A kind of runway Surface airworthiness method of discrimination based on pattern-recognition
Technical field
The invention belongs to artificial intelligence, big data analysis and computer fields, and in particular to a kind of based on pattern-recognition
Runway Surface airworthiness method of discrimination,
Background technique
Runway Surface causes under coefficient of friction because of the frequent landing of aircraft there are rubber scratch or sleet
Drop, seriously affects aviation safety.It needs to be detected by means of runway coefficient of friction currently, the airworthiness of runway Surface differentiates
Vehicle, generally from sport car repacking.The floor manager personnel on airport drive round-trip repeatedly high speed row on coefficient of friction detection vehicle to runway
It sails, generates road face coefficient of friction, conventional method investment is huge and time-consuming and laborious, and the airport busy in navigational matters, is often difficult to look for
This work is completed to time slot.
Summary of the invention
To solve the above problems, the present invention provides a kind of runway Surface airworthiness differentiation side based on pattern-recognition
Method.
The technical scheme adopted by the invention is that:
A kind of runway Surface airworthiness method of discrimination based on pattern-recognition, it is characterised in that: the following steps are included:
A. tolerance fuzzy clustering algorithm is used, data sample is clustered and is pre-processed;B. optimize hidden layer mind in RBF neural
Through first quantity, to obtain optimal real-time;C. RBF neural is trained;D. using multidimensional compensation method to algorithm
Part optimize, realize and complete most accurately to differentiate that result is preferred with least hidden layer neuron number, the step
The specific implementation of rapid A are as follows:
The mark image near airfield runway is acquired, and using mark image after testing as sample set, it is assumed that n d
Tieing up the sample set that sample is constituted is Dset={ IM1,IM2,…,IMn, IMi∈Rd, wherein IMiFor a mark image sample;?
During airworthiness differentiates, cluster result is divided into C class;Degree of membership of the sample in tolerance fuzzy clustering is denoted as gij, indicate sample
This IMiIt is subordinate to the degree of membership of j classification;Introducing tolerance θ makes sample in parameter set PcRelative to central value McVariance be D (IMi,
McPc)=(IMi-Mc)T(IMi-Mc)-θ;
After introducing tolerance, 0 two kinds of situations of >=0 and < will occur in the variance of sample, define two set SetNAnd SetpRespectively
Indicate Sample Similarity in this case:
Therefore, the smallest category set Min of Sample Similarityi={ c, D (IMi,MjPj)≥D(IMi,McPc), for not
Same fuzzy coefficient α has following degree of membership formula:
Based on the above, tolerance fuzzy clustering processing can be carried out to sample set, steps are as follows:
(1) central value collection is initialized
(2) sample in circular treatment training set calculates its Euclidean distance Diste, to find corresponding training set sample
Minimum central valueMeet following condition:
(1≤i≤C), wherein t indicates the number of iterations;
(3) each cyclic process, while updating per one-dimensional central value:
Wherein, ρtIt is a system function with property of successively decreasing, enables
Wherein, 0 < ρ0< 1 can use random number;T is training total degree;T is current iteration step number;
To guarantee that network has artificial intelligence interface in use, tolerance θ is introduced using error function θ (t), error function
Determination to can be experience specified, can also be accustomed to being fitted according to user, in the initialization of system, enableN
For total sample number, δtFor system control function, sought using gradient descent method:
(4) according to classification degree of membershipGenerate the tolerance fuzzy clustering of sample:
3. the degree of membership formula of sample can be obtained by 4. 5. bringing formula into formula 2. are as follows:
Preferably, the specific implementation of the step B are as follows:
In an iterative process, after the iterative steps for reaching setting, all types of cluster results is calculated, if some type of
It clusters sample number and is less than threshold value ε, then just the classification is fitted in neighborhood class: assuming that iteration finishes central value generated
Collection is combined into M={ M1,M2,…,Md, after a clustering processing, set of types of the sample number less than threshold value ε is combined into Setε
={ ST1,ST2,…,STk(1≤k≤C), Then by door
Limit value ε set of types below is deleted, and the sample under it is taken out, and calculates the Euclidean distance with the central value in set of types:
After finding out minimum euclidean distance, the index of set of types locating for central value can be obtained, sample set is finally fitted to neighbour
The neighborhood fitting of neuron is completed in the class set of domain.
