CN114137518B - Radar high-resolution range profile open set identification method and device - Google Patents
Radar high-resolution range profile open set identification method and device Download PDFInfo
- Publication number
- CN114137518B CN114137518B CN202111199838.4A CN202111199838A CN114137518B CN 114137518 B CN114137518 B CN 114137518B CN 202111199838 A CN202111199838 A CN 202111199838A CN 114137518 B CN114137518 B CN 114137518B
- Authority
- CN
- China
- Prior art keywords
- neural network
- convolutional neural
- sample set
- layer
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 54
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 105
- 238000012549 training Methods 0.000 claims abstract description 73
- 238000012360 testing method Methods 0.000 claims abstract description 54
- 238000007781 pre-processing Methods 0.000 claims abstract description 13
- 238000012545 processing Methods 0.000 claims description 27
- 238000010606 normalization Methods 0.000 claims description 18
- 239000011159 matrix material Substances 0.000 claims description 7
- 230000005484 gravity Effects 0.000 claims description 6
- 238000010276 construction Methods 0.000 claims description 5
- 238000005096 rolling process Methods 0.000 claims description 5
- 230000009466 transformation Effects 0.000 claims description 3
- 230000006870 function Effects 0.000 description 14
- 238000004088 simulation Methods 0.000 description 14
- 230000008569 process Effects 0.000 description 7
- 230000035945 sensitivity Effects 0.000 description 6
- 230000004913 activation Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000012706 support-vector machine Methods 0.000 description 3
- 206010063385 Intellectualisation Diseases 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000013519 translation Methods 0.000 description 2
- 240000004760 Pimpinella anisum Species 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000001427 coherent effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S13/00—Systems using the reflection or reradiation of radio waves, e.g. radar systems; Analogous systems using reflection or reradiation of waves whose nature or wavelength is irrelevant or unspecified
- G01S13/88—Radar or analogous systems specially adapted for specific applications
- G01S13/89—Radar or analogous systems specially adapted for specific applications for mapping or imaging
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Remote Sensing (AREA)
- Theoretical Computer Science (AREA)
- Radar, Positioning & Navigation (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Electromagnetism (AREA)
- Computer Networks & Wireless Communication (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a radar high-resolution range profile open set identification method and device based on a convolutional neural network, wherein the method comprises the following steps: acquiring a radar high-resolution range profile and establishing a training sample set and a test sample set; preprocessing data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set; constructing a convolutional neural network model; training a convolutional neural network model by utilizing the preprocessed training sample set to obtain a trained convolutional neural network; and performing open-set recognition on the trained convolutional neural network by using the preprocessed test sample set to obtain a radar high-resolution range profile open-set recognition result based on the convolutional neural network. The method provided by the invention can be used for identifying and classifying the known class targets in the library, and meanwhile, the unknown class targets outside the library can be refused to be judged, so that the target identification accuracy is improved, and the automation and intelligent level of the radar is further improved.
Description
Technical Field
The invention belongs to the technical field of radars, and particularly relates to a radar high-resolution range profile open set identification method and device based on a convolutional neural network.
Background
The range resolution of the radar is proportional to the received pulse width after matched filtering, and the range unit length of the radar transmitting signal meets the following conditions: Δr=cτ/2=c/2B, where Δr is the distance unit length of the radar transmit signal, c is the speed of light, τ is the pulse width of the matching reception, and B is the bandwidth of the radar transmit signal. The large radar transmit signal bandwidth provides high range resolution (HRR, high resolution range). In practice, the range resolution of a radar is high or low relative to an observed target, and when the observed target has a dimension L along the radar line of sight, if L < < Δr, the corresponding radar echo signal width is approximately the same as the radar transmission pulse width (the received pulse after the matching process), and is generally referred to as "point" target echo, such radar is a low resolution radar, and if Δr < < L, the target echo becomes a "one-dimensional range profile" extending in distance according to the target characteristics, such radar is a high resolution radar (<indicatesmuch smaller).
The high resolution radar transmits a broadband coherent signal (a chirped or stepped frequency signal) and receives echo data by back scattering of the transmitted electromagnetic wave by the target. Typically the echo characteristics are calculated using a simplified scatter point model, i.e. using a Born first order approximation that ignores multiple scatter. The fluctuation and peak presented in the high-resolution radar echo reflect the distribution condition of radar scattering cross-sectional areas (RCS) of scattering bodies (such as a nose, a wing, a tail rudder, an air inlet hole, an engine and the like) on a target at a certain radar view angle along the radar sight line (RLOS), and reflect the relative geometrical relationship of scattering points in the radial direction, which is commonly called as high-resolution range profile (HRRP, high resolution range profile). Thus, HRRP samples contain important structural features of the target, which is valuable for target identification and classification.
The traditional target recognition method aiming at the high-resolution range profile data mainly adopts a support vector machine to directly classify targets, or adopts a characteristic extraction method based on a limiting Boltzmann machine to firstly project the data into a high-dimensional space and then classify the data by using a classifier. However, the method only uses the time domain characteristics of the signals, and the target recognition accuracy is not high.
