CN110084413A - Safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network - Google Patents
Safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network Download PDFInfo
- Publication number
- CN110084413A CN110084413A CN201910306543.9A CN201910306543A CN110084413A CN 110084413 A CN110084413 A CN 110084413A CN 201910306543 A CN201910306543 A CN 201910306543A CN 110084413 A CN110084413 A CN 110084413A
- Authority
- CN
- China
- Prior art keywords
- safety
- civil
- civil aviation
- layer
- model
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000013507 mapping Methods 0.000 claims abstract description 5
- 238000013139 quantization Methods 0.000 claims description 24
- 230000009467 reduction Effects 0.000 claims description 23
- 239000011159 matrix material Substances 0.000 claims description 21
- 230000008569 process Effects 0.000 claims description 18
- 238000005457 optimization Methods 0.000 claims description 13
- 238000000513 principal component analysis Methods 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 9
- 210000002569 neuron Anatomy 0.000 claims description 6
- 230000000644 propagated effect Effects 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 4
- 238000007619 statistical method Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 3
- 239000006185 dispersion Substances 0.000 claims description 3
- 230000009977 dual effect Effects 0.000 claims description 3
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 230000005484 gravity Effects 0.000 claims description 3
- 230000000717 retained effect Effects 0.000 claims description 3
- 238000013058 risk prediction model Methods 0.000 claims description 2
- 239000004615 ingredient Substances 0.000 claims 1
- 230000000694 effects Effects 0.000 abstract description 3
- 238000013178 mathematical model Methods 0.000 abstract description 3
- 238000004458 analytical method Methods 0.000 abstract 1
- 230000008901 benefit Effects 0.000 description 2
- 238000005315 distribution function Methods 0.000 description 2
- 210000005036 nerve Anatomy 0.000 description 2
- 238000007476 Maximum Likelihood Methods 0.000 description 1
- 238000005299 abrasion Methods 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 238000007664 blowing Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 239000011521 glass Substances 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
- 230000035939 shock Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component 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
- G06N3/088—Non-supervised learning, e.g. competitive learning
-
- 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"
-
- 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
- G06Q50/40—Business processes related to the transportation industry
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Strategic Management (AREA)
- Economics (AREA)
- General Physics & Mathematics (AREA)
- Entrepreneurship & Innovation (AREA)
- Data Mining & Analysis (AREA)
- Marketing (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Evolutionary Computation (AREA)
- Game Theory and Decision Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Quality & Reliability (AREA)
- General Health & Medical Sciences (AREA)
- Development Economics (AREA)
- Operations Research (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Educational Administration (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Primary Health Care (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network that the invention discloses a kind of, this method excavates the mapping relations between civil aviaton's unsafe incidents and safety of civil aviation risk index, using principal component analytical method and depth confidence network, safety of civil aviation risk forecast model is established.This method solve affecting parameters in civil aviaton's Risk Forecast Method in the prior art are very few, it is poor with the increase simple mathematical model prediction effect of data volume the problems such as.
Description
Technical field
The invention belongs to aviation safety technical fields, and in particular to one kind is capable of the prediction side of safety of civil aviation risk index
Method.
Background technique
Safety of civil aviation risk profile is the important content of safety of civil aviation risk management, the management to safety of civil aviation, especially
Identification potential risk and reasonable distribution civil aviation mangement resource and airline operation cost control all have a major impact.It is dangerous
The identification of event is it can be found that the danger source that accident occurs, finds out the risk factor that there may be, for promoting safety of civil aviation
Management effect and efficiency are significant.The height of risk index not only reflects the tight of unsafe incidents generation to a certain extent
Severe and frequency even more reflect safety of civil aviation level.Currently by simple mathematical model can only forecasting risk index, cannot
Deep relationship between further investigated risk index and unsafe incidents, and then can not accomplish in safety of civil aviation management to risk
The identification in source and the control in advance of potential risk, it is therefore necessary to excavate the deep layer between civil aviaton's unsafe incidents and risk profile
Secondary connection improves the accuracy of risk profile.
Summary of the invention
Goal of the invention: existing to solve the purpose of the present invention is to provide a kind of safety of civil aviation risk index prediction technique
Affecting parameters are very few in civil aviaton's Risk Forecast Method in technology, with the increase simple mathematical model prediction effect difference etc. of data volume
Problem.
