluation of structural body safety. The method is specifically divided into two flows: structural body feature extraction based on an attention mechanism and model integration based on score difference. The specific implementation steps are described as follows:
(1) structural body feature extraction based on an attention mechanism:
(1.1) attention-attracting mechanism: in engineering safety monitoring practice, the physical characteristics of single-domain spatial structures, stress conditions and the like obtained by structural body division are different, and meanwhile, the stress condition of the structural body can be gradually changed along with the time. In different time periods, the influence of the single domain on the global evaluation is dynamically variable, and corresponding weights are distributed by using an attention mechanism, so that the feature expression is more reasonable, and the comprehensive evaluation accuracy is improved.
(1.2) encoding-decoding structure: the invention provides an Encode-decoder (RED) based on RNN, the structure is shown in figure 3, and the structure is in time
Figure BDA0002067172610000054
The single domain evaluation result set is
Figure BDA0002067172610000055
Each single-domain evaluation result is composed in the order of superscript, and is rewritten into the source sequence x ═ d1,d2,…,dm) Encoding and decoding to obtain the target sequence y ═ y (y)1,y2,…,yn). Similar to translations between different languages, the source and target sequences may not be equal in length, i.e., m ≠ n.
(1.3) RED encoding process: the Encoder (Encoder) reads each component in the input sequence x cyclically, updating the hidden state at step t:
Figure BDA0002067172610000056
where f is a nonlinear activation function. The coding process is to learn the attention weight distribution p (y) of the target sequence under the source sequence1,y2,…,yn|d1,d2,…,dm). Using RNN enables learning the attention weight p (d) on the source sequence by training the next component in the predicted source sequencet|dt-1,…,d1). Combining the attention weights at each step yields the attention distribution of the source sequence:
Figure BDA0002067172610000057
when the encoder loops to the end of the input sequence, the hidden state of the RNN will be updated to the corresponding abstract Feature Expression (FE) for the entire input sequence x, which will be the input to the decoder.
(1.4) RED decoding process: decoder (Decoder) for predicting a given hidden state based on the representation FE of the characteristics of the input sequence
Figure BDA0002067172610000061
Output sequence component y oftAnd finally training to obtain an output sequence y. The hidden states at step t are:
Figure BDA0002067172610000062
similar to the encoding process, the output sequence component y is calculatedtThe attention weight of (1) is:
Figure BDA0002067172610000063
wherein g and f in the encoding process are both nonlinear activation functions.
(1.5) encoding the objective function: jointly training the encoder and decoder to maximize the joint attention weight distribution of the output sequence given the input sequence, the objective function is then:
Figure BDA0002067172610000064
wherein the training set length is λ (corresponding to the preamble)
Figure BDA0002067172610000065
)，p(yi|xi) Is the target sequence yiIn the source sequence xiAttention weight distribution of the following. The encoding and decoding processes can be trivial, and a gradient-based approach can be used to minimize the objective function, thereby estimating the parameters.
The input and output sizes of RED are all variable. When the number m of single domains is large, dimension reduction can be performed on the single domain by encoding and decoding so that m > n. If m is of a moderate size, the output sequence size m ≈ n can be set. The encoding and decoding process is that the source sequence x is equal to (d)1,d2,…,dm) Weights are assigned and converted into more reasonable sequences y ═ y1,y2,…,yn) Therefore, the method is beneficial to improving the global comprehensive judgment training efficiency (dimensionality reduction) and accuracy (attention).
(2) Model integration based on score difference: the common comprehensive evaluation model in actual operation and maintenance of the dam comprises an Expert Weighting Method (EWM) and a support vector machine dam deformation evaluation based on wavelet transformationProfessional models such as a model (Wavelet Support Vector Machine, WSVM), and Neural Networks (NN). As shown in FIG. 4, the present invention provides a grading method based on score difference (Score-Difference based StackingMethod, SD-Stacking), integrating EWM, WSVM and NN as a primary model, selecting a Decision Tree (DT) as a secondary model, measuring the learning performance of the primary model by using learning scores, distributing weights for the primary model output by using score differences, improving the input of the secondary model, improving the integration performance of the model, and realizing comprehensive evaluation of the dam universe.
(2.1) primary model scoring: for the selected K elementary models { model { number1,model2,…,modelKDefine its learning performance score as { score1,score2,…,scoreK}. Let the scoring function be: rate (model)κ)＝scoreκK is 1,2, …, K. And if the comprehensive evaluation belongs to the classification task, the accuracy rate ACC and F1 are used as scoring functions. The ACC describes the overall classification accuracy of the model, i.e. the proportion of all correctly classified samples. F1 is the harmonic mean of the precision (probability that a certain class is correctly identified) and recall (proportion that the correct class is identified) and can measure the stability of the model. ACC and F1 both range from [0,1 ]]The closer the value is to 1, the better the corresponding performance is.
(2.2) score-based weight assignment: the sample data set is divided into two parts of Training and Testing. Averagely distributing Training set Training to K primary models to obtain { Training1,training2,…,trainingK}. Training and testing were performed on each of the primary models, memory modelκIn trailingκAnd the outputs of Testing are A respectivelyκAnd Bκ. Constructing a training set of secondary models using the outputs of the primary models [ A ]1；A2；…；AK]And test set mean (B)1,B2,…,BK). Based on the primary model scores, the weights of K primary models are recorded as { w }1,w2,…,wKDefine a scoring weight function as:
Figure BDA0002067172610000071
the weights are an expression of the learning performance of the different models on the sample data set and are thus adaptive. Rewriting the training set and test set of the secondary model to [ w ] using weights1A1；w2A2；…；wKAK]And mean (w)1B1,w2B2,…,wKBK)。
(2.3) training and comprehensive judgment of a secondary model: the second-order model (meta-model) is in [ w ]1A1；w2A2；…；wKAK]And mean (w)1B1,w2B2,…,wKBK) Training and testing. The learning performance of different primary models is distinguished by using the grading difference, and the secondary model can better utilize the characteristics of the primary models. When the learning performance of the primary model is close to that of the secondary model, the corresponding weight difference is small, and the performance improvement of the secondary model is limited. And when the learning performance of the primary model is obviously graded, the weight difference is increased, and at the moment, the strong model can be obviously promoted by utilizing the grading difference to inhibit the expression of the weak model, so that the precision of the secondary model is improved.

