CN110109063A - A kind of radiation source repetition modulation identification method based on deepness belief network - Google Patents

Abstract

The invention discloses a method for identifying the type of re-frequency modulation of a radiation source based on a deep belief network. The specific steps are as follows: according to the typical five kinds of radiation source signals, the representative features are extracted to form eigenvectors; in order to make the re-frequency modulation type identification network converge quickly, the eigenvectors are normalized; in order to meet the deep neural network To meet the needs of network output, one-hot coding is used to digitize the features of five output categories; the feature vector is used as input, and the binary vector of five output categories is used as output to establish a radar repetition frequency modulation based on deep belief network. way to identify the model. The invention can accurately identify the repetitive frequency modulation type of the radar pulse signal, and has better anti-false ability and anti-missing ability. Compared with other classifiers, the invention has the same recognition effect for several typical repetitive frequency signals a certain degree of improvement.

Description

A deep belief network-based identification method for radio frequency modulation type

Technical field:

The invention relates to the field of radar signal identification, in particular to the identification of the repetition frequency modulation type of radar signals under complex electromagnetic environment.

Background technique:

Due to the rapid development of radar technology and the wide application of radar equipment with complex systems, current radars usually have multiple functions and can meet diverse mission requirements. As a result, a radar has multiple working modes, and under different working modes , radars have obvious differences in signal parameters and signal styles. That is to say, the radar working mode is a working state in which the radar has a relatively stable working period and working characteristics in order to complete a specific task. In addition, due to the influence of radar technology system and noise, even if the radar is in the same working mode, the variation law of its signal parameters is not obvious. At the same time, due to the increasingly complex electronic environment and the enhancement of jamming technology and anti-reconnaissance technology, it is difficult to identify and process radar signals of complex systems. How to accurately identify the typical type of re-frequency modulation of a radiation source in a complex electromagnetic environment is a problem that needs to be solved at present.

Invention content:

The purpose of the present invention is to propose a method for identifying the type of re-frequency modulation of a radiation source based on a deep belief network.

Specific steps are as follows:

Step 1: According to the typical five types of radiation source signals, extract representative features to form feature vectors;

Step 2: In order to make the re-frequency modulation type identification network converge quickly, the eigenvectors are normalized;

Step 3: In order to meet the needs of the output of the deep neural network, one-hot encoding is used to digitize the features of the five output categories;

Step 4: Taking the feature vector as input and the binary vector of five output categories as output, establish a radar re-frequency modulation recognition model based on deep belief network.

The specific steps for extracting features in the first step are as follows:

Feature 1: The repetition rate feature of the PRI sequence is:

Among them, N represents the number of intervals, M represents the number of occurrences of data corresponding to each interval, and M _max is the highest number of occurrences in a single interval.

Feature 2: The small window repetition rate feature of the PRI sequence is:

Among them, N1 represents the number of windowed intervals, M1 represents the number of occurrences of data corresponding to each interval, and M1 _max is the highest number of occurrences in a single interval.

Feature 3: The small window fitting slope feature of the PRI sequence is:

k=p ₁ x ⁿ +p ₂ x ^n-1 +...+p _n x+p _n+1

Feature 4: The small window deviation rate feature of the PRI sequence is:

where km _max is the maximum time interval value, km _min is the minimum time interval value, K _i is the value of each interval in the sequence, n is the sequence length, and win is the windowing scale.

The above four features make up the feature vector:

x=(per, winper, k, b}

In the second step, the feature vector is normalized, and the specific steps are as follows:

Among them, x represents the input vector set, x={per, winper, k, b}, x _min is the minimum value of the feature parameter, and x _max is the maximum value of the feature parameter.

The specific steps of performing one-hot encoding on the output category in the third step are as follows:

Using one-hot encoding, first map the five types to integer values. Then, represent each integer value as a binary vector. After encoding, the repetition frequency fixed type can be expressed as "00001", the repetition frequency staggered type can be expressed as "00010", the repetition frequency pulse group type can be expressed as "00100", the repetition frequency jitter type can be expressed as "01000", and the repetition frequency can be expressed as "01000". The slip type can be represented as "10000".