Technical solution: the present invention adopts the following technical scheme:
Safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network, including safety of civil aviation risk profile
Modelling phase and safety of civil aviation risk index forecast period；The safety of civil aviation risk forecast model establishment stage includes such as
Lower step:
(1) the civil aviaton's unsafe incidents for extracting S, are divided into K class, to annual each unsafe incidents amount of progress
Change, obtains S civil aviaton unsafe incidents quantization matrix X=(X1,X2…Xk…XK)；Wherein XkFor the dangerous thing of S Niank class civil aviaton
Part quantization matrix, k=1,2 ..., K；Wherein xqFor S dimensional vector, q=1,2 ..., Qk, QkFor
The unsafe incidents number that kLei civil aviaton unsafe incidents are included；
Obtain the corresponding safety of civil aviation risk index Y=(y of S1,y2,…ys…yS)T；
(2) principal component analysis dimensionality reduction is used to every a kind of unsafe incidents, if kLei civil aviaton unsafe incidents quantization matrix
XkIt is after dimensionality reductionCorresponding low-dimensional feature space is Φk；
(3) establish the depth confidence model of safety of civil aviation risk profile, the model include limitation Bohr of L stacking hereby
Graceful machine, it is described L stacking limitation Boltzmann machine in first limit Boltzmann machine hidden layer be the l+1 limitation glass
The visible layer of the graceful machine of Wurz；The visible layer of first limitation Boltzmann machine has K in the limitation Boltzmann machine of the L stacking
The hidden layer of a node, l-th limitation Boltzmann machine has 1 node；L=1,2 ..., L-1；
(4) the depth confidence model of safety of civil aviation risk profile is trained, Optimized model parameter obtains safety of civil aviation
Risk forecast model；
The safety of civil aviation risk index forecast period includes:
Civil aviaton's unsafe incidents in year to be predicted are obtained, and classified according to the method in step (1) to it, measured
Change, obtain K class civil aviaton's unsafe incidents quantization vector in year to be predicted, every a kind of civil aviaton's unsafe incidents quantization vector is reflected
It is mapped to corresponding low-dimensional feature space in step 2, the vector z after obtaining dimensionality reductionk, the vector after K dimensionality reduction is input to training
In good safety of civil aviation risk forecast model, the output of model is the prediction of the safety of civil aviation risk index in year to be predicted
Value.
Value in step (1) after unsafe incidents quantization forms vector xq, xqIn s-th of element value be s, kth
The ratio of q-th unsafe incidents frequency and pilot time number, s=1,2 ..., S in class civil aviaton unsafe incidents.
To kLei civil aviaton unsafe incidents quantization matrix X in the step (2)kUsing PCA dimensionality reduction, it is after dimensionality reduction Tk<Qk, wherein TkFor the main component number retained after kth class unsafe incidents dimensionality reduction.It calculates
The corresponding specific gravity of each main component after dimensionality reduction: ω*=(ω1 *,ω2 *,…ωK *), wherein To retain after kLei civil aviaton unsafe incidents quantization matrix dimensionality reduction
T-th of main component corresponding to characteristic value；K=1,2 ..., K, t=1,2 ..., Tk。
The value of the number of nodes J of visible layer and hidden layer is located in the middle in the depth confidence model established in step (3)
Are as follows:
Or:
Or:
Wherein K is the classification number of civil aviaton's unsafe incidents, and d is the constant within [0,10].
In the present invention, the depth confidence model established in step (3) is stacked, this 4 by 4 limitation Boltzmann machines
The number of nodes of layer network can be expressed as [K, J1,J2,J3, 1], J1,J2,J3Respectively first, second, third limitation Bohr is hereby
The number of nodes of graceful machine hidden layer.It is to minimize following safety of civil aviation to the target of depth confidence model training in the step (4)
Risk profile energy function:
Wherein, parameter θ={ ω, a, b }, ω={ ωijBe visible layer and hidden layer weight；V={ viIt is visible layer
Input vector；H={ hiIt is hidden layer output vector；B={ bjBe hidden layer bias coeffcient, a={ aiIt is the inclined of visible layer
Lean on coefficient.
The training to depth confidence model includes that unsupervised training and backward finely tune train two stages；
Wherein unsupervised training is using civil aviaton's unsafe incidents quantization matrix after dimensionality reduction in step (2)As training sample, successively it is respectively trained since first limitation Boltzmann machine every
One limitation Boltzmann machine, primarily determines the parameter of depth confidence model；The training process of depth confidence model is by right
Sdpecific dispersion algorithm carries out gibbs sampler to data, is compared after reconstructed sample with original sample and determines frequency of training to subtract
Small reconstructed error completes the process of parameter optimization；
Forward direction stack limitation Boltzmann machine belong to unsupervised learning, unsupervised learning to model parameter have supervision
The initialization of learning parameter, is equivalent to and provides the priori knowledge of input data for supervised learning, and model training result is available
It advanced optimizes.Optimization process is gradually to finely tune mould to low layer using known label from depth confidence network the last layer
Shape parameter, referred to as after to fine tuning learn.It will be wanted according to " China Civil Aviation pacifies security information Statistical Analysis Report " is collected
Predict that the safety of civil aviation risk index in time as sample label, is trained after utilization to fine tuning, adjustment limitation Boltzmann machine
Parameter keeps corresponding error sufficiently small, so that model result is reached global optimization to complete Reverse optimization process.