Claims (3)

1. A project safety evaluation method based on grading difference Stacking multi-model ensemble learning is characterized by comprising the following steps:
(1) extracting structural body features based on an attention mechanism;
(1.1) introducing an attention mechanism;
(1.2) constructing an encoding-decoding structure;
(1.3) RED encoding process;
(1.4) RED decoding process;
(1.5) acquiring a coding objective function;
(2) model integration based on score difference: integrating an expert weighting method EWM, a support vector machine dam deformation evaluation model WSVM based on wavelet transformation and a neural network NN as primary models, and selecting a decision tree DT as a secondary model; the learning performance of the primary model is measured by using the learning score, the score difference is used as the output of the primary model to distribute weight, and the input of the secondary model is improved;
the specific steps of the model integration based on the score difference in the step (2) are as follows:
(2.1) primary model scoring: for the selected K elementary models { model { number1,model2,…,modelKDefine its learning performance score as { score1,score2,…,scoreK}; let the scoring function be: rate (model)κ)＝scoreκK ═ 1,2, …, K; if the comprehensive evaluation belongs to the classification task, the accuracy rates ACC and F1 are used as scoring functions; the ACC describes the overall classification accuracy of the model, namely the proportion of all correctly classified samples; f1 is the harmonic mean value of the precision rate, namely the probability that a certain category is correctly identified, and the recall rate, namely the proportion that the correct category is identified, and can measure the stability of the model; ACC and F1 both range from [0,1 ]]The closer the value is to 1, the better the corresponding performance is;
(2.2) score-based weight assignment: dividing the sample data set into two parts of Training and Testing; averagely distributing Training set Training to K primary models to obtain { Training1,training2,…,trainingK}; training and testing were performed on each of the primary models, memory modelκIn trailingκAnd the outputs of Testing are A respectivelyκAnd Bκ(ii) a Constructing a training set of secondary models using the outputs of the primary models [ A ]1；A2；…；AK]And test set mean (B)1,B2,…,BK) (ii) a Based on the primary model scores, the weights of K primary models are recorded as { w }1,w2,…,wKDefine a scoring weight function as:
Figure FDA0003011346220000011
the weight is the embodiment of the learning performance of different models on the sample data set, so the method is self-adaptive; rewriting the training set and test set of the secondary model to [ w ] using weights1A1；w2A2；…；wKAK]And mean (w)1B1,w2B2,…,wKBK)；
(2.3) training and comprehensive judgment of a secondary model: the second-level model, i.e., the meta-model, is in [ w ]1A1；w2A2；…；wKAK]And mean (w)1B1,w2B2,…,wKBK) Training and testing; the learning performances of different primary models are distinguished by using the grading difference, and when the learning performances of the primary models are close to each other, the corresponding weight difference is small, so that the improvement on the performance of the secondary model is limited; when the learning performance of the primary model is obviously graded, the weight difference is increased, and at the moment, the strong model can be obviously improved by utilizing the grading difference, so that the weak model expression is inhibited, and the precision of the secondary model is improved;
the specific steps of the attention mechanism introduced in the step (1.1) are as follows: an RNN-based Encoder-Decoder structure is utilized, an attention mechanism is introduced to extract structural body characteristics, and the importance degree of different single domains in the global judgment of the structural body is distinguished;
the specific steps of constructing the encoding-decoding structure in the step (1.2) are as follows: the single-domain evaluation result set of the structure body in time T is DT＝[d1,d2,…,dm]The evaluation results of the single domains are composed in the order of superscript, and the result is rewritten into the source sequence x ═ d1,d2,…,dm) Encoding and decoding to obtain the target sequence y ═ y (y)1,y2,…,yn) (ii) a The lengths of the source sequence and the target sequence are unequal, namely m is not equal to n;
the RED encoding process in step (1.3) comprises the following specific steps: the Encoder reads each component in the input sequence x cyclically, updating the hidden state at step t:
Figure FDA0003011346220000021
wherein f is a nonlinear activation function; the coding process is to learn the attention weight distribution p (y) of the target sequence under the source sequence1,y2,…,yn|d1,d2,…,dm) (ii) a Using RNN can predict the next component in the source sequence by training, learn attention weights on the source sequencep(dt|dt-1,…,d1) (ii) a Combining the attention weights at each step yields the attention distribution of the source sequence:
Figure FDA0003011346220000022
when the encoder loops to the end of the input sequence, the hidden state of the RNN will be updated to the abstract feature representation FE corresponding to the entire input sequence x, which will be the input to the decoder.
2. The engineering safety evaluation method based on scoring difference Stacking multi-model ensemble learning as claimed in claim 1, wherein the RED decoding process in step (1.4) specifically comprises the following steps: decoder predicts given hidden state according to characteristic expression FE of input sequence
Figure FDA0003011346220000023
Output sequence component y oftFinally training to obtain an output sequence y; the hidden states at step t are:
Figure FDA0003011346220000024
similar to the encoding process, the output sequence component y is calculatedtThe attention weight of (1) is:
Figure FDA0003011346220000025
wherein g and f in the encoding process are both nonlinear activation functions.
3. The engineering safety evaluation method based on scoring difference Stacking multi-model ensemble learning as claimed in claim 1, wherein the specific steps of obtaining the coding objective function in step (1.5) are as follows: jointly training the encoder and decoder to maximize the joint attention weight distribution of the output sequence given the input sequence, the objective function is then:
Figure FDA0003011346220000031
whereinTraining set length of λ, p (y)i|xi) Is the target sequence yiIn the source sequence xiAttention weight distribution of the following.
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 CN110147614A (en) 2019-08-20
CN110147614B true 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)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
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

