RU2450356C2 - Method for automated robust classification radio signals according to structure-time parameters - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: method comprises steps where educational data are clustered based on a competition principle which enables to build a set of standard vectors of structure-time parameters of known classes of radio signals with automatic estimation of a value σ for each standard vector; using these data to build a probabilistic neural network; using the built probabilistic neural network to estimate probability density distribution of known classes of radio signals in the region of an unknown (classified) signal and the unknown signal is identified with a class having the highest probability density distribution in the region of the unknown signal, characterised by that before building the probabilistic neural network.

EFFECT: automation and increase in robustness of the process of classifying radio signals according to structure-time parameters.

Description

The invention relates to communication technology and can be used in radio monitoring systems for the automated classification of radio signals according to structural-temporal parameters.

The closest method of the same purpose to the claimed invention according to the maximum number of similar features is a method for classifying radio signals according to structural-temporal parameters using a device such as a probabilistic neural network [Callan R. Basic concepts of neural networks .: Transl. from English - M .: Williams, 2003. - 288 p.], Which is a parallel implementation of the Bayesian method. The essence of this method is that for each sample (vector of structural-temporal parameters) of the radio signal, a decision can be made based on the choice of the most probable class from which the sample could belong. Such a solution requires an estimate of the probability density function for each class of radio signals, which is calculated in accordance with the expression:

where x is a classified sample;

n _j is the number of samples of class j in the training set;

x _i is the i-th sample of class j;

σ is a parameter specifying the width of the potential functions (permissible variance) and determining their influence. The registered radio signal is identified with the class that has the most dense probability distribution in a given area of the attribute space.

The sequence of actions with the known method for classifying radio signals according to structural-temporal parameters is as follows.

1. The construction (training) of the classifier is carried out on the basis of a probabilistic neural network as a result of the following actions:

- measurement of structural and temporal parameters of radio signals of various classes, the formation of training data;

- building on the training data of a probabilistic neural network with the expert path chosen by the value of the width of the potential function σ.

2. The unknown signals are classified using the constructed classifier based on a probabilistic neural network as a result of the following actions:

- measurement of structural and temporal parameters of an unknown radio signal x;

- calculation of the probability density function for each known class of radio signals in the region of the radio signal x in accordance with expression (1);

- determination of the class of unknown signal x according to the principle: x∈j if g _j (x) = max.

The considered method for classifying radio signals according to structural-temporal parameters using a probabilistic neural network is adopted as a prototype.

The reasons that impede the achievement of the following technical result when using the prototype include the following disadvantages. In the known method, it is necessary to select the value of σ by expert means. The difficulty of choosing is that different classes of radio signals have different values of intra-class dispersion. The second reason is the low robustness of the known method.

The basis of the invention is the task of automating the choice of the width of the potential function for each class of radio signals and increasing the robustness of classification by constructing reference vectors of structural and temporal parameters of various classes of radio signals.

The problem is achieved in that, in contrast to the known method for classifying radio signals according to structural-temporal parameters using a device of the type probabilistic neural network, which consists in the fact that according to the training data, consisting of vectors of structural-temporal parameters of signals of various classes, a probabilistic neural networks with the expert-selected value of the width of the potential function σ, using the constructed probabilistic neural network, the density is estimated the probability distribution of the known classes of radio signals in the region of an unknown (classified) signal and the unknown signal is identified with the class having the most dense probability distribution in the region of the unknown signal, in the inventive method, before constructing a probabilistic neural network, the training data undergoes a clustering procedure based on the competition principle, which provides building a set of reference vectors of structural-temporal parameters of known classes of radio signals with automatic eskoy estimate for each reference vector values σ.

Thanks to the introduction of a set of essential distinguishing features into the known method, the claimed method allows to automate and increase the robustness of the process of classifying radio signals according to structural-temporal parameters.

The preliminary clustering procedure allows you to automatically generate a set of reference vectors of structural and temporal parameters of known classes of radio signals, which are the average value of the parameter vectors of known classes of radio signals. The probabilistic neural network constructed from the reference vectors is robust in solving the classification problem due to the fact that the generated reference vectors are the optimal estimate of the true structural and temporal parameters of radio signals of various classes.

In addition, the dispersion values of the formed clusters calculated from the results of clustering are used as the values of the width of the potential function when constructing a probabilistic neural network, which allows us to automate the classification process, since it is not necessary to select the values of the width of the potential function by experts.

A method of automated robust classification of radio signals according to structural-temporal parameters consists in sequentially performing the following actions.

1. The construction (training) of the classifier is carried out on the basis of a probabilistic neural network, which consists in the sequence of the following actions:

1.1 Structural and temporal parameters of radio signals are measured for the formation of training data {X} _i , i = 1 ... N, where N is the number of training samples.

1.2 A set of clustering vectors of structural-temporal parameters of radio signals {X} is being clustered. Clustering is performed as follows:

1.2.1 A set of reference vectors {Y} _i , i = 1 ... N, initialized by random values is generated.

1.2.2 Randomly select one of the vectors of the training set {X} _i .

1.2.3 The winner vector {Y} _{c is} determined, which is most similar to the input vector. By similarity is meant the distance between vectors in Euclidean space.

1.2.4. The correction of the set of reference vectors {Y} _i is performed according to the formula:

Y _i (t + 1) = Y _i (t) + h _ci (t) (X (t) -Y _i (t)),

where t is the era number. The function h (t) is called the neighborhood function, which is divided into a distance function and a learning speed function and is defined as:

h (t) = h (|| r _c -r _i ||, t) · α (t),

- функция расстояния, θ{t} - радиус обучения, d=||r_c-r_i|| - расстояние между вектором-победителем и i-тым вектором,

- функция скорости обучения, А и В - константы скорости. На первоначальном этапе выбираются достаточно большие значения скорости обучения А, В и радиуса обучения θ{t), что позволяет расположить векторы {Y} в соответствии с распределением примеров в учебных данных, а затем производится точная подстройка векторов {Y}, когда значения параметров скорости обучения много меньше начальных.where r determines the position of the vector in the sample,

is the distance function, θ {t} is the learning radius, d = || r _c -r _i || is the distance between the winning vector and the i-th vector,

is a function of the learning rate, A and B are the rate constants. At the initial stage, sufficiently large values of the learning speed A, B and the learning radius θ {t) are selected, which allows one to arrange the vectors {Y} in accordance with the distribution of examples in the training data, and then fine-tuning the vectors {Y} when the values of the speed parameters are performed learning is much less than the initial.

1.2.5. Returns to step 1.2.2 until t reaches the specified value T. For the successful implementation of the method T> 1000.

Based on the results of clustering, a set of reference vectors of the structural-temporal parameters of radio signals {Y}, j = 1 ... M, is formed, where M is the number of received reference samples, with M <N, and the corresponding variances σ _j .

1.3 A probabilistic neural network is constructed from the reference data {Y} with a corresponding σ _j vector (sample).

2. Classification of unknown signals is carried out using the constructed classifier based on a probabilistic neural network, which consists in the sequence of the following actions:

2.1. Measurement of the structural-temporal parameters of the unknown radio signal x.

2.2 The probability density function is calculated for each known class of radio signals in the region of the radio signal x in accordance with expression (1).

2.3 The unknown signal class x is determined by the principle: x∈j if g _j (x) = max.

The advantage of the invention is that, compared with the prototype, the claimed method allows to solve the problem of classifying radio signals according to structural-temporal parameters completely automatically and reduces the likelihood of making a false decision (classification) due to random fluctuations of radio signals by 10-20%, i.e. increase robustness classification.

Claims (1)

A method of automated robust classification of radio signals according to structural-temporal parameters, which consists in constructing a probabilistic neural network using the constructed value of the potential function σ using the constructed probabilistic method using the training data consisting of vectors of the structural-temporal parameters of signals of various classes neural network, the probability density distribution of known classes of radio signals in the unknown is estimated (we classify of the signal) and the unknown signal is identified with the class having the most dense probability distribution in the region of the unknown signal, characterized in that before the construction of the probabilistic neural network, the training data is clustered based on the principle of competition, which provides the construction of a set of reference vectors of structural-temporal parameters known classes of radio signals with automatic estimation for each reference vector of the value of σ.