In recent years, target recognition methods for radar high-resolution range profile data are mainly aimed at closed set recognition, and the recognition requires that the data types in a test sample set and the data types in a training sample set are consistent. However, in practical applications, the radar may capture not only high-resolution range images of objects within the library, but also high-resolution range images of objects of unknown classes outside the library. Under the condition, the existing closed set recognition algorithm cannot reject the unknown class data outside the library, but can misjudge the unknown class data as a class in the library, and the target recognition accuracy of the radar is greatly reduced under the condition.
Thus, some researchers began to study on the open set identification of radar high-resolution range profiles. On the basis of Support Vector Domain Description (SVDD), a Multi-core support vector domain description (Multi KERNEL SVDD) model is proposed by Chai Jing et al to more flexibly describe Multi-mode distribution of HRRP data in a high-dimensional feature space, so as to improve the identification and rejection performance of radar HRRP. Zhang Xuefeng et al propose a multi-classifier fusion algorithm based on a maximum correlation classifier (Maximum correlation classifier, MCC), a support vector machine (Support vector machine, SVM) and an association vector machine (RELEVANCE VECTOR MACHINE, RVM) to implement the rejection and recognition functions of the radar HRRP. However, both of the above algorithms require the reliance on a specific form of kernel function to extract features, which limits the ability of the model to extract sufficiently separable features, thereby affecting the accuracy of target recognition and the level of radar intelligence.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a radar high-resolution range profile open set identification method and device based on a convolutional neural network. The technical problems to be solved by the invention are realized by the following technical scheme:
a radar high-resolution range profile open set identification method based on a convolutional neural network comprises the following steps:
Acquiring a radar high-resolution range profile and establishing a training sample set and a test sample set; the training sample set comprises a plurality of target radar high-resolution range profiles of known classes, and the test sample set comprises a plurality of target radar high-resolution range profiles of known classes in a library and unknown classes outside the library;
preprocessing the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set;
Constructing a convolutional neural network model;
Training the convolutional neural network model by using the preprocessed training sample set to obtain a trained convolutional neural network;
And performing open set identification on the trained convolutional neural network by using the preprocessed test sample set to obtain a radar high-resolution range profile open set identification result based on the convolutional neural network.
In one embodiment of the invention, preprocessing the training sample set and the test sample set comprises:
and sequentially carrying out center of gravity alignment and normalization on the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set.
In one embodiment of the invention, constructing the convolutional neural network model includes:
constructing a convolutional neural network model with a four-layer structure; the four-layer structure is respectively a first layer of convolution layer, a second layer of convolution layer, a third layer of convolution layer and a fourth layer of full-connection layer, and each convolution layer is provided with the same convolution step length; wherein each convolution layer comprises a plurality of convolution kernels, and the size of each convolution kernel is the same.
In one embodiment of the present invention, the method further comprises constructing a loss function of the convolutional neural network, wherein the loss function is expressed as follows:
Wherein Θ (x) is the output result of the convolutional neural network, O i (i=1, …, N) is N prototypes randomly initialized according to gaussian distribution, d (Θ (x), O k) is the euclidean distance from Θ (x) to O k, λ is a hyper-parameter, r i＝d(Oi,Oc), representing each prototype O i to the center Is a distance of (3).
In one embodiment of the present invention, training the convolutional neural network model using the preprocessed training sample set to obtain a trained convolutional neural network, including:
Randomly dividing the preprocessed training sample set into q batches, wherein the data of each batch is n multiplied by D dimensional matrix data; wherein, Floor () represents a downward rounding, and P represents the number of high-resolution range profiles in the training sample set;
sequentially inputting the data of each batch into a convolutional neural network for processing to obtain an output result of the convolutional neural network;
and calculating the value of the loss function according to the output result of the convolutional neural network, and updating the parameter value of the convolutional neural network by using a random gradient method until the network converges, so as to obtain the trained convolutional neural network.
In one embodiment of the present invention, inputting each batch of data into a trained convolutional neural network in sequence for processing to obtain an output result of the convolutional neural network, including:
performing convolution and downsampling processing on the current input data by using a first convolution layer to obtain a first feature map;
rolling and downsampling the first feature map by using a second convolution layer to obtain a second feature map;
rolling and downsampling the second feature map by using a third convolution layer to obtain a third feature map;
Performing nonlinear transformation processing on the third feature map by using a fourth full-connection layer to obtain a processing result of current data;
repeating the steps until all the input data are processed, and obtaining the output result of the convolutional neural network.
In one embodiment of the present invention, performing open-set recognition on the trained convolutional neural network by using the preprocessed test sample set to obtain a radar high-resolution range profile open-set recognition result based on the convolutional neural network, including:
Let the probability that the convolutional neural network predicts that the sample x to be measured belongs to the category k be The expression is as follows:
Probability when a sample is to be tested When the sample to be detected is smaller than a preset threshold value, judging the sample to be detected as an out-of-library unknown class; otherwise, the class is judged as a known class in the library, and the specific class to which the class belongs is further obtained.