In the step 4, the specific steps of establishing a radar repetition frequency modulation identification model based on a deep belief network are as follows:

The deep classifier takes as input the four features of the impulsive signal and outputs a binary vector of five modulation types. The hidden layer adopts a single-layer RBM to extract features in an unsupervised manner, and the output layer adopts a feedforward neural network to classify in a supervised manner.

Beneficial effects:

The invention proposes a new method for identifying the type of pulse repetition frequency modulation of radiation source, which has the advantages of: the proposed features are representative; the selected neural network combines two learning methods of supervised learning and unsupervised learning, that is, training is saved Time also improves the recognition accuracy; the generated deep belief network-based RFM type identification model can accurately identify the type of radio frequency modulation in complex electromagnetic environments.

Description of drawings:

Fig. 1 is the flow chart of the present invention;

Figure 2 shows the comparison results of model recognition accuracy under different missing rates;

Figure 3 shows the comparison results of model recognition accuracy under different false rates.

Specific implementation plan:

The present invention will be described in detail below through specific implementation examples.

(1) Feature extraction is performed on the PRI pulse sequence. The four features are as follows:

Feature 1: Repeat Rate Feature:

For a fixed PRI signal, within a certain tolerance range, the frequency of certain intervals must be much higher than other intervals, and this variation characteristic is called the repetition rate. The current pulse sequence is taken out, and the time intervals in the sequence are counted according to the histogram distribution to obtain the repetition rate. Ideally, the repetition rate of the repetition frequency fixed sequence is 100%. When the measurement error is considered, the repetition rate is approximately 100%; the repetition rate of other repetition frequency modulation sequences such as repetition frequency staggering fluctuates, and the distance from 100% is far away. , so the repetition rate fixed pattern can be effectively identified by the repetition rate feature.

Among them, N represents the number of intervals, M represents the number of occurrences of data corresponding to each interval, and M _max is the highest number of occurrences in a single interval.

Feature 2: Small window repetition rate feature:

For group-variable signals, under a certain windowing scale, within a certain tolerance range, the frequency of occurrence must be higher than other intervals, and this change characteristic is called the small window repetition rate. The current pulse is opened and intercepted, the time interval in the small window is counted according to the histogram distribution, and the repetition rate of the small window is obtained. Ideally, the repetition rate of the small window of the repeated frequency pulse group sequence will exceed 50%, and the repetition rate characteristic will be higher than a certain fixed threshold, that is, not close to zero. When considering the measurement error, the small window repetition rate feature and the repetition rate feature still have such characteristics. Therefore, the pattern of repetitive pulse groups can be effectively identified through the squeeze between the repetition rate feature of the small window and the repetition rate feature.

For the repetition frequency staggered signal, its complete sub-cycle will appear repeatedly. Therefore, the repetition rate characteristics of each sub-cycle of the repetition frequency staggered signal are the same. After removing the repetition frequency fixed and repetition frequency pulse group signals, this feature can be used to distinguish The repeated frequency staggered signal is output.

Among them, N1 represents the number of windowed intervals, M1 represents the number of occurrences of data corresponding to each interval, and M1 _max is the highest number of occurrences in a single interval.

Feature 3: Small window fitting slope feature:

After identifying the fixed repetition frequency, repetition frequency pulse group and repetition frequency staggered signal, for the repetition frequency sliding signal, that is, the PRI monotonically increases or decreases to a limit value and then changes to another limit value, so in a short time Its variation law presents a monotonic variation mode, and the parameter of the small window fitting slope is to be selected.

Ideally, the first-order fitting slope of the re-frequency glide signal is significantly different from other types of signals. Even if the measurement error is considered, the range of its upper and lower jitter will not exceed the PRI value at this moment, so even if there is the influence of the measurement error. , does not have a significant impact on the final fit rate.

k=p ₁ x ⁿ +p ₂ x ^n-1 +...+p _n x+p _n+1

Feature 4: Small window deviation rate feature:

For the repetition frequency jitter signal, the PRI jitter range is generally 1% to 30%, and it exhibits random variation within this range. In fact, the change of the pulse interval has a certain regularity, that is, the variation range of a single repetition frequency jitter PRI will not more than 30% of the central value, so the deviation rate feature can be used to distinguish the jitter signal.

where km _max is the maximum time interval value, km _min is the minimum time interval value, K _i is the value of each interval in the sequence, n is the sequence length, and win is the windowing scale.