The backward fine tuning training stage includes the following steps:
Rear into trim process, for the sample in training set, the corresponding activation value σ (Y of input layer is set1), output
Layer generate error beWherein Y4Indicate that output layer, r4 are that the last one RBM is obtained in propagated forward
Reconstructed error.⊙ indicates Ha Deman product, then the 4th layer of propagated error are as follows:Remainder layer and so on.Using gradient descent method, then can be obtained:
……
Other each layers and so on.
If finally obtained error is sufficiently small, corresponding parameter value at this time is exported, if error cannot receive, after
It is continuous carry out after trained to fine tuning, t2 is the number of iterations, adjust the number of iterations so that after to fine tuning training error it is sufficiently small.Its
To the learning rate of trim process after middle η expression, value range is (0.1,0.5).
Bias coeffcient b={ the b of the hidden layerjInitial value be set as the random number of [0,1] range；；Visible layer
Bias coeffcient a={ aiInitial value be set as the random number of [0,1] range；；Learning rate ε value range is (0.1,0.5)；It can
See weights omega={ ω of layer and hidden layerijIt is [K × a J1,J1×J2,J2×J3,J3×1]TThe matrix of dimension, in first K
×J1In array, the parameter initialization of corresponding position is contribution amount obtained in principal component analysis；I.e. first limitation Bohr is hereby
In graceful machine, it is seen that the initial weight ω of j-th of neuron of k-th of neuron of layer to hidden layerkjAre as follows:
Wherein ζ is the random number of [0,1] range；K=1,2 ..., K, j
=1,2 ..., J1。
Through the above steps, safety of civil aviation risk forecast model is established, safety of civil aviation risk is carried out using the model and refers to
The step of number prediction are as follows:
Civil aviaton's unsafe incidents in year to be predicted are obtained, and classified according to the method in step 1 to it, quantified,
K class civil aviaton's unsafe incidents quantization vector in year to be predicted is obtained, every a kind of civil aviaton's unsafe incidents are quantified into DUAL PROBLEMS OF VECTOR MAPPING
Vector z to low-dimensional feature space corresponding in step 2, after obtaining dimensionality reductionk, by the vector z after K dimensionality reductionk, k=1,2 ...,
K is input in trained safety of civil aviation risk forecast model, and the output of model is the safety of civil aviation risk in year to be predicted
The predicted value of index.
The utility model has the advantages that compared with prior art, method disclosed by the invention has the advantage that
(1) the safety of civil aviation risk forecast model combined using principal component analysis with depth confidence network, can be to the people
Boat unsafe incidents are analyzed, and sample utilization rate is improved, and it is from a wealth of sources to be more in line with danger source in civil aviaton's risk management, thing
The features such as part occurrence frequency difference is obvious, and association is mutually coupled between accident；
(2) input unsafe incidents are carried out to principal component analysis in the risk forecast model that depth confidence network combines
Then unsupervised learning have the reversed fine tuning of supervision to learn, makes using training result and security risk index sample label
Safety of civil aviation risk forecast model more fully respectively can take the chief in terms of feature extraction and prediction, improve the adaptation of model
Property and prediction result stability；
(3) principal component analysis is conducive to relatively fewer in data volume with the risk forecast model that depth confidence network combines
And data are made full use of in the relatively large number of situation of unsafe incidents type, and unsafe incidents information is deeply extracted, it is on the one hand real
The prediction of existing safety of civil aviation risk index, on the other hand for simple network model is difficult to solve and complexity is not achieved in the quality of data
The similar problems that network model requires provide certain resolving ideas；
(4) civil aviaton has been excavated not with the risk forecast model that depth confidence network combines by establishing principal component analysis
Profound Nonlinear Mapping relationship between security incident and risk profile is realized and is referred to according to the forecasting risk of civil aviaton's unsafe incidents
Number increases the information content of result while improving prediction result accuracy, so that airline be facilitated to carry out risk management, civil aviaton
The raising of industry progress security risk level.
Detailed description of the invention
Fig. 1 is the flow chart of safety of civil aviation risk index prediction technique disclosed by the invention；
Fig. 2 is first limitation Boltzmann machine structure chart；
Fig. 3 is depth confidence network structure.
Specific embodiment
With reference to the accompanying drawings and detailed description, the present invention is furture elucidated.
