Disclosure of Invention
In order to solve the problems, the invention provides a method for judging the airworthiness of the runway surface of the airport based on pattern recognition.
The technical scheme adopted by the invention is as follows:
a method for distinguishing airworthiness of an airport runway pavement based on pattern recognition is characterized by comprising the following steps: the method comprises the following steps: A. clustering and preprocessing the data samples by adopting a tolerance fuzzy clustering algorithm; B. optimizing the number of hidden layer neurons in the RBF neural network to obtain optimal real-time performance; C. training the RBF neural network; D. and optimizing the local part of the algorithm by adopting a multi-dimensional compensation method to realize the optimal selection of the most accurate discrimination result by using the least number of the hidden layer neurons, wherein the specific implementation mode of the step A is as follows:
acquiring a trace image near an airport runway, taking the detected trace image as a sample set, and assuming that the sample set consisting of n D-dimensional samples is Dset={IM1,IM2,…,IMn,IMi∈RdIn which IMiA trace image sample; in the process of judging the navigability, dividing the clustering result into C types; the membership of the sample in the tolerance fuzzy clustering is recorded as gijRepresents a sample IMiMembership degree of membership j category; introducing a tolerance theta such that the samples are in the parameter set PcRelative to the central value McVariance of (2) is D (IM)i,McPc)=(IMi-Mc)T(IMi-Mc)-θ;
After introducing the tolerance, the variance of the sample is greater than or equal to 0 and less than 0, and two sets are definedNAnd SetpThe sample similarity in this case is represented respectively:
therefore, the class set Min with the least sample similarityi={c,D(IMi,MjPj)≥D(IMi,McPc) For different fuzzy coefficients a, the following membership formula is given:
based on the above, tolerance fuzzy clustering processing can be carried out on the sample set, and the steps are as follows:
(1) initializing a set of central values
(2) Circularly processing the samples in the training set and calculating the Euclidean distance Dist of the samples
eTo find the minimum central value of the corresponding training set sample
The following conditions are satisfied:
(1. ltoreq. i. ltoreq.C), where t represents the number of iterations;
(3) and (3) updating the central value of each dimension simultaneously in each circulation process:
where ρ istIs a system function with a decreasing nature, let
Wherein 0 < rho0If less than 1, random numbers can be taken; t is the total number of training times; t is the current iteration step number;
in order to ensure that the network has an artificial intelligence interface in use, the tolerance theta is introduced by adopting a tolerance function theta (t), the determination of the tolerance function can be empirically specified, and can also be fitted according to user habits, and in the initialization of the system, the tolerance function is ordered
N is the total number of samples, δ
tFor the system control function, a gradient descent method is adopted to solve:
(4) according to class membership
Generating a tolerance fuzzy cluster of samples:
bringing the formula (c) into the formula (c), and obtaining the membership degree of the sample according to the formula:
preferably, the specific implementation manner of step B is:
in the iteration process, after the set iteration step number is reached, calculating clustering results of various types, and if the number of clustering samples of a certain type is less than a threshold value epsilon, fitting the classification into a neighborhood class: assuming that the set of central values generated after the iteration is finished is M ═ M
1,M
2,…,M
dAfter the first clustering treatment, the type Set with the sample number smaller than the threshold value epsilon is Set
ε={ST
1,ST
2,…,ST
k}(1≤k≤C),
Deleting the type set below the threshold value epsilon, taking out the samples below the threshold value epsilon, and calculating the Euclidean distance between the type set and the central value in the type set:
and after the minimum Euclidean distance is calculated, the type set index where the central value is located can be obtained, and finally, the sample set is fitted to the neighborhood type set, so that the neighborhood fitting of the neuron is completed.