Preferably, the specific implementation of the step C are as follows:
The parameter of neural network is set are as follows:
(1) training sample set is inputted are as follows: the training set after tolerance clustering processing:
(2) desired output vector
(3) hidden layer-output layer connection weight is initialized:
Wherein, p, q are respectively output layer and hidden layer neuron number;
Initialization procedure use following formula:
(4) the iteration formula of connection weight, hidden layer neuron central value and width is as follows:
Wherein (t-1), t, (t+1) indicate three states of iteration, CjiIndicate hidden layer neuron central value, djiIt indicates
Width, α indicate step pitch, and β indicates that Studying factors, E (t) are assessing network function.
Preferably, the specific implementation of the step D are as follows:
(1) neuron of basic RBF neural hidden layer described in step C calculates one by oneAnd draw
Enter threshold tau, incorporates the sample that all distances are greater than threshold tau into compensation training collection, the size of τ is adjusted according to live atlas, is made
Obtaining compensation training and concentrating sample size is the 30% of the training set of basic RBF neural;
(2) the tolerance fuzzy clustering method in recycle step A clusters compensation training collection;
(3) by the basic RBF neural in new cluster result sample set input step C, its radial base letter is set
Number are as follows:
(4) new neural network is trained with the training set of basic RBF neural, after multidimensional compensation can be obtained
Result.
The invention has the benefit that
The rubber trace or sleet covering trace for a large amount of runway counterpoise groundings that the present invention only needs to generate using runway monitoring
Picture can complete the differentiation of runway Surface airworthiness by neural network learning and pattern-recognition in real time, without artificial
It is participated in other equipment, has the advantages that distinguishing speed is fast, high-efficient and safety and stability.
Detailed description of the invention
Fig. 1 is the basic block diagram of RBF neural.
Fig. 2 is the runway Surface airworthiness method of discrimination overall flow figure based on pattern-recognition.
Specific embodiment
As shown in Fig. 2, the present invention is based on the runway Surface airworthiness method of discrimination of pattern-recognition, including following step
It is rapid:
A. tolerance fuzzy clustering algorithm is used, data sample is clustered and is pre-processed:
This method mainly acquires the two-way counterpoise grounding of runway, the mark image near two-way holding point of taking off, to reach structure
Simply, training is succinct, and guarantees the high-precision of algorithm, and main body uses RBF neural, realizes succinct linear learning, reaches
To the high-precision effect of Nonlinear Learning.On the other hand, the quantity of the cluster and hidden layer neuron of data sample is to RBF mind
Performance through network has conclusive influence, and traditional clustering algorithm rule is excessively harsh, in the pixel samples of mark image
The clustering processing effect of collection is undesirable, therefore the present invention proposes tolerance fuzzy clustering algorithm, has the effect of very ideal, algorithm
It is as follows:
Using mark image largely after testing as sample set, it is assumed that such n d ties up the sample set that sample is constituted and be
Dset={ IM1,IM2,…,IMn, IMi∈Rd, wherein IMiFor a mark image sample;During airworthiness differentiates, it can incite somebody to action
Differentiate that result is divided into two classes or three classes, for convenience of stating, cluster result is divided into C class;Person in servitude of the sample in tolerance fuzzy clustering
Category degree is denoted as gij, indicate sample IMiIt is subordinate to the degree of membership of j classification;Tolerance θ is introduced, so that sample is in parameter set PcRelative in
Center value McVariance be D (IMi,McPc)=(IMi-Mc)T(IMi-Mc)-θ。
After introducing tolerance, 0 two kinds of situations of >=0 and < will occur in the variance of sample, define two set SetNAnd SetpRespectively
Indicate Sample Similarity in this case:
Therefore, the smallest category set Min of Sample Similarityi={ c, D (IMi,MjPj)≥D(IMi,McPc), for not
Same fuzzy coefficient α has following degree of membership formula:
Based on the above, tolerance fuzzy clustering processing can be carried out to sample set, steps are as follows:
(1) central value collection is initialized
(2) sample in circular treatment training set calculates its Euclidean distance Diste, to find corresponding training set sample
Minimum central valueMeet following condition:
(1≤i≤C);Wherein t indicates the number of iterations.
