CN110187321A - Extraction method of radar radiation source feature parameters in complex environment based on deep learning - Google Patents

Abstract

The invention discloses a kind of Radar emitter characteristic parameter extraction methods under complex environment based on deep learning, belong to electronic reconnaissance field.It include: that initial characteristics extract, Classification Neural building, sparse self-encoding encoder constructs and eigenmatrix splicing；Classification Neural identification is wanted the method combined with sparse self-encoding encoder neural network recognization by the present invention, analyse in depth and study the essence of emitter Signals, explore new characteristic parameter, building is more conducive to the feature vector of signal identification, improves the recognition capability of radar emitter signal under complex environment.

Description

Extraction method of radar radiation source feature parameters in complex environment based on deep learning

technical field

The invention belongs to the field of electronic reconnaissance, in particular to a method for extracting characteristic parameters of radar radiation sources in a complex environment based on deep learning.

Background technique

Target recognition is a key link in the field of electronic reconnaissance, in which the main task is to extract characteristic parameters of radiation source signals. The theoretical achievements in this area include: instantaneous autocorrelation method, wavelet transform method, fuzzy function ridge feature method, wavelet packet and entropy feature method, etc. They have studied radar target recognition from different angles and levels.

At present, the problems existing in the extraction of characteristic parameters of radar radiation sources are as follows:

Existing methods are mainly aimed at specific signals, and in complex environments where multiple signals are compounded, the current algorithms are not enough to meet actual needs.

For the existing feature parameters and identification methods, they are more effective in the case of simple electromagnetic environment, but the identification effect is not good under the condition of low signal-to-noise ratio (such as: SNR≤2dB), and can not meet the requirements of various intra-pulse modulation signals. Identification requirements in case of simultaneous presence.

The purpose of identification is to know the type of weapon that emits the signal and to judge its threat level. However, the identification of radiation sources is less considered at present.

From the above analysis, it can be seen that it is of great significance to deeply study the essence of the radiation source signal and explore the eigenvectors that are more suitable for signal identification.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for extracting characteristic parameters of radar radiation source in complex environment based on deep learning, which solves the problem that the extraction effect of characteristic parameters of radar radiation source in complex environment is not good and cannot meet the identification requirements.

The technical scheme adopted in the present invention is as follows:

A method for extracting characteristic parameters of radar radiation sources in a complex environment based on deep learning, comprising:

Initial feature extraction: extract the parameter information of radiation source and loading platform as initial features;

Classification neural network construction: input initial features, construct the upper-layer classification neural network structure of "initial features-neural network intermediate layer A-radiation source and loading platform category", and map initial features and radiation source and loading platform categories through the output of neural network intermediate layer A The feature matrix A of the relationship;

Sparse auto-encoder network construction: The initial features are used as input and output at the same time, and the lower-layer sparse auto-encoder network structure of "initial feature-encoder-neural network intermediate layer B-decoder" is constructed, and the initial features are output through the neural network intermediate layer B. Deeply refined intrinsic attribute feature matrix B;

Feature matrix splicing: splicing the relational feature matrix A that reflects the categories of initial features, radiation sources and loading platforms, and the feature matrix B that reflects the intrinsic properties of the initial features themselves, to obtain the final complex environmental feature parameters.

Further, the parameters of the initial feature for the radiation source include the carrier frequency, pulse width, angle of arrival, pulse repetition frequency, and antenna scanning period of the radar, combined with the pulse arrival time of signals such as communication and interference, pulse envelope parameters, pulse Modulation parameters, amplitude, spectrum parameters;

The parameters of the initial feature for the loading platform include the moving speed of the loading platform and the spatial position parameters.

Further, the information in the classification neural network propagates in one direction, and the training method of the middle layer A of the neural network adopts a supervised learning method.

Further, in the sparse self-encoder network, the encoder is used to perform dimensionality reduction processing on the initial features, and the kernel information of the initial features is refined; The features have the same amount of information, and the output error is fed back to the initial features, so as to train the intermediate layer B of the neural network, and output the feature matrix B after the initial features are deeply refined.

To sum up, due to the adoption of the above-mentioned technical solutions, the beneficial effects of the present invention are:

1. the present invention adopts the method that the classification neural network and the sparse self-encoding neural network are combined, the classification neural network obtains the characteristic matrix A that maps the relationship between the initial characteristic and the radiation source, the loading platform category according to the initial characteristic input, and the sparse self-encoding neural network according to The initial feature input is obtained to obtain the intrinsic attribute feature matrix B after the initial features are deeply refined, and the final feature parameters are obtained by splicing the two feature matrices. The feature vector that helps in signal identification improves the identification ability of radar radiation source signals in complex environments.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present invention more clearly, the following briefly introduces the accompanying drawings used in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore do not It should be regarded as a limitation of the scope. For those of ordinary skill in the art, other related drawings can also be obtained from these drawings without creative efforts, wherein:

Fig. 1 is the characteristic parameter extraction flow chart of the present invention;

2 is a schematic diagram of a sparse autoencoder network of the present invention;

FIG. 3 is a schematic diagram of the change of the relative entropy value.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention, that is, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the invention generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

Thus, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but is merely representative of selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present invention.