Preferably, the specific implementation manner of step C is:
the parameters of the neural network are set as follows:
(1) the input training sample set is: training set after tolerance clustering:
(2) desired output vector
(3) Initializing the connection weight of the hidden layer and the output layer:
wherein, p and q are the neuron numbers of an output layer and an implied layer respectively;
the initialization process of (a) adopts the following equation:
(4) the iterative formula of the connection weight, the hidden layer neuron center value and the width is as follows:
where (t-1), t, (t +1) represent the three states of the iteration, CjiRepresenting hidden layer neuron center values, djiRepresenting the width, alpha representing the step distance, beta representing the learning factor, and E (t) being the network evaluation function.
Preferably, the specific implementation manner of step D is:
(1) calculating the neurons of the hidden layer of the basic RBF neural network in the step C one by one
Introducing a threshold value tau, classifying all samples with the distance larger than the threshold value tau into a compensation training set, and adjusting the size of the tau according to a field atlas to make the compensationThe number of samples in the compensation training set is 30% of the training set of the basic RBF neural network;
(2) clustering the compensation training set by using the tolerance fuzzy clustering method in the step A;
(3) inputting the new clustering result sample set into the basic RBF neural network in the step C, and setting the radial basis function as follows:
(4) and (4) training the new neural network by using the training set of the basic RBF neural network to obtain a result after multi-dimensional compensation.
The invention has the beneficial effects that:
according to the method, the airworthiness of the runway surface of the airport runway can be judged in real time only by utilizing a large number of images of rubber traces or rain and snow covering traces of the runway grounding belt generated by runway monitoring and through neural network learning and pattern recognition, manual work and other equipment are not needed, and the method has the advantages of high judging speed, high efficiency, safety and stability.
Detailed Description
As shown in fig. 2, the method for judging the airworthiness of the runway surface of the airport based on pattern recognition comprises the following steps:
A. clustering and preprocessing the data samples by adopting a tolerance fuzzy clustering algorithm:
the method mainly collects trace images near a runway bidirectional grounding belt and a bidirectional takeoff waiting point, and adopts the RBF neural network as a main body to realize concise linear learning and achieve the high-precision effect of nonlinear learning in order to achieve the purposes of simple structure, concise training and guarantee of high precision of the algorithm. On the other hand, the clustering performance of the data samples and the number of neurons in the hidden layer have decisive influence on the performance of the RBF neural network, while the traditional clustering algorithm is too rigorous in rule, and the clustering processing effect on the pixel sample set of the trace image is not ideal, so that the tolerance fuzzy clustering algorithm provided by the invention has very ideal effect, and the algorithm is as follows:
a large number of detected trace images are used as a sample set, and the sample set consisting of n D-dimensional samples is assumed to be Dset={IM1,IM2,…,IMn,IMi∈RdIn which IMiA trace image sample; in the process of judging the navigability, the judging result can be divided into two or three categories, and for convenient expression, the clustering result is divided into a category C; the membership of the sample in the tolerance fuzzy clustering is recorded as gijRepresents a sample IMiMembership degree of membership j category; introducing a tolerance theta such that the samples are in the parameter set PcRelative to the central value McVariance of (2) is D (IM)i,McPc)=(IMi-Mc)T(IMi-Mc)-θ。
After introducing the tolerance, the variance of the sample is greater than or equal to 0 and less than 0, and two sets are definedNAnd SetpThe sample similarity in this case is represented respectively:
therefore, the class set Min with the least sample similarityi={c,D(IMi,MjPj)≥D(IMi,McPc) For different fuzzy coefficients a, the following membership formula is given:
based on the above, tolerance fuzzy clustering processing can be carried out on the sample set, and the steps are as follows:
(1) initializing a set of central values
(2) Circularly processing the samples in the training set and calculating the Euclidean distance Dist of the samples
eTo find the minimum central value of the corresponding training set sample
The following conditions are satisfied:
(i is more than or equal to 1 and less than or equal to C); where t represents the number of iterations.