Another embodiment of the present invention provides a radar high-resolution range profile open set recognition device based on a convolutional neural network, including:
The data acquisition module is used for acquiring a radar high-resolution range profile and establishing a training sample set and a test sample set; the training sample set comprises a plurality of target radar high-resolution range profiles of known classes, and the test sample set comprises a plurality of target radar high-resolution range profiles of known classes in a library and unknown classes outside the library;
The preprocessing module is used for preprocessing the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set;
The model construction module is used for constructing a convolutional neural network model;
The training module is used for training the convolutional neural network model by utilizing the preprocessed training sample set to obtain a trained convolutional neural network;
And the target identification module is used for carrying out open set identification on the trained convolutional neural network by utilizing the preprocessed test sample set to obtain a radar high-resolution range profile open set identification result based on the convolutional neural network.
The invention has the beneficial effects that:
1. The radar high-resolution range profile open set recognition method provided by the invention can combine the primary characteristics of each layer by adopting the convolutional neural network technology, so that the characteristics of higher layers are obtained for recognition, the recognition rate is obviously improved, the method can be used for recognizing and classifying known class targets in a library, meanwhile, unknown class targets outside the library can be refused to be judged, the target recognition accuracy is improved, and the automation and intellectualization level of the radar is further improved;
2. The invention adopts a multi-layer convolution neural network structure and performs energy normalization and alignment pretreatment on the data, so that the high-layer characteristics of the high-resolution range profile data can be mined, the amplitude sensitivity, the translation sensitivity and the gesture sensitivity of the high-resolution range profile data are removed, and compared with the traditional direct classification method, the method has stronger robustness.
The present invention will be described in further detail with reference to the accompanying drawings and examples.
Drawings
Fig. 1 is a schematic flow chart of a radar high-resolution range profile open set recognition method based on a convolutional neural network provided by the embodiment of the invention;
fig. 2 is a schematic structural diagram of a radar high-resolution range profile open set recognition device based on a convolutional neural network according to an embodiment of the present invention;
FIG. 3 is a simulation test result provided by an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to specific examples, but embodiments of the present invention are not limited thereto.
Example 1
Referring to fig. 1, fig. 1 is a schematic flow chart of a method for identifying a radar high-resolution range profile open set based on a convolutional neural network, which includes the following steps:
step 1: acquiring a radar high-resolution range profile and establishing a training sample set and a test sample set; the training sample set comprises a plurality of target radar high-resolution range profiles of known classes, and the test sample set comprises a plurality of target radar high-resolution range profiles of known classes in a library and unknown classes outside the library.
Firstly, P radar high-resolution range profile raw data of N categories are obtained to serve as a training sample set, wherein N is more than or equal to 3, and P is more than or equal to 900;
Then, Q radar high-resolution range profile raw data of N training sample categories and L radar high-resolution range profile raw data of M unknown categories are obtained as a test sample set, wherein Q is more than or equal to 900, M is more than or equal to 1, and L is more than or equal to 300.
Step 2: and preprocessing the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set.
In this embodiment, the data in the training sample set and the test sample set are sequentially subjected to center of gravity alignment and normalization processing, so as to obtain a preprocessed training sample set and a preprocessed test sample set.
Specifically, the original data in the training sample set or the test sample set is recorded as x 0, and firstly, the original data x 0 is subjected to center of gravity alignment to obtain data x' 0 with the aligned center of gravity; then, normalization processing is carried out on the data x' 0 aligned according to the center of gravity, and normalized data x is obtained, wherein the expression is as follows:
The preprocessed training sample set and the test sample set are respectively P multiplied by D, (Q+L) multiplied by D dimensional matrixes, wherein D represents the total number of distance units contained in the original data of the radar high-resolution range profile.
Step 3: and constructing a convolutional neural network model.
In this embodiment, the convolutional neural network model is configured to have a four-layer structure, and includes three convolutional layers and a full-connection layer, which are respectively denoted as a first convolutional layer, a second convolutional layer, a third convolutional layer, and a fourth full-connection layer. Each convolution layer has the same convolution step length, and each convolution layer comprises a plurality of convolution kernels, and the sizes of the convolution kernels are the same.
Specifically, for the first layer convolution layer:
Setting the first layer convolution layer to comprise C convolution kernels, and marking the C convolution kernels of the first layer convolution layer as K, wherein the size of K is set to be 1 xw multiplied by 1, w represents each convolution kernel window in the first layer convolution layer, and 1< w < D; c is a positive integer greater than 0; setting the convolution step length of the first layer of convolution layer as L; simultaneously setting the size of a kernel window of the downsampling process of the first layer of convolution layer to be m multiplied by m, wherein m is 1< m < D, D represents the total number of distance units respectively contained in each type of high-resolution distance imaging data in the training sample, and m is a positive integer greater than 0; setting the step length of the downsampling process of the first layer of convolution layer as I, wherein the I and the m are equal in value.