As shown in Figure 1, the invention discloses a kind of safety of civil aviation risk indexs based on PCA and depth confidence network to predict
Method, including safety of civil aviation risk forecast model establishment stage and safety of civil aviation risk index forecast period；Wherein safety of civil aviation
Risk forecast model establishment stage includes the following steps 1-4:
Step 1, the civil aviaton's unsafe incidents for extracting S, are divided into K class,
According to " unsafe incidents type and sample (2013 editions) ", in unsafe incidents type, first layer can be divided into
Including running in the air；Takeoff and anding；Aircraft is related；Ground handling；Airport/blank pipe；Weather and multiple major class including other；
The second layer defines the specific sample of unsafe incidents, mainly hits including bird, thrashing, a variety of specific including aborting etc.
Event.The inclusion relation between first layer and the second layer is defined simultaneously, as the unsafe incidents for including in weather major class have thunder
It hits, shock by electricity；Wind shear；Bucketing；Ice hit including four kinds, in aircraft correlation major class comprising blowing out, tire falls off, pricks
It is broken；Thrashing, failure, card resistance；Element falling, damage, abrasion；Aircraft is on fire to smolder；Engine cut-off；It is unsatisfactory for flying
Six kinds including condition.
Unsafe incidents are divided into K class, every one kind includes multiple unsafe incidents according to the division of first layer by the present invention
Specific sample.The civil aviaton's unsafe incidents for extracting S quantify annual each unsafe incidents, obtain S civil aviaton uneasiness
Total event quantization matrix X=(X1,X2…Xk…XK)；Wherein XkFor S Niank class civil aviaton unsafe incidents quantization matrix, k=1,
2,…,K；Wherein xqFor S dimensional vector, q=1,2 ..., Qk, QkFor the dangerous thing of kLei civil aviaton
The unsafe incidents number that part is included；Value after unsafe incidents quantization forms vector xq, xqIn the value of s-th of element be
The ratio of q-th unsafe incidents frequency and pilot time number in s, kLei civil aviaton unsafe incidents, s=1,
2,…,S。
Obtain the corresponding safety of civil aviation risk index Y=(y of S1,y2,…ys…yS)T；
According to " China Civil Aviation pacifies security information Statistical Analysis Report ", the corresponding S that available civil aviation authority calculates
Safety of civil aviation risk index Y=(y1,y2,…ys…yS)T。
Step 2 uses principal component analysis dimensionality reduction to every a kind of unsafe incidents, if kLei civil aviaton unsafe incidents quantify
Matrix XkIt is after dimensionality reductionCorresponding low-dimensional feature space is Φk；
To kLei civil aviaton unsafe incidents quantization matrix XkUsing PCA dimensionality reduction, it is after dimensionality reduction Tk<Qk, wherein TkFor the main component number retained after kLei civil aviaton unsafe incidents dimensionality reduction.
Calculate the corresponding specific gravity of each main component after dimensionality reduction: ω*=(ω1 *,ω2 *,…ωK *) wherein, To retain after kLei civil aviaton unsafe incidents quantization matrix dimensionality reduction
T-th of main component corresponding to characteristic value；K=1,2 ..., K, t=1,2 ..., Tk。
Step 3, the depth confidence model for establishing safety of civil aviation risk profile
The limitation Boltzmann machine number of depth confidence model influences whether calculation amount, according to input data in the present invention
Amount, by limitation Boltzmann machine number setting at 3-5, in the present embodiment, depth confidence model is by 4 limitation Boltzmann machines
It stacks.The visible layer of first RBM (Restricted Boltzmann Machines limits Boltzmann machine) has K
Node, the hidden layer of the last one RBM have 1 node.The number of nodes J of visible layer and hidden layer is located in the middle referring to BP nerve
The node setting empirical value of network carrys out value, and common empirical equation has:
Or:
Or:
Wherein K is the classification number of civil aviaton's unsafe incidents, and d is the constant within [0,10].The number of nodes of 4 layer network can
To be expressed as [K, J1,J2,J3, 1], J1,J2,J3The node of respectively first, second, third limitation Boltzmann machine hidden layer
Number.In general, the number of nodes of middle layer has following relationship: K < J1,1<J3<J2.The structure of first RBM as shown in Fig. 2,
The structure chart of entire depth confidence network is as shown in Figure 3.
Step 4 is trained the depth confidence model of safety of civil aviation risk profile, and Optimized model parameter obtains civil aviaton
Security risk prediction model；
Target to depth confidence model training is to minimize following safety of civil aviation risk profile energy function:
Wherein, parameter θ={ ω, a, b }, ω={ ωijBe visible layer and hidden layer weight；V={ viIt is visible layer
Input vector；H={ hiIt is hidden layer output vector；B={ bjBe hidden layer bias coeffcient, a={ aiIt is the inclined of visible layer
Lean on coefficient.By continuing to optimize parameter ω, b and a, so that safety of civil aviation risk profile energy function be made to obtain minimum value.