The above four features make up the feature vector:

x={per, winper, k, b}

(2) Normalize the four features:

In order to make the classifier work more convenient and faster, the four features of (1) are normalized, and the feature value becomes a decimal between (0,1).

Among them, x represents the input vector set, x={per, winper, k, b}, x _min is the minimum value of the feature parameter, and x _max is the maximum value of the feature parameter.

(3) One-hot encoding for five output categories:

Since the output is a signal class, one-hot encoding is used to solve such discrete value problems. The five types are first mapped to integer values; then, each integer value is represented as a binary vector. After encoding, the repetition frequency fixed type can be expressed as "00001", the repetition frequency staggered type can be expressed as "00010", the repetition frequency pulse group type can be expressed as "00100", the repetition frequency jitter type can be expressed as "01000", and the repetition frequency can be expressed as "01000". The slip type can be represented as "10000".

(4) Establish a recognition model of radar repetition frequency modulation based on deep belief network:

The model construction uses a deep belief network model, a single-layer RBM for the hidden layer, and a feedforward neural network for the output layer. The working process is as follows: The first step uses unsupervised training of the RBM network to ensure that the feature vectors are mapped to different feature spaces. retain feature information. The mini-batch data processing mode is used during training, and the convergence process is accelerated by the CD (Contrastive Divergence) algorithm. The relevant parameters of the trained RBM network will be used as the input for the supervised training of the output layer. The second step is fine-tuning, setting the BP network at the output layer of the DBN, receiving the output feature vector of the RBM as its input feature vector, and training the entity relation classifier supervised. The supervised adjustment work is mainly completed through the BP algorithm. The top-down adjustment of the entire BP network ensures that the error is controlled within the minimum range and reduces the possibility of the error being transmitted upwards, thereby ensuring that the feature vector mapping of the entire DBN is optimal. , the RBM network parameters after supervised training are used as the input of the BP layer. After the above two processes are completed, a training process of mixed supervision is completed. This mixed training and learning can improve the recognition effect of the system's re-frequency modulation type.

The deep classifier takes as input the four features of the impulsive signal and outputs a binary vector of five modulation types. The hidden layer adopts a single-layer RBM to extract features in an unsupervised manner, and the output layer adopts a feedforward neural network to classify in a supervised manner.

Example 1: Analysis of the stability experiment of the recognition model under the missing condition

Parameter setting in the experiment: when the receiver measurement error is 0.05% (δ=0.05%), in order to accurately test the anti-missing ability of the model, the false rate is set to 0 (η=0), and then the missing rate is set in turn 10%, 20%, 30%, 40%, namely ε=10%, ε=20%, ε=30%, ε=40%.

The classifier was trained 500 times, and the recognition accuracy of this method under four different missing rate conditions was obtained. The results are shown in Figure 2. When the missing rate is 10%, the model recognition accuracy is 99.2%; when the missing rate is 20%, the model recognition accuracy is 96.2%; when the missing rate is 30%, the model recognizes The accuracy rate is 93.2%; when the missing rate is 40%, the recognition accuracy of the model is 92.0%.

It can be seen from the observation that the recognition model of the re-frequency modulation method has a high recognition accuracy in the case of different missing rates. Although the increase of the missing rate will lead to an increase of invalid data values, the invalid values caused by the missing are likely to be multiples of the original value, which can be filtered by data cleaning. Therefore, with the increase of the missing rate, the recognition accuracy does not increase. A sharp drop occurred. Therefore, the recognition model has high stability under the condition of missing pulse interference.

Example 2: Analysis of the stability experiment of the recognition model under false conditions

Parameter setting in the experiment: when the receiver measurement error is 0.05% (δ=0.05%), in order to accurately test the anti-false ability of the model, the missing rate is set to 0 (∈=0), and then the false rate is set in turn It is 10%, 20%, 30%, 40%, namely η=10%, η=20%, η=30%, η=40%.