Setting the activation function of the first convolution layer asX represents the sample data after the preprocessing,Representing the convolution operation, b represents the full 1 offset of the third layer convolution layer.
For the second layer convolution layer:
Setting the convolution kernel comprising C 'convolution kernels, and marking the C' convolution kernels of the second layer of convolution layer as K ', wherein the size of the C' convolution kernels is the same as the size of the convolution kernels of the first layer of convolution layer; the convolution step length of the second layer of convolution layer is marked as L ', w is less than or equal to L ' < D-w, and the L ' is equal to the convolution step length L of the first layer of convolution layer; simultaneously setting the size of a kernel window of the downsampling process of the second convolution layer to be m 'x m',1<m '< D, wherein m' is a positive integer greater than 0; the step length of the second layer downsampling process is equal to the value of I ', I ' and m '.
Setting the activation function of the second convolution layer to A first feature map representing the output of the first layer of convolutional layers,Representing the convolution operation, b' represents the full 1 offset of the second layer convolution layer.
For the third layer of convolution layer:
Setting the three layers of convolution layers to be C 'convolution kernels, wherein the C' convolution kernels of the third layer of convolution layers are K 'and the size of the C' convolution kernels is the same as the size of each convolution kernel window in the second layer of convolution layers; setting the convolution step length of the third layer of convolution layer as L 'which is equal to the convolution step length L' of the second layer of convolution layer; simultaneously setting the size of a kernel window of the downsampling treatment of the third convolution layer to be m '. Times.m', 1<m '< D, and m' being a positive integer greater than 0; the step sizes of the third layer downsampling process are I ', and the I ' and m ' are equal.
Setting the activation function of the second convolution layer to A second feature map representing the output of the second convolutional layer,Representing the convolution operation, b "represents the full 1 offset of the third layer of convolution layers.
For the fourth fully connected layer:
setting a weight matrix of random initialization In the form of a B x U dimensional matrix,Floor () represents downward rounding, D represents total number of distance units contained in each type of high-resolution distance imaging data in the training sample, B is more than or equal to D, and B is a positive integer greater than 0; setting the activation function asA third feature map representing the output of a third layer of convolutional layers,Representing an all 1 bias of the fourth fully connected layer, anIs U x 1 dimension.
After the model of the convolutional neural network is built, the method further comprises the step of building a loss function of the convolutional neural network, wherein the loss function is expressed as follows:
Wherein Θ (x) is the output result of the convolutional neural network, O i (i=1, …, N) is N prototypes randomly initialized according to gaussian distribution, d (Θ (x), O k) is the euclidean distance from Θ (x) to O k, λ is a hyper-parameter, r i＝d(Oi,Oc), representing each prototype O i to the center Is a distance of (3).
Step 4: training a convolutional neural network model by utilizing the preprocessed training sample set to obtain a trained convolutional neural network, which specifically comprises the following steps:
41 Randomly dividing the preprocessed training sample set into q batches, wherein the data of each batch is n multiplied by D dimensional matrix data; wherein, Floor () represents a floor, P represents the number of high resolution range profiles in the training sample set.
42 The data of each batch are sequentially input into the convolutional neural network for processing, and an output result of the convolutional neural network is obtained.
42-1) After inputting the data into the convolutional neural network, performing convolution and downsampling processing on the current input data by using a first layer of convolutional layer to obtain a first characteristic diagram.
Specifically, the convolution step length L of the first layer of convolution layer is used for respectively convolving the input data x with the C convolution kernels of the first layer of convolution layer to obtain C convolved results of the first layer of convolution layer, and the C convolved results are recorded as C characteristic graphs y of the first layer of convolution layer:
Carrying out Gaussian normalization processing on the C feature graphs y of the first layer of convolution layer to obtain C feature graphs of the first layer of convolution layer after Gaussian normalization processing
For a pair ofEach feature map in the first layer is respectively subjected to downsampling treatment to obtain C feature maps after downsampling treatment of the first layer of convolution layerI.e. a first feature map, expressed as:
Wherein, C feature graphs representing a first layer convolution layer after Gaussian normalization within a kernel window size m x m of the first layer downsampling processIs set at the maximum value of (c),C feature maps of the first layer convolution layer after the Gaussian normalization processing are shown.
42-2) Performing convolution and downsampling processing on the first feature map by using the second convolution layer to obtain a second feature map.