Energy function is abbreviated as E, to energy function indexation and regularization later it can be concluded that hidden layer h and visible layer
Joint probability distribution function between v means the probability being observed in the state of ENERGY E:
WhereinZ (θ) is abbreviated as Z.
Pass through the joint probability distribution letter between the hidden layer h and visible layer v to safety of civil aviation risk profile energy function
Several definition can derive the conditional value at risk between hidden layer h and visible layer v, so that it is pre- to define safety of civil aviation risk
Network node in model is surveyed to be activated the probability function handed on.
By joint probability distribution function it follows that
So as to find out the conditional probability distribution of h and v are as follows:
For convenience of derivation, if the neuron in h and v is two-dimensional random unit, it is possible to derive in known visible layer
In the case where hide member hjThe probability of (i.e. value be 1) of being activated isIt can similarly obtain
In the case of known hidden layer, it is seen that first viThe probability of (i.e. value be 1) of being activated is
Boltzmann machine is limited by repetitive exercise come learning parameter θ={ ω, a, b } value, to extract training set
Feature.It is indicated with maximum likelihood function are as follows:
Using stochastic gradient rise method Optimal Parameters θ, if ε is the learning rate of pre-training, iterative formula are as follows:
Then approximation sample is carried out to reconstruct data using contrast divergence algorithm, the iteration that can obtain parameter θ updates public affairs
Formula:
Δωij=ε (< vihj>data-<vihj>recon)
Δai=ε (< vi>data-<vi>recon)
Δbi=ε (< hj>data-<hj>recon)
Wherein,<>dataFor the mathematic expectaion of training sample set；<>reconFor the mathematic expectaion of reconstruction model output.
The training of limitation Boltzmann machine is exactly to obtain the hiding feature of depth confidence network by iterating.
Training to depth confidence model includes that unsupervised training and backward finely tune train two stages；
Wherein unsupervised training is using civil aviaton's unsafe incidents quantization matrix after dimensionality reduction in step (2)As training sample, successively it is respectively trained since first limitation Boltzmann machine every
One limitation Boltzmann machine, primarily determines the parameter of depth confidence model；The training process of depth confidence model is by right
Sdpecific dispersion algorithm carries out gibbs sampler to data, is compared after reconstructed sample with original sample and determines frequency of training to subtract
Small reconstructed error completes the process of parameter optimization；
Forward direction stack limitation Boltzmann machine belong to unsupervised learning, unsupervised learning to model parameter have supervision
The initialization of learning parameter, is equivalent to and provides the priori knowledge of input data for supervised learning, and model training result is available
It advanced optimizes.Optimization process is gradually to finely tune mould to low layer using known label from depth confidence network the last layer
Shape parameter, referred to as after to fine tuning learn.It will be wanted according to " China Civil Aviation pacifies security information Statistical Analysis Report " is collected
Predict that the safety of civil aviation risk index in time as sample label, is trained after utilization to fine tuning, adjustment limitation Boltzmann machine
Parameter keeps corresponding error sufficiently small, to make model result reach global optimization to optimization process after completing.
The backward fine tuning training stage includes the following steps:
Rear into trim process, for the sample in training set, the corresponding activation value σ (Y of input layer is set1), output
Layer generate error beWherein Y4Indicate that output layer, r4 are that the last one RBM is obtained in propagated forward
Reconstructed error.⊙ indicates Ha Deman product, then the 3rd layer of propagated error are as follows:Remainder layer and so on.Using gradient descent method, then can be obtained:
……
Other each layers and so on.
If finally obtained error is sufficiently small, corresponding parameter value at this time is exported, if error cannot receive, after
It is continuous carry out after trained to fine tuning, t2 is the number of iterations, adjust the number of iterations so that after to fine tuning training error it is sufficiently small.Its
To the learning rate of trim process after middle η expression, value range is (0.1,0.5).