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8374974B2 (en) * 2003-01-06 2013-02-12 Halliburton Energy Services, Inc. Neural network training data selection using memory reduced cluster analysis for field model development
CN104200263B (en) * 2014-07-23 2017-05-10 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

Also Published As

Publication number Publication date
CN110147614A (en) 2019-08-20

Similar Documents

Publication Publication Date Title
CN112506990B (en) Hydrological data anomaly detection method based on spatiotemporal information
CN107992976B (en) Hot topic early development trend prediction system and prediction method
CN111563706A (en) Multivariable logistics freight volume prediction method based on LSTM network
CN101446283B (en) Evolutionary computation-based method for diagnosing compressor fault
US8615478B2 (en) Using affinity measures with supervised classifiers
CN106408423A (en) Method and system for risk assessment and method for constructing system for risk assessment
CN109685653A (en) A method of fusion deepness belief network and the monitoring of the credit risk of isolated forest algorithm
CN101480143A (en) Method for predicating single yield of crops in irrigated area
CN112756759B (en) Spot welding robot workstation fault judgment method
CN111062425B (en) Unbalanced data set processing method based on C-K-SMOTE algorithm
CN111339478B (en) Meteorological data quality assessment method based on improved fuzzy analytic hierarchy process
CN106778838A (en) A kind of method for predicting air quality
CN109271427A (en) A kind of clustering method based on neighbour's density and manifold distance
CN114036610A (en) Penetration depth prediction method based on data enhancement
CN113362913A (en) Method for predicting and optimizing gasoline octane number loss based on random forest regression
CN108062566A (en) A kind of intelligent integrated flexible measurement method based on the potential feature extraction of multinuclear
CN104732067A (en) Industrial process modeling forecasting method oriented at flow object
CN110147614B (en) Engineering safety evaluation method based on grading difference Stacking multi-model ensemble learning
CN110491443B (en) lncRNA protein correlation prediction method based on projection neighborhood non-negative matrix decomposition
CN109886316B (en) Transformer state parameter combination prediction method based on cloud system similarity weight distribution
CN110837853A (en) Rapid classification model construction method
CN116303626A (en) Well cementation pump pressure prediction method based on feature optimization and online learning
CN112599194B (en) Method and device for processing methylation sequencing data
Kumarasiri et al. Rainfall forecasting: an artificial neural network approach
CN113657441A (en) Classification algorithm based on weighted Pearson correlation coefficient and combined with feature screening

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