Specifically, the C feature maps after the downsampling processing of the first layer convolution layer are processed by using the convolution step L' of the second layer convolution layerRespectively convolving the C 'convolution kernels K' of the second layer convolution layer with the first feature map to obtain C 'convolved results of the second layer convolution layer, and marking the results as C' feature maps of the second layer convolution layer
C' feature maps for a second convolution layerCarrying out Gaussian normalization processing to obtain C' feature graphs of the second-layer convolution layer after Gaussian normalization processing
For a pair ofEach feature map in the second layer is respectively subjected to downsampling treatment, so that C' feature maps after downsampling treatment of the second layer convolution layer are obtainedI.e. a second feature map, expressed as:
Wherein, C ' feature graphs representing the convolutions of the second layer after Gaussian normalization within the kernel window size m ' x m ' of the second layer downsampling processIs set at the maximum value of (c),C' feature maps of the second layer convolution layer after the Gaussian normalization processing are shown.
42-3) Rolling and downsampling the second feature map by using the third convolution layer to obtain a third feature map.
Specifically, the C characteristic graphs after the downsampling treatment of the second layer convolution layer are processed by using the convolution step length L' of the third layer convolution layer(I.e. the second feature map) is convolved with C "convolution kernels K" of the third layer of convolution layers, obtaining C 'convolutions of the third layer of convolution layer and marking the results as C' characteristic graphs of the third layer of convolution layer
C' feature maps for a third layer of convolution layersCarrying out Gaussian normalization processing to obtain C' feature graphs of a third layer of convolution layer after Gaussian normalization processing
For a pair ofEach feature map in the third layer is respectively subjected to downsampling treatment, so that C' feature maps after downsampling treatment of the third layer convolution layer are obtainedI.e. a third feature map, expressed as:
Wherein, C ' feature graphs representing a third layer of convolution layers after Gaussian normalization within a kernel window size of m ' x m ' for the third layer downsampling processIs set at the maximum value of (c),C' feature maps of the third layer of convolution layer after the Gaussian normalization treatment are shown.
42-4) Performing nonlinear transformation processing on the third feature map by using a fourth full-connection layer to obtain a processing result of the current dataThe expression is as follows:
Wherein, A randomly initialized weight matrix representing the fourth fully connected layer,Representing the full 1 bias of the fourth fully connected layer.
42-5) Repeating the steps until all input data are processed, and obtaining the output result of the convolutional neural network.
43 Calculating the value of the loss function according to the output result of the convolutional neural network, and updating the parameter value of the convolutional neural network by using a random gradient method until the network converges to obtain the trained convolutional neural network.
Specifically, the step 42) is performedSubstituting the output result of the convolutional neural network into a loss function expression to obtain a loss function value, and updating the parameter value of the convolutional neural network by adopting the existing random gradient method until the network converges to obtain the trained convolutional neural network. The random gradient method is mature prior art, and this embodiment is not specifically described.
According to the embodiment, the multilayer convolutional neural network structure is adopted, the data are subjected to energy normalization and alignment pretreatment, the high-layer characteristics of the high-resolution range profile data can be mined, the amplitude sensitivity, the translation sensitivity and the gesture sensitivity of the high-resolution range profile data are removed, and compared with a traditional direct classification method, the method has stronger robustness.
Step 5: and performing open set identification on the trained convolutional neural network by using the preprocessed test sample set to obtain a radar high-resolution range profile open set identification result based on the convolutional neural network.
51 Assuming that the probability that the convolutional neural network predicts that the sample x to be detected belongs to the category k isThe expression is as follows:
52 Probability when a sample is to be tested When the sample to be detected is smaller than a preset threshold value, judging the sample to be detected as an out-of-library unknown class; otherwise, the class is judged as a known class in the library, and the specific class to which the class belongs is further obtained.
Specifically, according to training results of known class data in the library, a threshold value tau is set, and when the probability of the sample to be tested is highWhen the sample to be detected is smaller than the threshold tau, the sample to be detected is judged to be of an out-of-library unknown class; otherwise, the specific category k is determined as the known category in the library according to the formula in the step 51).
The radar high-resolution range profile open set recognition method provided by the embodiment can combine the primary characteristics of each layer by adopting the convolutional neural network technology, so that the characteristics of a higher layer are obtained for recognition, the recognition rate is remarkably improved, the method can be used for recognizing and classifying known class targets in a library, meanwhile, unknown class targets outside the library can be refused to be judged, the target recognition accuracy is improved, and the automation and intellectualization level of the radar is further improved.
Example two
On the basis of the first embodiment, the present embodiment further provides a radar high-resolution range profile open set recognition device based on the convolutional neural network. Referring to fig. 2, fig. 2 is a schematic structural diagram of a radar high-resolution range profile open set recognition device based on a convolutional neural network, which includes:
the data acquisition module 1 is used for acquiring radar high-resolution range profiles and establishing a training sample set and a test sample set; the training sample set comprises a plurality of target radar high-resolution range profiles of known classes, and the test sample set comprises a plurality of target radar high-resolution range profiles of known classes in a library and unknown classes outside the library;
The preprocessing module 2 is used for preprocessing the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set;
The model construction module 3 is used for constructing a convolutional neural network model;
the training module 4 is used for training the convolutional neural network model by using the preprocessed training sample set to obtain a trained convolutional neural network;
And the target recognition module 5 is used for performing open set recognition on the trained convolutional neural network by utilizing the preprocessed test sample set to obtain a radar high-resolution range profile open set recognition result based on the convolutional neural network.
