Radar target identification method and system based on feature extraction and matching network

Abstract

The invention discloses a radar target recognition method based on feature extraction and a matching network. The radar target recognition method comprises the following steps: a, performing EMD (Empirical Mode Decomposition) on radar echoes to obtain a first eigenmode function imf1 and the sum imfr of other eigenmode functions; b, a CLEAN algorithm is used for the sum of all eigenmode functions obtained after EMD decomposition, and a fuselage component xF and a micro-motion component xJEM are obtained; c, converting all the components imf1, imfr, xF and xJEM into images by using a Grubm angle field; and d, splicing all the images, and identifying the spliced images by using a matching network to obtain a classification result. According to the invention, part of features are extracted by fusing domain expert knowledge and then classified by using the deep neural network, so that the learning difficulty can be reduced and the neural network prediction precision can be improved.

Description

Technical Field

[0001] The present invention relates to the field of electronic information technology, and in particular to a radar target recognition method and system based on feature extraction and matching network. Background Art

[0002] In modern military operations, air superiority often means absolute initiative on the battlefield. Helicopters are agile and suitable for complex terrain, making them irreplaceable equipment in the air battlefield. Propeller aircraft have superior low-altitude and ultra-low-altitude maneuverability, and mainly implement low-speed multi-purpose operations for battlefield control and battlefield support, with considerable comprehensive combat capabilities. Jet aircraft have high aircraft speed and maneuverability. In combat, they are mainly responsible for seizing air superiority in the theater at high altitude and carrying out precise attacks on enemy fixed targets. They are an important core force in air combat. The three types of aircraft each undertake different military tasks in the air battlefield and have made great contributions to winning the war. Therefore, it is of great significance to study the classification and recognition of the three types of aircraft: helicopters, propellers, and jets. Radar target recognition methods based on traditional classifiers generally include two tasks: first, the characteristic information of the target echo is used to extract the target features through feature engineering, and then a suitable classifier is selected to classify the extracted feature information.

[0003] The above method has the following defects: the validity of the target features and the performance of the classifier are both important factors affecting the final recognition results, and the recognition effect is greatly affected by the environment. The generalization is also insufficient, which makes it difficult to meet the requirements of adaptation in actual recognition scenarios. Summary of the invention

[0004] The present invention aims to provide a radar target recognition method based on feature extraction and matching network, which uses feature engineering to extract key characteristic frequencies in radar echoes, innovatively uses Gram angle field to convert one-dimensional echo data into narrow-band radar echo data, uses Empirical Mode Decomposition (EMD) algorithm to separate the fuselage component and micro-motion component of jet aircraft and helicopter, and uses CLEAN algorithm to remove the maximum harmonic in the signal, thereby separating the fuselage component and micro-motion component of propeller aircraft. Then, all separated signals are converted into images through Gram angle field, and all images are spliced ​​and put into matching network for inference to obtain their types.

[0005] Deep learning technology, which has excellent generalization and adaptability, has solved the above problems of traditional recognition methods and has been widely used in the field of radar target recognition. Through the powerful adaptive learning ability of deep learning, we can find a feature extraction model that can be widely applied to a variety of radar targets and scenes, and combine feature extraction with target recognition through a deep network to obtain an integrated learning and classification model. Therefore, it is necessary to explore new methods to solve the problem of target recognition with a small amount of radar echo data.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A radar target recognition method based on feature extraction and matching network comprises the following steps:

[0008] a. Decompose the radar echo EMD to obtain its first eigenmode function imf1 and the sum of the remaining eigenmode functions imf r ;

[0009] b. Use the CLEAN algorithm to obtain the fuselage component x by summing up all the eigenmode functions obtained after EMD decomposition F and the micromotion component x JEM ;

[0010] c. All components imf1, imf r 、x F and x JEM All are transformed into images using the Gram angular field method;

[0011] d. Splice all images together and use the matching network to identify the spliced ​​images and obtain the classification results;

[0012] Preferably, in step a, the EMD decomposition of the radar echo x(t) is expressed as Among them, imf i represents the i-th eigenmode signal, r L is the remaining term after L decompositions, and the decomposition process is as follows:

[0013] a.1. Initialization: let r(t) = x(t);

[0014] a.2. Extract all local maxima of r(t), interpolate the maximum sequence, and obtain the upper envelope r of r(t) max (t);

[0015] a.3. Extract all local minima of r(t), interpolate the minimum sequence, and obtain the lower envelope r of r(t) min (t);

[0016] a.4. For each moment rmax (t) and r min (t) calculates its average value and obtains the instantaneous average value m(t): m(t) = (r max (t)+r min (t)) / 2;

[0017] a.5. Let r(t) = r(t) - m(t);

[0018] a.6. Determine whether r(t) satisfies the two conditions of IMF. If so, the first IMF is obtained, denoted as imf1. Otherwise, repeat steps a.2 to a.5 on r(t) obtained in step a.5 until r(t) satisfies the two conditions of IMF and obtains imf1.