Bias coeffcient b={ the b of the hidden layerjInitial value be set as the random number of [0,1] range；Visible layer it is inclined
Lean on coefficient a={ aiInitial value be set as the random number of [0,1] range；Learning rate ε value range is (0.1,0.5)；It can be seen that
Weights omega={ ω of layer and hidden layerijIt is [K × a J1,J1×J2,J2×J3,J3×1]TThe matrix of dimension.In order to accelerate to instruct
The optimal speed of parameter during white silk, in first K × J1In array, the parameter initialization of corresponding position is in principal component analysis
Obtained contribution amount；In i.e. first limitation Boltzmann machine, it is seen that j-th of nerve of k-th of neuron of layer to hidden layer
The initial weight ω of memberkjAre as follows:
Wherein ζ is the random number of [0,1] range；K=1,2 ..., K, j
=1,2 ..., J1。。
Through the above steps, safety of civil aviation risk forecast model is established, safety of civil aviation risk is carried out using the model and refers to
The step of number prediction are as follows:
Civil aviaton's unsafe incidents in year to be predicted are obtained, and classified according to the method in step 1 to it, quantified,
K class civil aviaton's unsafe incidents quantization vector in year to be predicted is obtained, every a kind of civil aviaton's unsafe incidents are quantified into DUAL PROBLEMS OF VECTOR MAPPING
Vector z to low-dimensional feature space corresponding in step 2, after obtaining dimensionality reductionk, by the vector z after K dimensionality reductionk, k=1,2 ...,
K is input in trained safety of civil aviation risk forecast model, and the output of model is the safety of civil aviation risk in year to be predicted
The predicted value of index.
Claims (8)
1. the safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network, which is characterized in that including safety of civil aviation
Risk forecast model establishment stage and safety of civil aviation risk index forecast period；The safety of civil aviation risk forecast model establishes rank
Section includes the following steps:
(1) the civil aviaton's unsafe incidents for extracting S, are divided into K class, quantify to annual each unsafe incidents, obtain
To S civil aviaton unsafe incidents quantization matrix X=(X1,X2…Xk…XK)；Wherein XkFor S Niank class civil aviaton unsafe incidents amount
Change matrix, k=1,2 ..., K；Wherein xqFor S dimensional vector, q=1,2 ..., Qk, QkFor kth class
The unsafe incidents number that civil aviaton's unsafe incidents are included；
Obtain the corresponding safety of civil aviation risk index Y=(y of S1,y2,…ys…yS)T；
(2) principal component analysis dimensionality reduction is used to every a kind of unsafe incidents, if kLei civil aviaton unsafe incidents quantization matrix XkDrop
It is after dimensionCorresponding low-dimensional feature space is Φk；
(3) the depth confidence model of safety of civil aviation risk profile is established, the model includes the limitation Boltzmann of L stacking
Machine, it is described L stacking limitation Boltzmann machine in first limit Boltzmann machine hidden layer be the l+1 limitation Bohr
The hereby visible layer of graceful machine；The visible layer of first limitation Boltzmann machine has K in the limitation Boltzmann machine of the L stacking
The hidden layer of node, l-th limitation Boltzmann machine has 1 node；L=1,2 ..., L-1；
(4) the depth confidence model of safety of civil aviation risk profile is trained, Optimized model parameter obtains safety of civil aviation risk
Prediction model；
The safety of civil aviation risk index forecast period includes:
Civil aviaton's unsafe incidents in year to be predicted are obtained, and classified according to the method in step (1) to it, quantified, are obtained
K class civil aviaton's unsafe incidents to year to be predicted quantify vector, and every a kind of civil aviaton's unsafe incidents quantization DUAL PROBLEMS OF VECTOR MAPPING is arrived
Corresponding low-dimensional feature space in step 2, the vector z after obtaining dimensionality reductionk, the vector after K dimensionality reduction is input to trained
In safety of civil aviation risk forecast model, the output of model is the predicted value of the safety of civil aviation risk index in year to be predicted.
2. the safety of civil aviation risk index prediction technique according to claim 1 based on PCA Yu depth confidence network, special
Sign is that the value in step (1) after unsafe incidents quantization forms vector xq, xqIn s-th of element value be s, kth
The ratio of q-th unsafe incidents frequency and pilot time number, s=1,2 ..., S in class civil aviaton unsafe incidents.
3. the safety of civil aviation risk index prediction technique according to claim 1 based on PCA Yu depth confidence network, special
Sign is, to kLei civil aviaton unsafe incidents quantization matrix X in the step (2)kUsing PCA dimensionality reduction, it is after dimensionality reduction Tk<Qk, TkFor the main component number retained after kth class dimensionality reduction；Each master after calculating dimensionality reduction
Want the corresponding specific gravity of ingredient: ω*=(ω1 *,ω2 *,…ωK *), wherein To retain after kLei civil aviaton unsafe incidents quantization matrix dimensionality reduction
T-th of main component corresponding to characteristic value；K=1,2 ..., K, t=1,2 ..., Tk。
4. the safety of civil aviation risk index prediction technique according to claim 1 based on PCA Yu depth confidence network, special
Sign is, the value of the number of nodes J of visible layer and hidden layer is located in the middle in the depth confidence model established in step (3)
Are as follows:
Or:
Or:
Wherein K is the classification number of civil aviaton's unsafe incidents, and d is the constant within [0,10].