(3) each cyclic process, while updating per one-dimensional central value, the method is as follows:
Wherein, ρtIt is a system function with property of successively decreasing, can enables in this example
Wherein, 0 < ρ0< 1 can use random number;T is training total degree;T is current iteration step number.
To guarantee that network has artificial intelligence interface in use, tolerance θ is introduced using error function θ (t), error function
Determination to can be experience specified, can also be accustomed to being fitted according to user.In the initialization of system, enableN
For total sample number.
δtFor system control function, sought using gradient descent method:
(4) according to classification degree of membershipThe tolerance fuzzy clustering of sample is generated, algorithm is as follows:
3. the degree of membership formula of sample can be obtained by 4. 5. bringing formula into formula 2. are as follows:
B. optimize hidden layer neuron quantity in RBF neural, to obtain optimal real-time:
The quantity of hidden layer neuron affects the real-time of training algorithm, and excessive neuronal quantity will not only be brought
Higher accuracy seriously affects timeliness instead, therefore the present invention optimizes the number of hidden layer neuron using feedback mechanism
Amount obtains optimal real-time under the premise of guaranteeing certain accuracy, the method is as follows:
In an iterative process, after the iterative steps for reaching setting, all types of cluster results is calculated, if some type of
It clusters sample number and is less than threshold value ε, then just the classification is fitted in neighborhood class.
Assuming that it is M={ M that iteration, which finishes center value set generated,1,M2,…,Md, after a clustering processing, sample
Set of types of this number less than threshold value ε is combined into Setε={ ST1,ST2,…,STk(1≤k≤C),
Then threshold value ε set of types below is deleted, and the sample under it is taken out, is calculated and the central value in set of types
Euclidean distance:
After finding out minimum euclidean distance, the index of set of types locating for central value can be obtained, sample set is finally fitted to neighbour
The neighborhood fitting of neuron is completed in the class set of domain.
C. RBF neural is trained:
Tolerance fuzzy clustering algorithm is completed, after data sample is clustered and is pre-processed, the basic knot of neural network
Structure is as shown in Figure 2.
The parameter of neural network is set are as follows:
(1) training sample set is inputted are as follows: the training set after tolerance clustering processing:
(2) desired output vector
(3) hidden layer-output layer connection weight is initialized:
Wherein, p, q are respectively output layer and hidden layer neuron number.
Initialization procedure use following formula:
(4) the iteration formula of connection weight, hidden layer neuron central value and width is as follows:
Wherein (t-1), t, (t+1) indicate three states of iteration;CjiIndicate hidden layer neuron central value;djiIt indicates
Width.α indicates step pitch, and β indicates that Studying factors, E (t) are assessing network function.
D. the part of algorithm is optimized using multidimensional compensation method, realization has been come with least hidden layer neuron number
Result is differentiated at most accurate.
(1) neuron of basic RBF neural hidden layer described in step C calculates one by oneAnd draw
Enter threshold tau, incorporates the sample that all distances are greater than threshold tau into compensation training collection, the size of τ is adjusted according to live atlas, is made
Obtaining compensation training and concentrating sample size is the 30% of the training set of basic RBF neural.
(2) the tolerance fuzzy clustering method in recycle step A clusters compensation training collection;
(3) by the basic RBF neural in new cluster result sample set input step C, its radial base letter is set
Number are as follows:
(4) new neural network is trained with the training set of basic RBF neural, after multidimensional compensation can be obtained
Result.
In conclusion completing a kind of runway Surface airworthiness differentiation side based on pattern-recognition of the present invention
Method, overall procedure are as shown in Figure 2.The present invention only needs to obtain runway monitoring image, and it is suitable can to complete runway Surface in real time
The differentiation of boat property has the advantages that distinguishing speed is fast, high-efficient and safety and stability without manually participating in other equipment.