The radar high-resolution range profile open set recognition device provided in this embodiment can implement the radar high-resolution range profile open set recognition method provided in the first embodiment, and detailed processes are not described here again.
Therefore, the radar high-resolution range profile open set recognition device provided by the embodiment has the advantages of being capable of recognizing and classifying known class targets in the library, meanwhile being capable of refusing to judge unknown class targets outside the library, and being high in target recognition accuracy.
Example III
The beneficial effects of the invention are verified and illustrated by simulation tests.
1. Simulation conditions
The hardware platform of the simulation experiment of this embodiment is:
A processor: intel (R) Core (TM) i9-10980XE, with a dominant frequency of 3.00GHz and 256GB of memory.
The software platform of the simulation experiment of this embodiment is: ubuntu 20.04 operating system and python 3.9.
The data used in the simulation test are measured data of high-resolution range profiles of 10 types of civil aviation aircrafts, wherein the types of the 10 types of civil aviation aircrafts are A319, A320, A330-2, A330-3, B737-8, CRJ-900, A321, A350-941, B737-7 and B747-89L respectively. The first 6 kinds of aircrafts are used as known object classes in the library, the second 4 kinds of aircrafts are used as unknown object classes outside the library, and a training sample set and a testing sample set are manufactured. Wherein the training sample set is 30921 samples in total, and the number of the class samples in each library is about 5000; the test sample set includes a total of 18424 samples of 6 known classes within the library and a total of 12359 samples of 4 unknown classes outside the library, each class sample being about 3000.
Prior to performing the experiment, all raw data were preprocessed according to step 2 in the first embodiment described above, and then the open set identification experiment was performed using a convolutional neural network.
2. Simulation content and result analysis
The simulation experiment compares the method of the invention with the traditional two-stage refusal judgment and identification method and the softMax threshold method.
The traditional two-stage refusing identification method is mainly used for further classifying the objects judged to be in the library by using SVDD, OCSVM, isolation-Forest and other methods based on refusing the objects outside the library. The SoftMax threshold rule considers that the target features in the library extracted by the convolutional neural network are presented as a plurality of clusters in a feature space, the clusters can be respectively wrapped by a plurality of hyperspheres, if the features of the sample to be detected fall into the hyperspheres, the features belong to the category, and if the features do not fall into any hypersphere, the features are judged as targets outside the library.
The simulation experiment utilizes the area AUC under the working characteristic curve (ROC) of a subject to evaluate the refusing capability of different methods to the targets outside the library, wherein the larger the AUC value is, the stronger the refusing capability to the targets outside the library is represented.
Referring to fig. 3, fig. 3 is a comparison chart of simulation test results provided by the embodiment of the present invention, and as can be seen from fig. 3, in the simulation test, the rejection capability of the present invention to the object outside the library is strongest, and next, the rejection capability of the present invention to the object outside the library is generally obtained by using the SoftMax threshold method, which is 3 conventional methods.
Because the simulation experiment uses more data types, the Macro Average F1-Score is used for comprehensively evaluating the open set recognition capability of different methods, wherein the larger the value of the F1-Score is, the stronger the open set recognition capability is represented. Simulation experiment results are shown in the following table.
| Method of | AUC | F1-Score |
| SVDD+SVM | 0.5552 | 0.4394 |
| OCSVM+SVM | 0.5443 | 0.3216 |
| Isolation Forest+SVM | 0.4825 | 0.4283 |
| SoftMax thresholding | 0.7141 | 0.4627 |
| The invention is that | 0.8329 | 0.5418 |
It can be seen that in the simulation experiment, the comprehensive open set recognition capability of the invention is strongest and is obviously superior to other 4 methods.
In summary, the invention proves the effectiveness of the invention, whether in terms of refusing judgment capability to objects outside the warehouse or comprehensively considering the open set recognition capability, the invention obtains the optimal result.
The foregoing is a further detailed description of the invention in connection with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the invention, and these should be considered to be within the scope of the invention.