5. the safety of civil aviation risk index prediction technique according to claim 3 based on PCA Yu depth confidence network, special
Sign is that the depth confidence model established in step (3) is stacked, the section of 4 layer network by 4 limitation Boltzmann machines
Points can be expressed as [K, J1,J2,J3, 1], J1,J2,J3Respectively first, second, third limitation Boltzmann machine hidden layer
Number of nodes.
6. the safety of civil aviation risk index prediction technique according to claim 5 based on PCA Yu depth confidence network, special
Sign is, is to minimize following safety of civil aviation risk profile energy to the target of depth confidence model training in the step (4)
Function:
Wherein, parameter θ={ ω, a, b }, ω={ ωijBe visible layer and hidden layer weight；V={ viIt is that visible layer inputs
Vector；H={ hiIt is hidden layer output vector；B={ bjBe hidden layer bias coeffcient, a={ aiBe visible layer bias system
Number.
7. the safety of civil aviation risk index prediction technique according to claim 6 based on PCA Yu depth confidence network, special
Sign is that the training to depth confidence model includes that unsupervised training and backward finely tune train two stages；
Wherein unsupervised training is using civil aviaton's unsafe incidents quantization matrix after dimensionality reduction in step (2)As training sample, successively it is respectively trained since first limitation Boltzmann machine every
One limitation Boltzmann machine, primarily determines the parameter of depth confidence model；The training process of depth confidence model is by right
Sdpecific dispersion algorithm carries out gibbs sampler to data, is compared after reconstructed sample with original sample and determines frequency of training to subtract
Small reconstructed error completes the process of parameter optimization；
The limitation Boltzmann machine that forward direction stacks belongs to unsupervised learning, the model parameter supervised learning that unsupervised learning is arrived
The initialization of parameter, is equivalent to and provides the priori knowledge of input data for supervised learning, and model training result can be obtained into one
Step optimization；Optimization process is gradually to be joined to low layer fine tuning model from depth confidence network the last layer using known label
Number, referred to as after to fine tuning learn；It will be predicted according to " China Civil Aviation pacifies security information Statistical Analysis Report " is collected
The safety of civil aviation risk index in time is trained after utilization to fine tuning as sample label, the parameter of adjustment limitation Boltzmann machine,
Keep corresponding error sufficiently small, so that model result is reached global optimization to complete Reverse optimization process；
The backward fine tuning training stage includes the following steps:
Rear into trim process, for the sample in training set, the corresponding activation value σ (Y of input layer is set1), output layer generates
Error beWherein Y4Indicate output layer, r4For the reconstruct that the last one RBM is obtained in propagated forward
Error；⊙ indicates Ha Deman product, then the 3rd layer of propagated error are as follows:Its
Remaining layer and so on；Using gradient descent method, then can be obtained:
Other each layers and so on；
If finally obtained error is sufficiently small, export corresponding parameter value at this time, if error cannot receive, continue into
After row to fine tuning train, t2 is the number of iterations, adjust the number of iterations so that after to fine tuning training error it is sufficiently small；Wherein η table
To the learning rate of trim process after showing, value range is (0.1,0.5).
8. the safety of civil aviation risk index prediction technique according to claim 7 based on PCA Yu depth confidence network, special
Sign is, the bias coeffcient b={ b of the hidden layerjInitial value be set as the random number of [0,1] range；Visible layer it is inclined
Lean on coefficient a={ aiInitial value be set as the random number of [0,1] range；Learning rate ε value range is (0.1,0.5)；It can be seen that
Weights omega={ ω of layer and hidden layerijIt is [K × a J1,J1×J2,J2×J3,J3×1]TThe matrix of dimension, first K ×
J1In array, the parameter initialization of corresponding position is contribution amount obtained in principal component analysis；I.e. first limitation Boltzmann
In machine, it is seen that the initial weight ω of j-th of neuron of k-th of neuron of layer to hidden layerkjAre as follows:
Wherein ζ is the random number of [0,1] range；K=1,2 ..., K, j=1,
2,…,J1。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910306543.9A CN110084413A (en) | 2019-04-17 | 2019-04-17 | Safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910306543.9A CN110084413A (en) | 2019-04-17 | 2019-04-17 | Safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110084413A true CN110084413A (en) | 2019-08-02 |
Family
ID=67415343