The classifier was trained for 500 times, and the recognition accuracy of the method in this paper under four different false rate conditions was obtained. The results are shown in Figure 3. When the false rate is 10%, the model recognition accuracy is 94.0%; when the false rate is 20%, the model recognition accuracy is 93.2%; when the false rate is 30%, the model recognition accuracy The accuracy rate is 92.4%; when the false rate is 40%, the recognition accuracy of the model is 87.8%.

Through observation, it can be seen that the recognition model of the re-frequency modulation method has high recognition accuracy under different false rates. Although the increase of false pulses will lead to the appearance of invalid values, this invalid value is usually higher The accurate value should be small and can be filtered through data cleaning. Therefore, as the false rate increases, the recognition accuracy does not drop significantly. Therefore, the recognition model has high stability under the condition of false pulse interference.

Example 3: Analysis of the validity experiment of the recognition model under the condition of coexistence of absence and falsehood

As can be seen from the table below, the hybrid supervised neural network model proposed in this paper has certain advantages. Compared with the best decision tree-based method, the accuracy rates of repetition frequency staggering, repetition frequency pulse grouping, repetition frequency jittering, and repetition frequency slipping are increased by 3.4%, 2%, 9.5%, and 4.7%, respectively.

Table 2: Comparison of the effect of this method and other typical methods

.

Claims (5)

1. a kind of radiation source repetition modulation identification method based on deepness belief network, it is characterised in that: including following step It is rapid:

Step 1: typical five kinds of emitter Signals are directed to, representational feature, composition characteristic vector are extracted；

Step 2: in order to make repetition modulation identification network fast convergence, feature vector is normalized；

Step 3: five kinds of output classifications are carried out feature using one-hot coding by the needs in order to meet deep neural network output Digitlization；

Step 4: being input with feature vector, using five kinds of other binary vectors of output class as output, establishes and is believed based on depth Read the radar repetition Modulation Mode Recognition model of network.

2. a kind of radiation source repetition modulation identification method based on deepness belief network according to claim 1, It is characterized in that: five kinds of emitter Signals features and extracting method described in step 1, specific as follows:

The repetitive rate feature of feature 1:PRI sequence are as follows:

Wherein N represents section number, and M represents the number of each section corresponding data appearance, M_maxFor the highest in single section Frequency of occurrence.

The small window repetitive rate feature of feature 2:PRI sequence are as follows:

Wherein N1 represents windowing section number, and M1 represents the number of each section corresponding data appearance, M1_maxFor single section Interior highest frequency of occurrence.

The small window fit slope feature of feature 3:PRI sequence are as follows:

K=p₁xⁿ+p₂x^n-1+…+p_nx+p_n+1

The small window deviation ratio feature of feature 4:PRI sequence are as follows:

Wherein k_maxFor maximum time interval numerical value, k_minFor the smallest time interval numerical value, K_iIndicate each interval in sequence Value, n is sequence length, and win is windowing scale.

Above four features composition characteristic vector:

X={ per, winper, k, b }.

3. a kind of radiation source repetition modulation identification method based on deepness belief network according to claim 1, It is characterized in that: feature vector being normalized described in step 2, the specific method is as follows:

Wherein, x indicates input vector set, x={ per, winper, k, b }, x_minIt is characterized the minimum value of parameter, x_maxFor spy Levy parameter maximum value.

4. a kind of radiation source repetition modulation identification method based on deepness belief network according to claim 1, It is characterized in that: using one-hot coding described in step 3, five seed types being first mapped to integer value；Then, by each integer value table It is shown as binary vector；After coding, repetition fixed type is represented by " 00001 ", and PRF staggering type is represented by " 00010 ", repetition arteries and veins set type are represented by " 00100 ", and repetition wobble type is represented by " 01000 ", and repetition is sliding to become type It is represented by " 10000 ".

5. a kind of radiation source repetition modulation identification method based on deepness belief network according to claim 1, It is characterized in that: with feature vector for input described in step 4, using the other binary vector of five types as output, establishing based on deep Spend the radar repetition Modulation Mode Recognition model of belief network, the specific steps are as follows:

Depth sorting device is using four kinds of features of pulse signal as input, using the binary vector of five kinds of modulation types as defeated Out.Hidden layer selects the RBM of single layer, carries out feature extraction by unsupervised mode, output layer takes feedforward neural network, to have The mode of supervision is classified.