[0019] Preferably, in step b, the remainder r obtained by EMD decomposition is L It corresponds to the near-ground clutter component, so the signal after clutter suppression can be obtained as The signal can be represented as harmonic components and in the following form:

[0020] The maximum harmonic component x corresponding to the fuselage component is obtained by the CLEAN algorithm F and the micromotion component x JEM ,specific

[0021] Here are the steps:

[0022] b.1. Input time domain signal x EMD ;

[0023] b.2. Calculate the Doppler spectrum of the time domain signal X = FFT (x EMD )

[0024] b.3. Find the maximum point in X and calculate its Doppler frequency f1 according to the position corresponding to the maximum point, and record the corresponding amplitude ρ1 and phase

[0025] b.4. Reconstruct the maximum harmonic component in the signal based on the result of b.3 and obtain the time corresponding to the fuselage component.

[0026] Domain echo; b.5. From the original signal x EMD Subtract the maximum harmonic component reconstructed in step b.4 from the time domain echo x corresponding to the micro-motion component JEM =x EMD -x F .

[0027] Preferably, in step c, the input signal is first scaled to between [-1, 1] using the following formula: where xmin and x max are the maximum and minimum values ​​of the signal, is the scaled x(t);

[0028] Then convert the rectangular coordinate system to the polar coordinate system using the following formula

[0029]

[0030] By calculating the sum of the cosine functions between each time point, the two-dimensional matrix transformed by the Gram angle field is obtained as shown below

[0031]

[0032] Preferably, the matching network introduces an attention mechanism to calculate the predicted output as follows:

[0033]

[0034] where x i ,y i Support set samples and labels, Represents the query set D q A sample of Represents the predicted output of the matching network. Its attention mechanism can be expressed as follows:

[0035]

[0036] Where cosine represents the cosine distance, f(·), g(·) are the embedding functions implemented by the convolutional neural network, and x i represents the samples of the support set, represents a sample of the query set;

[0037] The objective function of the matching network is as follows:

[0038]

[0039] Among them, θ is the parameter of the matching network, D represents the meta-training set, and D s and D q They represent the support set and query set divided by the meta-training set respectively.

[0040] The present invention also discloses a radar target recognition system based on feature extraction and matching network, including: an EMD decomposition module, used for performing EMD decomposition on the radar echo to obtain the sum of its first intrinsic mode function imf1 and the rest of the intrinsic mode functions imf r ;

[0041] The CLEAN algorithm module is used to perform the CLEAN algorithm on the sum of all eigenmode functions obtained after EMD decomposition to obtain the fuselage component x F and the micromotion component x JEM ;

[0042] Gram angular field module, used to convert all components imf1, imf r 、x F and x JEM Transform to image;

[0043] Image stitching module, used to stitch all components imf1, imf r 、x F and x JEM The transformed images are stitched;

[0044] The matching network module is used to use the matching network to identify the spliced ​​images and obtain classification results.

[0045] The beneficial effects of the present invention are as follows:

[0046] 1. The present invention integrates domain expert knowledge to extract some features and then uses a deep neural network for classification, which can reduce the learning difficulty and improve the prediction accuracy of the neural network.

[0047] 2. The present invention uses a matching network instead of traditional LSTM or CNN as a classifier, which improves the small sample learning ability and can achieve better results even when there is little military data. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of radar target recognition method based on feature extraction and matching network.

[0049] Figure 2 This is the structural diagram of the radar target recognition system based on feature extraction and matching network. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings.

[0051] Example 1

[0052] This embodiment discloses a radar target recognition method, such as Figure 1 As shown, this radar target recognition method is a radar target recognition method based on feature extraction and matching network, which is as follows:

[0053] 1. Use Figure 1The radar target recognition system architecture based on feature extraction and matching network shown in the figure is a radar target recognition system architecture based on feature extraction and matching network. First, the radar echo EMD is decomposed to obtain its first intrinsic mode function imf1 and the sum of the remaining intrinsic mode functions imf r Using CLEAN on the sum of all eigenmode functions obtained after EMD decomposition, the fuselage component x F and the micromotion component x JEM . All components imf1, imf r 、x F and x JEM All images are transformed into images using the Gram angular field, and then all images are spliced ​​together. Finally, the matching network is used to identify the spliced ​​images and obtain the classification results.