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910306543.9A Pending CN110084413A (en) | 2019-04-17 | 2019-04-17 | Safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110084413A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329974A (en) * | 2020-09-03 | 2021-02-05 | 中国人民公安大学 | LSTM-RNN-based civil aviation security event behavior subject identification and prediction method and system |
| CN112429252A (en) * | 2020-11-24 | 2021-03-02 | 中国人民解放军空军预警学院 | Flight emergency prediction method based on PCA algorithm |
| CN112633361A (en) * | 2020-12-20 | 2021-04-09 | 中国人民解放军空军预警学院 | Flight emergency prediction method and device based on LSTM neural network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106709820A (en) * | 2017-01-11 | 2017-05-24 | 中国南方电网有限责任公司电网技术研究中心 | Electrical power system load prediction method and device based on depth belief network |
| CN107609630A (en) * | 2017-08-02 | 2018-01-19 | 广东建设职业技术学院 | A kind of depth confidence network parameter optimization method and system based on artificial bee colony |
-
2019
- 2019-04-17 CN CN201910306543.9A patent/CN110084413A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106709820A (en) * | 2017-01-11 | 2017-05-24 | 中国南方电网有限责任公司电网技术研究中心 | Electrical power system load prediction method and device based on depth belief network |
| CN107609630A (en) * | 2017-08-02 | 2018-01-19 | 广东建设职业技术学院 | A kind of depth confidence network parameter optimization method and system based on artificial bee colony |
Non-Patent Citations (4)
| Title |
|---|
| XIAOMEI NI 等: "Civil aviation safety evaluation based on deep belief network and principal component analysis", 《SAFETY SCIENCE》 * |
| 倪晓梅 等: "考虑不平衡数据的民航不安全事件分析", 《航空计算技术》 * |
| 王攀 等: "《优化与控制中的软计算方法研究》", 31 January 2017 * |
| 苏欣等著: "《Android手机应用网络流量分析与恶意行为检测研究》", 31 October 2016 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112329974A (en) * | 2020-09-03 | 2021-02-05 | 中国人民公安大学 | LSTM-RNN-based civil aviation security event behavior subject identification and prediction method and system |
| CN112329974B (en) * | 2020-09-03 | 2024-02-27 | 中国人民公安大学 | LSTM-RNN-based civil aviation security event behavior subject identification and prediction method and system |
| CN112429252A (en) * | 2020-11-24 | 2021-03-02 | 中国人民解放军空军预警学院 | Flight emergency prediction method based on PCA algorithm |
| CN112429252B (en) * | 2020-11-24 | 2022-05-03 | 中国人民解放军空军预警学院 | Flight emergency prediction method based on PCA algorithm |
| CN112633361A (en) * | 2020-12-20 | 2021-04-09 | 中国人民解放军空军预警学院 | Flight emergency prediction method and device based on LSTM neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109492822B (en) | Air pollutant concentration time-space domain correlation prediction method | |
| CN106355151B (en) | A kind of three-dimensional S AR images steganalysis method based on depth confidence network | |
| US20210174262A1 (en) | Deep unsupervised learning approach , device and storage medium for airspace complexity evaluation | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN106778921A (en) | Personnel based on deep learning encoding model recognition methods again | |
| Zhao et al. | Automatic recognition of loess landforms using Random Forest method | |
| CN102800093B (en) | Based on the multiple-target remote sensing image dividing method decomposed | |
| CN110084413A (en) | Safety of civil aviation risk index prediction technique based on PCA Yu depth confidence network | |
| CN108345860A (en) | Personnel based on deep learning and learning distance metric recognition methods again | |
| CN109583635B (en) | Short-term load prediction modeling method for operational reliability | |
| CN105701502A (en) | Image automatic marking method based on Monte Carlo data balance | |
| CN108805167A (en) | L aplace function constraint-based sparse depth confidence network image classification method | |
| CN109948716A (en) | A kind of airport delay prediction technique based on region residual sum LSTM network | |
| CN111126820B (en) | Method and system for preventing electricity stealing | |
| CN102185735A (en) | Network security situation prediction method | |
| CN107423705A (en) | SAR image target recognition method based on multilayer probability statistics model | |
| CN115564114B (en) | Airspace carbon emission short-term prediction method and system based on graph neural network | |
| CN102663264A (en) | Semi-supervised synergistic evaluation method for static parameter of health monitoring of bridge structure | |
| CN114444378A (en) | Short-term power prediction method for regional wind power cluster | |
| CN110009030A (en) | Sewage treatment method for diagnosing faults based on stacking meta learning strategy | |
| CN111680875A (en) | Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model | |
| CN114492191A (en) | Heat station equipment residual life evaluation method based on DBN-SVR | |
| CN110443448A (en) | A kind of aircraft seat in the plane classification prediction technique and system based on two-way LSTM | |
| CN104318515A (en) | Hyper-spectral image wave band dimension descending method based on NNIA evolutionary algorithm | |
| CN104008402A (en) | Foundation cloud picture recognition method based on improved SOM algorithm |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190802 |
|
| RJ01 | Rejection of invention patent application after publication |