(3) And (3) updating the central value of each dimension simultaneously in each circulation process, wherein the method comprises the following steps:
where ρ istIs a system function of decreasing nature, in this example, let
Wherein 0 < rho0If less than 1, random numbers can be taken; t is the total number of training times; and t is the current iteration step number.
In order to ensure that the network has an artificial intelligence interface in use, the tolerance theta is introduced by adopting a tolerance function theta (t), and the determination of the tolerance function can be empirically specified or can be fitted according to user habits. In the initialization of the system, let
And N is the total number of samples.
δtFor the system control function, a gradient descent method is adopted to solve:
(4) according to class membership
And (3) generating tolerance fuzzy clustering of the samples, wherein the algorithm is as follows:
bringing the formula (c) into the formula (c), and obtaining the membership degree of the sample according to the formula:
B. and optimizing the number of hidden layer neurons in the RBF neural network to obtain optimal real-time performance:
the number of neurons in the hidden layer affects the real-time performance of the training algorithm, and excessive number of neurons does not bring higher accuracy, but seriously affects the timeliness, so that the feedback mechanism is adopted to optimize the number of neurons in the hidden layer, and the optimal real-time performance is obtained on the premise of ensuring certain accuracy, and the method comprises the following steps:
in the iteration process, after the set iteration step number is reached, the clustering results of all types are calculated, and if the number of the clustering samples of a certain type is smaller than the threshold value epsilon, the classification is fitted into the neighborhood.
Assuming that the set of central values generated after the iteration is finished is M ═ M
1,M
2,…,M
dAfter the first clustering treatment, the type Set with the sample number smaller than the threshold value epsilon is Set
ε={ST
1,ST
2,…,ST
k}(1≤k≤C),
Deleting the type set below the threshold value epsilon, taking out the samples below the threshold value epsilon, and calculating the Euclidean distance between the type set and the central value in the type set:
and after the minimum Euclidean distance is calculated, the type set index where the central value is located can be obtained, and finally, the sample set is fitted to the neighborhood type set, so that the neighborhood fitting of the neuron is completed.
C. Training the RBF neural network:
after the tolerance fuzzy clustering algorithm is completed and the data samples are clustered and preprocessed, the basic structure of the neural network is shown in fig. 2.
The parameters of the neural network are set as follows:
(1) the input training sample set is: training set after tolerance clustering:
(2) desired output vector
(3) Initializing the connection weight of the hidden layer and the output layer:
wherein, p and q are the numbers of neurons in the output layer and the hidden layer respectively.
The initialization process of (a) adopts the following equation:
(4) the iterative formula of the connection weight, the hidden layer neuron center value and the width is as follows:
where (t-1), t, (t +1) represent the three states of the iteration; cjiRepresenting hidden layer neuron center values; djiIndicating the width. Alpha represents the step distance, beta represents the learning factor, and E (t) is the network evaluation function.
D. And optimizing the local part of the algorithm by adopting a multi-dimensional compensation method, and realizing the most accurate judgment result by using the least number of the hidden layer neurons.
(1) Calculating the neurons of the hidden layer of the basic RBF neural network in the step C one by one
And introducing a threshold tau, classifying all samples with the distance greater than the threshold tau into a compensation training set, and adjusting the size of the tau according to a field atlas to ensure that the number of the samples in the compensation training set is 30% of that of the training set of the basic RBF neural network.
(2) Clustering the compensation training set by using the tolerance fuzzy clustering method in the step A;
(3) inputting the new clustering result sample set into the basic RBF neural network in the step C, and setting the radial basis function as follows:
(4) and (4) training the new neural network by using the training set of the basic RBF neural network to obtain a result after multi-dimensional compensation.
In summary, the method for judging airworthiness of the runway surface based on pattern recognition is completed, and the general flow is shown in fig. 2. The method can complete the discrimination of the airworthiness of the runway surface of the airport in real time only by acquiring the runway monitoring image without manual work and other equipment, and has the advantages of high discrimination speed, high efficiency, safety and stability.