[0054] 2. Feature extraction based on EMD-CLEAN algorithm

[0055] 1) Characteristic analysis: The flying speed of a jet aircraft is much faster than that of a helicopter. In its radar echo, its fuselage component corresponds to the high-frequency component. The slower flying speed of a helicopter and the fast-rotating propeller cause its fuselage component to correspond to the low-frequency component. The fuselage component of a propeller aircraft is much larger than the micro-motion component.

[0056] The maximum harmonic component can be used to correspond to the fuselage component.

[0057] 2) EMD decomposition (extracting the fuselage and micro-motion components of jet aircraft and helicopters): The EMD decomposition of the original signal x(t) can be expressed as Among them, imf i represents the i-th eigenmode signal, r L is the remaining term after L decompositions. The decomposition process is as follows:

[0058] ① Initialization: let r(t) = x(t);

[0059] ② Extract all local maxima of r(t), interpolate the maximum sequence, and obtain the upper envelope r of r(t) max (t);

[0060] ③ Extract all local minima of r(t), interpolate the minimum sequence, and obtain the lower envelope r of r(t) min (t);

[0061] ④ For each moment r max (t) and r min (t) calculate its average value and obtain the instantaneous average value m(t):

[0062] m(t)=(r max (t)+r min(t)) / 2;

[0063] ⑤Let r(t)=r(t)-m(t);

[0064] ⑥ Determine whether r(t) satisfies the two conditions of IMF. If so, the first IMF is obtained, recorded as imf1; otherwise, repeat the operations of steps ② to ⑤ for r(t) obtained in step ⑤ until r(t) satisfies the two conditions of IMF and obtains imf1;

[0065] 3) CLEAN algorithm extracts the maximum harmonic: the remainder r obtained by EMD decomposition L It corresponds to the near-ground clutter component, so the signal after clutter suppression can be obtained as The signal can be expressed as harmonic components and the form The CLEAN algorithm can obtain the maximum harmonic component x corresponding to the fuselage component. F and the micromotion component x JEM , the specific steps are as follows:

[0066] ① Input time domain signal x EMD ;

[0067] ② Calculate the Doppler spectrum of the time domain signal X = FFT (x EMD );

[0068] ③ Find the maximum point from X and calculate its Doppler frequency f1 according to the position corresponding to the maximum point, and record the corresponding amplitude ρ1 and phase

[0069] ④ Reconstruct the maximum harmonic component in the signal according to the result of step ③ to obtain the time domain echo corresponding to the fuselage component;

[0070] ⑤From the original signal x EMD Subtract the maximum harmonic component reconstructed in step ④ from the time domain echo x corresponding to the micro-motion component JEM =x EMD -x F .

[0071] 3. Calculation of Gram angle field: First, scale the input signal to between [-1,1] using the following formula:

[0072]

[0073] where x min and x max are the maximum and minimum values ​​of the signal, is the scaled x(t). Then the rectangular coordinate system is converted to the polar coordinate system using the following formula

[0074]

[0075] By calculating the sum of the cosine functions between each time point, the two-dimensional matrix transformed by the Gram angle field can be obtained as shown below

[0076]

[0077] 4. Matching network processing: The matching network introduces an attention mechanism to calculate the predicted output as follows:

[0078]

[0079] where x i ,y i Support set samples and labels, Represents the query set D q A sample of Represents the predicted output of the matching network. Its attention mechanism can be expressed as follows

[0080]

[0081] Where cosine represents the cosine distance, f(·), g(·) are the embedding functions implemented by the convolutional neural network, and x i represents the samples of the support set, Represents a sample of the query set. The objective function of the matching network is as follows:

[0082]

[0083] Among them, θ is the parameter of the matching network, D represents the meta-training set, and D s and D q They represent the support set and query set divided by the meta-training set respectively.

[0084] Example 2

[0085] On the basis of Example 1, this embodiment discloses a radar target recognition system based on feature extraction and matching network. The radar target recognition system based on feature extraction and matching network adopts the radar target recognition method based on feature extraction and matching network as in Example 1. Figure 2 As shown, the system includes:

[0086] The EMD decomposition module is used to perform EMD decomposition on the radar echo to obtain the first intrinsic mode function imf1 and the sum of the remaining intrinsic mode functions imf r ;

[0087] The CLEAN algorithm module is used to perform the CLEAN algorithm on the sum of all eigenmode functions obtained after EMD decomposition to obtain the fuselage component xF and the micromotion component x JEM ;

[0088] Gram angular field module, used to convert all components imf1, imf r 、x F and x JEM Transform to image;

[0089] Image stitching module, used to stitch all components imf1, imf r 、x F and x JEM The transformed images are stitched;

[0090] The matching network module is used to use the matching network to identify the spliced ​​images and obtain classification results.

[0091] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, technicians familiar with the field may make various corresponding changes and deformations based on the present invention, but these corresponding changes and deformations should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A radar target recognition method based on feature extraction and matching network, characterized in that: The following steps are involved: a. Decompose the radar echo EMD to obtain its first eigenmode function imf1 and the sum of the remaining eigenmode functions imf r ; b. Use the CLEAN algorithm to obtain the fuselage component x by summing up all the eigenmode functions obtained after EMD decomposition F and the micromotion component x JEM ; c. All components imf1, imf r 、x F and x JEM All are transformed into images using the Gram angular field method; d. Splice all images and use the matching network to identify the spliced ​​images to obtain the classification results.

2. The radar target recognition method based on feature extraction and matching network according to claim 1 is characterized in that: In step a, the EMD decomposition of the radar echo x(t) is expressed as Among them, imf i represents the i-th eigenmode signal, r L is the remaining term after L decompositions, and the decomposition process is as follows: a.

1. Initialization: let r(t) = x(t); a.

2. Extract all local maxima of r(t), interpolate the maximum sequence, and obtain the upper envelope r of r(t) max (t); a.

3. Extract all local minima of r(t), interpolate the minimum sequence, and obtain the lower envelope r of r(t) min (t); a.

4. For each moment r max (t) and r min (t) calculates its average value and obtains the instantaneous average value m(t): m(t) = (r max (t)+r min (t)) / 2; a.

5. Let r(t) = r(t) - m(t); a.

6. Determine whether r(t) satisfies the two conditions of IMF. If so, the first IMF is obtained, denoted as imf1. Otherwise, repeat steps a.2 to a.5 on r(t) obtained in step a.5 until r(t) satisfies the two conditions of IMF and obtains imf1.

3. The radar target recognition method based on feature extraction and matching network according to claim 2, It is characterized in that In step b, the remainder r obtained by EMD decomposition L It corresponds to the near-ground clutter component, so the signal after clutter suppression can be obtained as The signal can be represented as harmonic components and in the following form: The maximum harmonic component x corresponding to the fuselage component is obtained by the CLEAN algorithm F and the micromotion component x JEM , the specific steps are as follows: b.

1. Input time domain signal x EMD ; b.

2. Calculate the Doppler spectrum of the time domain signal X = FFT (x EMD ) b.

3. Find the maximum point from X and calculate its Doppler frequency f1 according to the position corresponding to the maximum point. Record the corresponding amplitude ρ1 and phase b.

4. Reconstruct the maximum harmonic component in the signal based on the result of b.3 to obtain the time domain echo corresponding to the fuselage component; b.

5. From the original signal x EMD Subtract the maximum harmonic component reconstructed in step b.4 from the time domain echo x corresponding to the micro-motion component JEM =x EMD -x F .

4. The radar target recognition method based on feature extraction and matching network according to claim 3 is characterized in that: In step c, the input signal is first scaled to between [-1,1] using the following formula: where x min and x max are the maximum and minimum values ​​of the signal, is the scaled x(t); Then convert the rectangular coordinate system to the polar coordinate system using the following formula By calculating the sum of the cosine functions between each time point, the two-dimensional matrix transformed by the Gram angle field is obtained as shown below 5. The radar target recognition method based on feature extraction and matching network according to claim 3 is characterized in that: The matching network introduces an attention mechanism to calculate the predicted output as follows: where x i ,y i Support set samples and labels, Represents the query set D q A sample of represents the predicted output of the matching network. Its attention mechanism can be expressed as follows: Where cosine represents the cosine distance, f(·), g(·) are the embedding functions implemented by the convolutional neural network, and x i represents the samples of the support set, represents a sample of the query set; The objective function of the matching network is as follows: Among them, θ is the parameter of the matching network, D represents the meta-training set, and D s and D q They represent the support set and query set divided by the meta-training set respectively.

6. Radar target recognition system based on feature extraction and matching network, characterized in that: include: The EMD decomposition module is used to perform EMD decomposition on the radar echo to obtain its first intrinsic mode function imf1 and the sum imf of the remaining intrinsic mode functions. r ; The CLEAN algorithm module is used to perform the CLEAN algorithm on the sum of all eigenmode functions obtained after EMD decomposition to obtain the fuselage component x F and the micromotion component x JEM ; Gram angular field module, used to convert all components imf1, imf r 、x F and x JEM Transform to image; Image stitching module, used to stitch all components imf1, imf r 、x F and x JEM The transformed images are stitched together; The matching network module is used to use the matching network to identify the spliced ​​images and obtain classification results.

Images