Claims (4)
1. A radar high-resolution range profile open set identification method based on a convolutional neural network is characterized by comprising the following steps:
Acquiring a radar high-resolution range profile and establishing a training sample set and a test sample set; the training sample set comprises a plurality of target radar high-resolution range profiles of known classes, and the test sample set comprises a plurality of target radar high-resolution range profiles of known classes in a library and unknown classes outside the library;
preprocessing the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set;
the construction of the convolutional neural network model specifically comprises the following steps:
Constructing a convolutional neural network model with a four-layer structure; the four-layer structure is respectively a first layer of convolution layer, a second layer of convolution layer, a third layer of convolution layer and a fourth layer of full-connection layer, and each convolution layer is provided with the same convolution step length; each convolution layer comprises a plurality of convolution kernels, and the sizes of the convolution kernels are the same;
constructing a loss function of a convolutional neural network, wherein the expression is as follows:
Wherein Θ (x) is the output result of the convolutional neural network, O i (i=1, …, N) is N prototypes randomly initialized according to gaussian distribution, d (Θ (x), O k) is the euclidean distance from Θ (x) to O k, λ is a hyper-parameter, r i＝d(Oi,Oc), representing each prototype O i to the center Is a distance of (2);
Training the convolutional neural network model by using the preprocessed training sample set to obtain a trained convolutional neural network, which specifically comprises the following steps:
Randomly dividing the preprocessed training sample set into q batches, wherein the data of each batch is n multiplied by D dimensional matrix data; wherein, Floor () represents a downward rounding, and P represents the number of high-resolution range profiles in the training sample set;
sequentially inputting the data of each batch into a convolutional neural network for processing to obtain an output result of the convolutional neural network;
Calculating the value of a loss function according to the output result of the convolutional neural network, and updating the parameter value of the convolutional neural network by using a random gradient method until the network converges to obtain a trained convolutional neural network;
Performing open set identification on the trained convolutional neural network by using the preprocessed test sample set to obtain a radar high-resolution range profile open set identification result based on the convolutional neural network, wherein the method specifically comprises the following steps of:
Let the probability that the convolutional neural network predicts that the sample x to be measured belongs to the category k be The expression is as follows:
Probability when a sample is to be tested When the sample to be detected is smaller than a preset threshold value, judging the sample to be detected as an out-of-library unknown class; otherwise, the class is judged as a known class in the library, and the specific class to which the class belongs is further obtained.
2. The radar high resolution range profile open set identification method of claim 1, wherein preprocessing the training sample set and the test sample set comprises:
and sequentially carrying out center of gravity alignment and normalization on the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set.
3. The radar high-resolution range profile open set recognition method according to claim 1, wherein the step of sequentially inputting the data of each batch into the trained convolutional neural network for processing to obtain an output result of the convolutional neural network comprises the steps of:
performing convolution and downsampling processing on the current input data by using a first convolution layer to obtain a first feature map;
rolling and downsampling the first feature map by using a second convolution layer to obtain a second feature map;
rolling and downsampling the second feature map by using a third convolution layer to obtain a third feature map;
Performing nonlinear transformation processing on the third feature map by using a fourth full-connection layer to obtain a processing result of current data;
repeating the steps until all the input data are processed, and obtaining the output result of the convolutional neural network.
4. A radar high-resolution range profile open set recognition device based on a convolutional neural network is characterized by comprising:
The data acquisition module (1) is used for acquiring radar high-resolution range profiles and establishing a training sample set and a test sample set; the training sample set comprises a plurality of target radar high-resolution range profiles of known classes, and the test sample set comprises a plurality of target radar high-resolution range profiles of known classes in a library and unknown classes outside the library;
the preprocessing module (2) is used for preprocessing the data in the training sample set and the test sample set to obtain a preprocessed training sample set and a preprocessed test sample set;
the model construction module (3) is used for constructing a convolutional neural network model;
The training module (4) is used for training the convolutional neural network model by utilizing the preprocessed training sample set to obtain a trained convolutional neural network;
The target recognition module (5) is used for performing open set recognition on the trained convolutional neural network by utilizing the preprocessed test sample set to obtain a radar high-resolution range profile open set recognition result based on the convolutional neural network;
wherein the model construction module (3) is specifically configured to:
Constructing a convolutional neural network model with a four-layer structure; the four-layer structure is respectively a first layer of convolution layer, a second layer of convolution layer, a third layer of convolution layer and a fourth layer of full-connection layer, and each convolution layer is provided with the same convolution step length; each convolution layer comprises a plurality of convolution kernels, and the sizes of the convolution kernels are the same;
constructing a loss function of a convolutional neural network, wherein the expression is as follows:
Wherein Θ (x) is the output result of the convolutional neural network, O i (i=1, …, N) is N prototypes randomly initialized according to gaussian distribution, d (Θ (x), O k) is the euclidean distance from Θ (x) to O k, λ is a hyper-parameter, r i＝d(Oi,Oc), representing each prototype O i to the center Is a distance of (2);