It should be noted that relational terms such as the terms "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any relationship between these entities or operations. any such actual relationship or sequence exists. Moreover, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion such that a process, method, article or device comprising a list of elements includes not only those elements, but also includes not explicitly listed or other elements inherent to such a process, method, article or apparatus. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in a process, method, article or apparatus that includes the element.

A method for extracting characteristic parameters of radar radiation sources in a complex environment based on deep learning, comprising:

Initial feature extraction: extract the parameter information of radiation source and loading platform as initial features;

Classification neural network construction: input initial features, construct the upper-layer classification neural network structure of "initial features-neural network intermediate layer A-radiation source and loading platform category", and map initial features and radiation source and loading platform categories through the output of neural network intermediate layer A The feature matrix A of the relationship;

Sparse auto-encoder network construction: The initial features are used as input and output at the same time, and the lower-layer sparse auto-encoder network structure of "initial feature-encoder-neural network intermediate layer B-decoder" is constructed, and the initial features are output through the neural network intermediate layer B. Deeply refined intrinsic attribute feature matrix B;

Feature matrix splicing: splicing the relational feature matrix A that reflects the categories of initial features, radiation sources and loading platforms, and the feature matrix B that reflects the intrinsic properties of the initial features themselves, to obtain the final complex environmental feature parameters.

Specifically, the feature parameter extraction process is shown in Figure 1:

Step 1: First extract the initial feature information: for the radiation source, consider the radar's carrier frequency, pulse width, angle of arrival, pulse repetition frequency, antenna scanning period and other parameters, combined with the pulse arrival time and pulse envelope parameters of signals such as communication and interference , pulse modulation parameters, amplitude, spectrum and other measurement parameters to form the initial characteristic parameters; for the loading platform, its moving speed, spatial position and other characteristics are directly selected as the initial characteristic parameters;

The initial characteristic parameter information of the radiation source and the state platform is synthesized, and the initial characteristic input quantity is obtained by synthesis.

Step 2. Use the initial features as the input to construct the upper classification neural network structure of "Initial Features - Neural Network Intermediate Layer A - Radiation Source and Loading Platform Category". The information in the upper classification neural network is propagated in one direction, not reversed The intermediate layer A of the neural network is trained in a supervised learning manner, and finally the intermediate layer A of the neural network outputs a feature matrix A that maps the relationship between the initial feature, the radiation source, and the loading platform category;

Specifically, each neuron in the classification neural network is divided into different groups according to the order of receiving information, each group is regarded as a neural layer, and the neurons in each neural layer receive the output signals of the neurons in the previous neural layer. , and continue to output the signal to the neuron in the next neural layer; each neural layer, as a high-dimensional representation of the data information in the input signal, can be regarded as a nonlinear function. complex mapping of input signals to output signals;

The neurons in the classification neural network not only receive the signals output by other neurons, but also receive their own feedback signals according to the comparison between the radiation source and the loading platform, and modify the parameters of the middle layer A of the neural network through the feedback signals to make the middle layer better. Extract feature parameters.

Step 3. Use the initial features as both input and output to construct the lower-layer sparse auto-encoder network structure of "initial feature-encoder-neural network middle layer B-decoder". Perform dimensionality reduction processing to extract the kernel information of the initial features; the decoder trains the encoder to determine whether the information extracted by the encoder is accurate, whether it obtains features with the same amount of information as the initial features, and feeds the output errors back to the initial features, so as to The intermediate layer B of the neural network is trained, and the feature matrix B after the initial features are deeply refined is output.

Specifically, the sparse autoencoder is an unsupervised machine learning algorithm that continuously adjusts the parameters of the autoencoder by calculating the error between the output of the autoencoder and the original input, and finally trains the model; the autoencoder can be used to compress the input information. , extract useful input features.

The autoencoder is divided into an encoder and a decoder. The encoder converts d-dimensional features into p-dimensional features, and the decoder reconstructs p-dimensional features back to d-dimensional features. When p<d is satisfied, the autoencoder is used as a dimension reduction feature Extraction, coupled with constraints such as encoding sparsity and range of values, yields meaningful output.

As shown in Figure 2: Use {x1, x2, x3, ...} to represent the input unlabeled data, that is, the initial feature input, and the goal is to output h _{W, b} (x) ≈ x, that is, sparse self-encoding is trying to approximate an identity function such that the output close to the input x;

When the number of neurons in the middle layer B of the neural network is less than the input amount, the auto-encoder neural network is forced to learn a "compressed" representation of the input data; when the number of neurons in the middle layer B of the neural network is large, the auto-encoder neural network is The network imposes sparsity constraints to achieve the effect of compressing input information and extracting input features.

It is understandable that the so-called sparsity can be explained as: when the output of a neuron is close to 1, it is considered to be activated, and when the output is close to 0, it is considered to be inhibited, so that the neuron is inhibited most of the time. The limit is called the sparsity limit.

Let b _j denote the activation degree of neuron j in the middle layer B of the neural network, and b _j [x(i)] denote the activation degree of neuron j under the given input x(i), then the activation degree of neuron j average activation It is expressed as follows:

In the formula, m represents the number of input x.

Then the sparsity parameter ρ of the neural network constraints:

where ρ is usually a small value close to 0.

Further, in order to achieve the sparsity limit, one needs to limit Keeping the value of ρ in a small range, set the penalty factor as follows:

In the formula, S ₂ represents the number of neurons in the middle layer B of the neural network;

Based on relative entropy, the penalty factor is expressed as:

in, is a mean with ρ and a is the relative entropy between two Bernoulli random variables with mean;

when hour, and with The difference from ρ increases monotonically.

Set the sparsity parameter ρ=0.2, along with The change of is shown in Figure 3; it can be seen that the relative entropy is in reaches its minimum value of 0 when When it is close to 0 or 1, the relative entropy becomes very large and tends to ∞, so minimizing the penalty factor can make close to ρ.

Thus, the overall cost function of the sparse autoencoder neural network is obtained as:

In the formula, β represents the weight that controls the sparsity penalty factor, and this algorithm can greatly reduce the initial data dimension.

The autoencoder passes through the encoder to the intermediate layer B of the neural network and then to the decoder, and the layers are fully connected to each other. By minimizing the reconstruction error, the network parameters can be effectively learned to obtain the desired feature parameters.

Step 4: splicing the feature matrix A reflecting the relationship between the initial features, the radiation source and the loading platform, and the feature matrix B reflecting the intrinsic properties of the initial feature itself, to obtain the final complex environment feature parameters.

The present invention combines classification neural network recognition and sparse autoencoder neural network recognition, deeply analyzes and studies the essence of radiation source signals, explores new feature parameters, constructs feature vectors that are more helpful for signal recognition, and improves the complexity of The ability to identify radar radiation source signals in the environment.

The above are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements and improvements made by any person skilled in the art within the spirit and principles of the present invention, etc. , should be included within the protection scope of the present invention.

Claims (4)

1. A radar radiation source characteristic parameter extraction method under a complex environment based on deep learning is characterized by comprising the following steps:

extracting initial features: extracting parameter information of a radiation source and a loading platform as initial characteristics;

constructing a classification neural network: inputting initial characteristics, constructing an upper-layer classification neural network structure of 'initial characteristics-neural network intermediate layer A-radiation source and loading platform category', and outputting a characteristic matrix A for mapping the relationship between the initial characteristics and the radiation source and the loading platform category through the neural network intermediate layer A;

constructing a sparse self-encoder network: the initial features are simultaneously used as input and output quantities, a lower-layer sparse self-encoder network structure of 'initial feature-encoder-neural network intermediate layer B-decoder' is constructed, and an internal attribute feature matrix B of which the initial features are deeply refined is output through the neural network intermediate layer B;

splicing the feature matrixes: and splicing a relation characteristic matrix A reflecting the initial characteristics, the radiation source and the loading platform types with a characteristic matrix B reflecting the inherent attributes of the initial characteristics to obtain the final complex environment characteristic parameters.

2. The method for extracting the characteristic parameters of the radar radiation source in the complex environment based on the deep learning as claimed in claim 1, wherein: the parameters of the initial characteristic for the radiation source comprise carrier frequency, pulse width, arrival angle, pulse repetition frequency and antenna scanning period of the radar, and pulse arrival time, pulse envelope parameter, intra-pulse modulation parameter, amplitude and spectrum parameter of signals such as communication and interference are combined;

the parameters of the initial characteristics for the loading platform comprise the moving speed and the spatial position of the loading platform.

3. The method for extracting the characteristic parameters of the radar radiation source in the complex environment based on the deep learning as claimed in claim 1, wherein: the information in the classified neural network is transmitted towards one direction, and the training mode of the neural network intermediate layer A adopts a supervised learning mode.

4. The method for extracting the characteristic parameters of the radar radiation source in the complex environment based on the deep learning as claimed in claim 1, wherein: in the sparse self-encoder network, an encoder is used for carrying out dimensionality reduction on the initial characteristic and refining kernel information of the initial characteristic; the decoder is used for training the encoder, judging whether the information extracted by the encoder is accurate or not, whether the characteristic with the same information quantity as the initial characteristic is obtained or not, and feeding back the output error to the initial characteristic so as to train the intermediate layer B of the neural network and output a characteristic matrix B of which the initial characteristic is deeply extracted.