the training module (4) is specifically configured to:
Randomly dividing the preprocessed training sample set into q batches, wherein the data of each batch is n multiplied by D dimensional matrix data; where n=floor (P/q), floor () represents a rounding down, and P represents the number of high-resolution range profiles in the training sample set;
sequentially inputting the data of each batch into a convolutional neural network for processing to obtain an output result of the convolutional neural network;
Calculating the value of a loss function according to the output result of the convolutional neural network, and updating the parameter value of the convolutional neural network by using a random gradient method until the network converges to obtain a trained convolutional neural network;
the object recognition module (5) is specifically configured to:
Let the probability that the convolutional neural network predicts that the sample x to be measured belongs to the category k be The expression is as follows:
Probability when a sample is to be tested When the sample to be detected is smaller than a preset threshold value, judging the sample to be detected as an out-of-library unknown class; otherwise, the class is judged as a known class in the library, and the specific class to which the class belongs is further obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111199838.4A CN114137518B (en) | 2021-10-14 | 2021-10-14 | Radar high-resolution range profile open set identification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111199838.4A CN114137518B (en) | 2021-10-14 | 2021-10-14 | Radar high-resolution range profile open set identification method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114137518A CN114137518A (en) | 2022-03-04 |
| CN114137518B true CN114137518B (en) | 2024-07-12 |
Family
ID=80395049
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111199838.4A Active CN114137518B (en) | 2021-10-14 | 2021-10-14 | Radar high-resolution range profile open set identification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114137518B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116089821B (en) * | 2023-02-23 | 2023-08-15 | 中国人民解放军63921部队 | Method for monitoring and identifying state of deep space probe based on convolutional neural network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
| AU2020104006A4 (en) * | 2020-12-10 | 2021-02-18 | Naval Aviation University | Radar target recognition method based on feature pyramid lightweight convolutional neural network |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107728143B (en) * | 2017-09-18 | 2021-01-19 | 西安电子科技大学 | Radar high-resolution range profile target identification method based on one-dimensional convolutional neural network |
| CN108520199B (en) * | 2018-03-04 | 2022-04-08 | 天津大学 | Human body action open set identification method based on radar image and generation countermeasure model |
| CN109376574B (en) * | 2018-08-14 | 2021-05-14 | 西安电子科技大学 | CNN-based (probabilistic neural network-based) HRRP (high-resolution Radar) target identification method for radar capable of refusing judgment |
| CN111458688B (en) * | 2020-03-13 | 2024-01-23 | 西安电子科技大学 | Three-dimensional convolution network-based radar high-resolution range profile target recognition method |
| CN112904299B (en) * | 2021-03-03 | 2023-06-30 | 西安电子科技大学 | Radar high-resolution range profile open set target identification method based on deep class segmentation |
-
2021
- 2021-10-14 CN CN202111199838.4A patent/CN114137518B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
| AU2020104006A4 (en) * | 2020-12-10 | 2021-02-18 | Naval Aviation University | Radar target recognition method based on feature pyramid lightweight convolutional neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114137518A (en) | 2022-03-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Shao et al. | Convolutional neural network-based radar jamming signal classification with sufficient and limited samples | |
| CN110516561B (en) | SAR image target identification method based on DCGAN and CNN | |
| CN111913156B (en) | Radar radiation source individual identification method based on deep learning model and feature combination | |
| CN111352086B (en) | Unknown target identification method based on deep convolutional neural network | |
| CN108256436B (en) | Radar HRRP target identification method based on joint classification | |
| CN104408469A (en) | Firework identification method and firework identification system based on deep learning of image | |
| CN112598046B (en) | Target tactical intent recognition method in multi-machine cooperative air combat | |
| CN112904299B (en) | Radar high-resolution range profile open set target identification method based on deep class segmentation | |
| CN104732244A (en) | Wavelet transform, multi-strategy PSO (particle swarm optimization) and SVM (support vector machine) integrated based remote sensing image classification method | |
| CN110109110B (en) | HRRP target identification method based on priori optimal variation self-encoder | |
| CN110703221A (en) | Urban low-altitude small target classification and identification system based on polarization characteristics | |
| CN111273288B (en) | Radar unknown target identification method based on long-term and short-term memory network | |
| CN114137518B (en) | Radar high-resolution range profile open set identification method and device | |
| Xiao et al. | Specific emitter identification of radar based on one dimensional convolution neural network | |
| CN111368653B (en) | Low-altitude small target detection method based on R-D graph and deep neural network | |
| Vinholi et al. | CNN-based change detection algorithm for wavelength-resolution SAR images | |
| Souza et al. | Automatic detection of ceratocystis wilt in eucalyptus crops from aerial images | |
| CN113887583A (en) | Radar RD image target detection method based on deep learning under low signal-to-noise ratio | |
| CN114821335B (en) | Unknown target discrimination method based on fusion of depth features and linear discrimination features | |
| CN116311067A (en) | Target comprehensive identification method, device and equipment based on high-dimensional characteristic map | |
| CN113469084B (en) | Hyperspectral image classification method based on contrast generation countermeasure network | |
| Tan et al. | Real-time cloud detection in high resolution images using maximum response filter and principle component analysis | |
| CN114296067A (en) | Pulse Doppler radar low-slow small target identification method based on LSTM model | |
| CN115861676A (en) | Extreme value distribution-based radar high-resolution range profile open-set identification method | |
| Zhang et al. | Rapid recognition of tomato’s Disease stages based on the kernel mutual subspace method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant |