CN111736125B - Radar target identification method based on attention mechanism and bidirectional stacking cyclic neural network - Google Patents
Radar target identification method based on attention mechanism and bidirectional stacking cyclic neural network Download PDFInfo
- Publication number
- CN111736125B CN111736125B CN202010256158.0A CN202010256158A CN111736125B CN 111736125 B CN111736125 B CN 111736125B CN 202010256158 A CN202010256158 A CN 202010256158A CN 111736125 B CN111736125 B CN 111736125B
- Authority
- CN
- China
- Prior art keywords
- hrrp
- sample
- rnn
- layer
- radar
- 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 50
- 230000002457 bidirectional effect Effects 0.000 title claims abstract description 42
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 32
- 230000007246 mechanism Effects 0.000 title claims abstract description 23
- 125000004122 cyclic group Chemical group 0.000 title claims abstract description 13
- 230000011218 segmentation Effects 0.000 claims abstract description 23
- 230000035945 sensitivity Effects 0.000 claims abstract description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims description 32
- 230000006870 function Effects 0.000 claims description 22
- 238000012360 testing method Methods 0.000 claims description 21
- 230000005484 gravity Effects 0.000 claims description 10
- 238000010606 normalization Methods 0.000 claims description 10
- 230000004913 activation Effects 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 7
- 230000005284 excitation Effects 0.000 claims description 6
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 230000000306 recurrent effect Effects 0.000 claims description 6
- 238000013519 translation Methods 0.000 claims description 6
- 238000005520 cutting process Methods 0.000 claims description 4
- 238000002592 echocardiography Methods 0.000 claims description 4
- 230000000694 effects Effects 0.000 claims description 4
- 210000005036 nerve Anatomy 0.000 claims description 4
- 230000006835 compression Effects 0.000 claims description 3
- 238000007906 compression Methods 0.000 claims description 3
- 230000002401 inhibitory effect Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 2
- 238000005457 optimization Methods 0.000 claims description 2
- 238000000605 extraction Methods 0.000 description 10
- 238000013135 deep learning Methods 0.000 description 5
- 230000002708 enhancing effect Effects 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 230000009466 transformation Effects 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000001627 detrimental effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
Images
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
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N3/045—Combinations of networks
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Remote Sensing (AREA)
- Radar, Positioning & Navigation (AREA)
- Computer Networks & Wireless Communication (AREA)
- Radar Systems Or Details Thereof (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a radar target identification method based on an attention mechanism and a bidirectional stacking cyclic neural network, which comprises the steps of firstly preprocessing to reduce sensitivity in an HRRP sample and establishing a dynamic adjustment layer; then selecting the sliding window size to split the HRRP, wherein the sliding window moving distance is smaller than the sliding window length; then, the importance degree of each segmentation sequence is adjusted through an importance network; modeling the time sequence correlation of the samples through the two-way stacking RNNs, and extracting the high-level characteristics of the samples; finally, a multi-level attention mechanism is adopted to adjust the importance degree of the hidden layer state and target classification is carried out through softmax.
Description
Technical Field
The invention belongs to the field of radar target identification, and particularly relates to a radar target identification method based on an attention mechanism and a bidirectional stacking cyclic neural network.
Background
The range resolution of high-resolution wideband radar is much smaller than the target size, and its echoes are also known as one-dimensional high-resolution range profiles (High Resolution Range Profile, HRRP) of the target. The HRRP contains extremely valuable structural information for classification, identification and the like, such as the radial size of a target, scattering point distribution and the like, and has wide engineering application prospect. Therefore, the HRRP-based radar automatic target recognition method gradually becomes a hot spot for research in the radar automatic target recognition field.
For most HRRP target recognition systems, since the original HRRP sample tends to have a high dimension, it is often difficult to directly embody the essential attribute of the recognition object, so feature extraction is a key ring. The main task of feature extraction is to provide some help (such as reducing data dimension, enhancing discrimination information, etc.) for the subsequent recognition task through some linear or nonlinear transformation. The effective characteristics not only can fully express data, but also can distinguish the differences of different categories, thereby improving the recognition precision.
The conventional feature extraction method can be divided into two parts: (1) feature extraction method based on dimension reduction; (2) Transformation (transducer) based feature extraction methods such as bispectrum, spectrogram, spectral amplitude features, and the like. These methods project HRRP signals into the frequency domain and then model and identify their frequency domain features. The traditional feature extraction method obtains good recognition performance in experiments, however, the following two problems exist: (1) The manner in which features are extracted is mostly unsupervised and lossy, meaning that the information of the separability will inevitably lose a part during the feature extraction, which is detrimental to the recognition of the back-end classifier. (2) The choice of feature extraction methods is highly dependent on the accumulation of knowledge and experience of researchers with HRRP data, and is difficult to achieve in some cases without prior information.
In order to solve the problems of the conventional method in feature extraction, in recent years, a deep learning-based method has been introduced into the field of radar target recognition. The radar high-resolution range profile recognition method based on deep learning can be roughly divided into the following three types: (1) A method of deep learning based on encoder-decoder architecture. (2) Deep learning method based on Convolutional Neural Network (CNN) structure. (3) a deep learning method based on a cyclic neural network. However, methods (1) and (2) feature extraction and modeling directly on the envelope information of the HRRP ensemble ignores sequence dependencies between HRRP distance units that may reflect the target physical structural features. The method (3) models based on sequence correlation, and although modeling and describing physical structural features, there are several problems: (1) The range bin with smaller amplitude may contain some very separable features, but these features are rarely used; (2) The unidirectional RNN can only utilize the current time and the structural information before the current time during prediction, and cannot well utilize the prior integral structural information contained in the HRRP.
Disclosure of Invention
In view of the above technical problems, the present invention is to provide a radar target recognition method based on an attention mechanism and a bidirectional stacking cyclic neural network, which includes firstly, preprocessing to reduce sensitivity in an HRRP sample and establishing a dynamic adjustment layer; then selecting the sliding window size to split the HRRP, wherein the sliding window moving distance is smaller than the sliding window length; then, the importance degree of each segmentation sequence is adjusted through an importance network; modeling the time sequence correlation of the samples through the two-way stacking RNNs, and extracting the high-level characteristics of the samples; finally, a multi-level attention mechanism is adopted to adjust the importance degree of the hidden layer state and target classification is carried out through softmax.
In order to solve the technical problems, the invention adopts the following technical scheme:
a radar target identification method based on an attention mechanism and a bidirectional stacking cyclic neural network comprises the following steps:
s1, collecting a data set, combining HRRP data sets collected by a radar according to the types of targets, respectively selecting a training sample and a test sample in different data segments by each type of sample, ensuring that the posture formed by the selected training set sample and the radar covers the posture formed by the test set sample and the radar in the process of selecting the training set sample and the test set sample, wherein the ratio of the number of samples of each type of targets to the number of samples of each type of targets is 8:2, and marking the selected data set as T = { (x) i ,y k )} i∈[1,n],k∈[1,c] Wherein x is i Represents the i-th sample, y k Indicating that the samples belong to the kth class, a class c target is collected altogether, and n indicates the total number of the samples;
s2, preprocessing an original HRRP sample set, wherein the strength of the HRRP comprises radar transmitting power, target distance, radar antenna gain and radar receiver gain factors which are determined together, and before target identification is carried out by using the HRRP, the method comprises the steps of 2 The original HRRP echo is processed by an intensity normalization method, so that the intensity sensitivity problem of the HRRP is improved, the HRRP is intercepted from radar echo data through a distance window, the position of a range image recorded in the intercepting process in a range gate is not fixed, and therefore translation sensitivity of the HRRP is caused, and the HRRP is subjected to gravity center pairThe alignment method improves the translation sensitivity problem of HRRP;
s3, because the amplitude difference of echoes in each distance unit in the HRRP is large, the data is directly sent into the convolution layer, so that the model is excessively focused on the distance unit with large amplitude, however, the distance unit with small amplitude possibly contains some characteristics with strong separability, radar target identification is facilitated, a dynamic adjusting layer is added to carry out integral dynamic range adjustment on the HRRP before the HRRP is segmented, and the adjusting layer can determine how to adjust the integral dynamic of the HRRP through model training on the premise that the relative relation of the sizes of the distance units is not changed, so that a better identification effect is achieved;
s4, selecting a sliding window with a fixed length to segment the HRRP sample processed by the method, wherein the segmented data format is the input format of a subsequent deep neural network;
s5, constructing an importance adjustment network to carry out channel adjustment on the processed data, automatically acquiring the importance degree of each characteristic channel by the importance network in a learning mode, and then improving useful characteristics and inhibiting characteristics with little use for the current task according to the importance degree;
s6, setting up deep nerve classification, adjusting parameters and optimizing, adopting a bidirectional cyclic nerve network, respectively inputting HRRP data into two independent RNN models in a positive and negative direction, and splicing the obtained hidden layers;
s7, performing preprocessing operations of steps S2, S3 and S4 of a training stage on the test data acquired by the step S1;
s8, sending the sample processed in the S7 into the model constructed in the S6 for testing to obtain a result, namely, finally classifying the output through the attention mechanism through a softmax layer, and testing the sample by the ith HRRPThe probability corresponding to a class k radar target in the target set may be calculated as:
wherein exp (·) represents the exponentiation, and c represents the number of categories.
Preferably, the step S2 further comprises the steps of:
s201, intensity normalization, assuming that the original HRRP is denoted as x raw =[x 1 ,x 2 ,…,x L ]Where L represents the total number of distance units contained within the HRRP, then the HRRP after intensity normalization is expressed as:
s202, aligning samples, translating the HRRP so that the center of gravity g of the HRRP moves to the vicinity of L/2, and thus, the distance units containing information in the HRRP are distributed in the vicinity of the center, wherein the calculation method of the center of gravity g of the HRRP is as follows:
wherein x is i Is the i-th dimension signal unit in the original HRRP.
Preferably, the S3 further includes: the HRRP sample is dynamically adjusted, namely, the sample is subjected to multiple powers, the data is subjected to the powers, so that the diversity of target class differences is reflected from multiple angles, the information contained in the radar HRRP is reflected in multiple different forms from multiple angles, the characteristics are conveniently extracted from multiple angles by a subsequent network for identification, and the output of a dynamic adjustment layer can be expressed as follows:
wherein M is the channel number of the dynamic adjustment layer, and the ith dynamic adjustment channelCan be expressed as:
wherein alpha is i The coefficients representing the power transform.
Preferably, the S4 further includes:
s401, carrying out sliding window segmentation on the HRRP sample subjected to dynamic adjustment, setting the length of the sliding window as N, and the sliding distance as d, wherein d is less than N, namely, two adjacent sections of signals after the segmentation have overlapping parts with the length of N-d, the overlapping segmentation is larger, the sequence characteristics in the HRRP sample are reserved, the subsequent deep neural network can learn the characteristics more useful for classification in the sample, the number of times of the segmentation corresponds to the dimension of a time point in the input format of the subsequent deep neural network, and the length N of the sliding window corresponds to the dimension of an input signal of each time point;
s402, the output after the sliding window segmentation may be expressed as:
wherein M is the number of sequences after segmentation, and the t-th segmentation sequence is
Wherein d is the sliding distance of the window, and N is the sliding window length.
Preferably, the S5 further includes:
s501, importance network carries on importance adjustment to the HRRP after segmentation, through learning the overall information of the convolution channel to emphasize the input sequence of some time points with more separable information selectively and restrain some other less important input sequences of time points, after the importance network adjustment, the model becomes more balanced, make more important, more useful characteristic can be highlighted, has improved model and represented HRRP ability, the importance adjustment is divided into compression characteristic and excitation characteristic two parts;
s502, compressing the feature part: the sample after sliding window cutting is The feature is composed of M sequences, each sequence is an N-dimensional vector, each sequence is compressed into a real weight x representing the importance of the sequence through a full connection layer and an activation function sq ,x slide The output through the full connection can be calculated by the following formula:
x sq =f(Wx slide +b)
s503, feature excitation section: selectively adjusting the extracted features through an expression formula to obtain adjusted features F E :
F E =x slide ⊙x sq
Wherein x is sq =[x sq (1),x sq (2),…,x sq (M)]It is an M-dimensional vector, +. slide Each element in each channel is multiplied by x sq The number in the corresponding dimension in this vector. As in feature F E The mth channel of (a) is adjusted to:
preferably, specifically, the step S6 further includes:
s601, the classification network is designed as a multi-layer stacked bidirectional RNN, assuming that the input is feature F RNN ,Wherein M is i Each time point dimension representing the ith bidirectional RNN, N representing the input sequence length, assuming its output is F output ,/>Where H is the number of hidden units, where the vector corresponding to the kth time point in the sequence can be expressed as:
wherein f (·) represents the activation function,representing a hidden layer output matrix corresponding to a forward RNN included in an ith bidirectional RNN,/and>represents the kth hidden layer state contained in the forward RNN contained in the ith bidirectional RNN, and similarly,/is>Represents a hidden layer output matrix corresponding to a backward RNN included in an ith bidirectional RNN,/v>Represents the kth hidden layer state, b, contained in the backward RNN contained in the ith bidirectional RNN Fi An output layer bias representing the ith bidirectional RNN;
s602, selecting hidden layers obtained by the last two-way RNNs at different moments for splicing according to an attention mechanism in a network, wherein the hidden layer state after the i-th layer is spliced is as follows:
finally, adding hidden layers after each layer is spliced to obtain hidden layer c after attention model processing ATT The method comprises the following steps:
wherein alpha is ik Representing the weight corresponding to the kth time point of the ith layer, M representing the number of hidden states contained in the forward RNN or the backward RNN of each layer in the bidirectional RNN model, namely the dimension of the time point, N 1 Indicating the number of layers of the network stack, N 0 Representing taking hidden states in several layers stacked bidirectional RNNs for c-solving, starting from the last layer ATT ,α ik The method of (2) is as follows:
wherein e ik The energy added for the forward and backward hidden states in the ith bidirectional RNN is expressed as:
e ik =U ATT tanh(W ATT h ik )
wherein the method comprises the steps ofThey are parameters for calculating the energy of the hidden units, l is the dimension of the hidden units, M is the dimension of the point in time;
s603, designing a loss function as cross entropy, learning parameters by calculating gradients of the loss function relative to the parameters by using training data, and fixing the learned parameters when the model converges, wherein the cost function based on the cross entropy is adopted and expressed as:
wherein N represents the number of training samples in a batch, e n Is a one-hot vector forTrue labels representing the nth training sample, P (i|x train ) Representing the probability that the training sample corresponds to the ith target.
The invention has the following beneficial effects:
(1) The dynamic adjustment layer is applied in the embodiment of the invention, and because some better separable characteristics possibly have difficulty in influencing the decision of the subsequent classifier due to the relative amplitude, the overall dynamic state of the HRRP is determined to be adjusted by model training on the premise that the relative relation of the sizes of the distance units is not changed through the dynamic adjustment layer, so that a better recognition effect is achieved.
(2) The embodiment of the invention applies an importance adjustment network, which can selectively emphasize the convolution channels with more separable information and inhibit the less useful convolution channels by learning the global information of the convolution channels. After adjustment, the model becomes more balanced from the spatial channel (convolution channel), so that more important and useful features can be highlighted, and the HRRP representation capability of the model is improved.
(3) The embodiment of the invention is different from the existing model built based on the HRRP structure, and the bidirectional circulating neural network is stacked and used in the embodiment of the invention, so that the model has a certain depth, the model organized in the mode can abstract the structural characteristics of a high layer step by step according to the context of data better, and hidden states in each bidirectional circulating neural network layer contain structural representations of different layers, thereby helping us to better apply the HRRP for recognition.
(4) The embodiment of the invention applies the attention model, and the weight for enhancing the judgment given by the middle signal aggregation area is considered in classification, so that the weight for giving the judgment by the noise areas at two sides is reduced.
Drawings
Fig. 1 is a flowchart of the steps of a radar target recognition method based on an attention mechanism and a bi-directional stacked recurrent neural network according to an embodiment of the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1, a flowchart of steps of a radar target recognition method based on an attention mechanism and a bi-directional stacked recurrent neural network according to an embodiment of the present invention is shown, specifically, the method includes the following steps:
s1, collecting a data set, combining HRRP data sets collected by a radar according to the types of targets, respectively selecting a training sample and a test sample in different data segments by each type of sample, ensuring that the posture formed by the selected training set sample and the radar covers the posture formed by the test set sample and the radar in the process of selecting the training set sample and the test set sample, wherein the ratio of the number of samples of each type of targets to the number of samples of each type of targets is 8:2, and marking the selected data set as T = { (x) i ,y k )} i∈[1,n],k∈[1,c] Wherein x is i Represents the i-th sample, y k Indicating that the samples belong to the kth class, a class c target is collected altogether, and n indicates the total number of the samples;
s2, preprocessing an original HRRP sample set, wherein the strength of the HRRP comprises radar transmitting power, target distance, radar antenna gain and radar receiver gain factors which are determined together, and before target identification is carried out by using the HRRP, the method comprises the steps of 2 The original HRRP echo is processed by the intensity normalization method, so that the intensity sensitivity problem of the HRRP is improved, the HRRP is intercepted from radar echo data through a distance window, the position of a distance image recorded in the intercepting process in a distance wave gate is not fixed, and therefore the translation sensitivity of the HRRP is caused, and the translation sensitivity problem of the HRRP is improved through a gravity center alignment method;
specifically, S2 further comprises the steps of:
s201, intensity normalization, assuming that the original HRRP is denoted as x raw =[x 1 ,x 2 ,…,x L ]Where L represents the total number of distance units contained within the HRRP, the HRRP after intensity normalization can be expressed as:
s202, aligning samples. Translating the HRRP moves its center of gravity g to around L/2 so that those range cells in the HRRP that contain information will be distributed around the center. The method for calculating the HRRP gravity center g comprises the following steps:
wherein x is i Is the i-th dimension signal unit in the original HRRP.
After the original HRRP sample is processed by an intensity normalization and gravity center alignment method, the amplitude is limited to be between 0 and 1, the scale is unified, and the value between 0 and 1 is very favorable for the subsequent neural network processing; HRRP echo signals with right or left distribution are adjusted to near the center point.
S3, because the amplitude difference of echoes in each distance unit in the HRRP is large, the data is directly sent into the convolution layer, so that the model is excessively focused on the distance unit with large amplitude, however, the distance unit with small amplitude possibly contains some characteristics with strong separability, radar target identification is facilitated, a dynamic adjusting layer is added to carry out integral dynamic range adjustment on the HRRP before the HRRP is segmented, and the adjusting layer can determine how to adjust the integral dynamic of the HRRP through model training on the premise that the relative relation of the sizes of the distance units is not changed, so that a better identification effect is achieved;
the step S3 further includes dynamically adjusting the HRRP sample, that is, performing a plurality of powers on the sample, performing a power processing on the data, so as to reflect diversity of target class differences from a plurality of angles, and reflecting information contained in the radar HRRP from a plurality of angles in a plurality of different forms, so that a subsequent network can conveniently extract features from a plurality of angles to identify, and the output of the dynamic adjustment layer can be expressed as:
wherein M is the channel number of the dynamic adjustment layer, and the ith dynamic adjustment channelCan be expressed as
Wherein alpha is i The coefficients representing the power transform.
S4, selecting a sliding window with a fixed length to segment the HRRP sample processed by the method, wherein the segmented data format is the input format of a subsequent deep neural network;
the S4 further includes:
s401, carrying out sliding window segmentation on the HRRP sample subjected to dynamic adjustment, setting the length of the sliding window as N, and the sliding distance as d, wherein d is less than N, namely, two adjacent sections of signals after the segmentation have overlapping parts with the length of N-d, the overlapping segmentation is larger, the sequence characteristics in the HRRP sample are reserved, the subsequent deep neural network can learn the characteristics more useful for classification in the sample, the number of times of the segmentation corresponds to the dimension of a time point in the input format of the subsequent deep neural network, and the length N of the sliding window corresponds to the dimension of an input signal of each time point;
s402, the output after the sliding window segmentation may be expressed as:
wherein M is the number of sequences after segmentation, and the t-th segmentation sequence is
Wherein d is the sliding distance of the window, and N is the sliding window length.
S5, constructing an importance adjustment network to carry out channel adjustment on the processed data, automatically acquiring the importance degree of each characteristic channel by the importance network in a learning mode, and then improving useful characteristics and inhibiting characteristics with little use for the current task according to the importance degree;
specifically, the S5 further includes:
s501, importance network carries on importance adjustment to the HRRP after segmentation, through learning the overall information of the convolution channel to emphasize the input sequence of some time points with more separable information selectively and restrain some other less important input sequences of time points, after the importance network adjustment, the model becomes more balanced, make more important, more useful characteristic can be highlighted, has improved model and represented HRRP ability, the importance adjustment is divided into compression characteristic and excitation characteristic two parts;
s502, compressing the feature part: the sample after sliding window cutting is The feature is composed of M sequences, each sequence is an N-dimensional vector, each sequence is compressed into a real weight x representing the importance of the sequence through a full connection layer and an activation function sq ,x slide The output through the full connection can be calculated by the following formula:
x sq =f(Wx slide +b)
s503, feature excitation section: selectively adjusting the extracted features through an expression formula to obtain adjusted features F E :
F E =x slide ⊙x sq
Wherein x is sq =[x sq (1),x sq (2),…,x sq (M)]It is an M-dimensional vector, +. slide Each element in each channel is multiplied by x sq The number in the corresponding dimension in this vector. As in feature F E The mth channel of (a) is adjusted to:
s6, deep nerve classification is built, parameter adjustment is carried out, optimization is carried out, a bidirectional cyclic neural network is adopted, HRRP data are respectively input into two independent RNN models in a positive and negative direction, and the obtained hidden layers are spliced.
The traditional RNN model is unidirectional, HRRP data can only be input along one direction when being input into the traditional model, so that the input at the current moment only has conditional dependence with the input data before the HRRP data, and the input information at the later moment cannot be effectively applied at the current moment. However, the HRRP contains the physical structure prior of the whole object, and only one-way information is considered to be unfavorable for modeling and identifying the HRRP characteristics. In particular, when the unidirectional RNN is applied, most of the observed data information is noise data when the time t is small, and it is difficult for the RNN to accurately model the target structural characteristics. Therefore, the embodiment of the invention adopts a bidirectional cyclic neural network, which inputs HRRP data into two independent RNN models respectively in a positive and negative direction and splices the obtained hidden layers, thus improving the defects of unidirectional RNNs and modeling the physical structural characteristics contained in the HRRP better. The embodiment of the invention uses a stacked bidirectional circulating neural network to enable the model to have a certain depth. The model organized in the mode can better abstract the structural characteristics of a high layer step by step depending on the context of data, and hidden states in each bidirectional circulating neural network layer contain structural representations of different layers, so that the model can help to better apply HRRP for identification. And applying the attention model on the basis, namely considering the weight for enhancing the judgment given by the middle signal aggregation area in classification, and reducing the weight for giving the judgment to the noise areas at two sides. Namely, the deep neural network model in the embodiment of the invention is formed by stacking five layers of bidirectional LSTM (long short time memory network) with attention mechanisms, and finally, the softmax layer is adopted to classify the output of the network.
In particular, the method comprises the steps of, the S6 further includes:
s601, assume that the input is feature F RNN ,Wherein M is i Each time point dimension representing the ith bidirectional RNN, N representing the input sequence length, assuming its output is F output ,Where H is the number of hidden units, where the vector corresponding to the kth time point in the sequence can be expressed as:
wherein f (·) represents the activation function,representing a hidden layer output matrix corresponding to a forward RNN included in an ith bidirectional RNN,/and>represents the kth hidden layer state contained in the forward RNN contained in the ith bidirectional RNN, and similarly,/is>Represents a hidden layer output matrix corresponding to a backward RNN included in an ith bidirectional RNN,/v>Represents the kth hidden layer state, b, contained in the backward RNN contained in the ith bidirectional RNN Fi Representing the output layer bias of the ith bidirectional RNN.
S602, selecting hidden layers obtained by the last two-way RNNs at different moments to splice, wherein the hidden layer state after the ith layer is spliced is as follows:
finally, adding hidden layers after each layer is spliced to obtain hidden layer c after attention model processing ATT The method comprises the following steps:
wherein alpha is ik Representing the weight corresponding to the kth time point of the ith layer, M representing the number of hidden states contained in the forward RNN or the backward RNN of each layer in the bidirectional RNN model, namely the dimension of the time point, N 1 Indicating the number of layers of the network stack, N 0 Representing taking hidden states in several layers stacked bidirectional RNNs for c-solving, starting from the last layer ATT 。α ik The method of (2) is as follows:
wherein e ik The energy added for the forward and backward hidden states in the ith bidirectional RNN can be expressed as:
e ik =U ATT tanh(W ATT h ik )
wherein the method comprises the steps ofThey are parameters for calculating the energy of the hidden units, l is the dimension of the hidden units, and M is the dimension of the point in time.
S603, performing splicing operation on the output of the attention mechanism, and then connecting a full-connection layer with the node number being the radar class number, namely, the output of the full-connection layer is a prediction result of a model, and the output can be expressed as:
output=f(C(c ATT )W o )
wherein C (& gt) is a splicing operation,c represents the number of categories and f (·) represents the softmax function.
S604, designing a loss function as cross entropy. The parameters are learned by calculating gradients of the loss function relative to the parameters using the training data, and the learned parameters are fixed as the model converges. The invention adopts a cost function based on cross entropy, and can be expressed as follows:
wherein N represents the number of training samples in a batch, e n Is one-hot vector representing the true label of the nth training sample, P (i|x train ) Representing the probability that the training sample corresponds to the ith target.
S605, initializing all weights and biases to be trained in the model, setting training parameters including learning rate, training data amount of each batch, training batch, and starting model training.
S7, performing preprocessing operations of steps S2, S3 and S4 of a training stage on the test data acquired by the step S1;
s8, sending the sample processed in the S7 into the model constructed in the S6 for testing to obtain a result, namely, finally classifying the output through the attention mechanism through a softmax layer, and testing the sample by the ith HRRPThe probability corresponding to a class k radar target in the target set may be calculated as:
wherein exp (·) represents the exponentiation, and c represents the number of categories.
Test HRRP sample x by maximum posterior probability test K classified to maximum target probability 0 In (a):
through the 8 steps, the radar target recognition model based on the attention mechanism and the bidirectional stacking cyclic neural network provided by the invention can be obtained.
It should be understood that the exemplary embodiments described herein are illustrative and not limiting. Although one or more embodiments of the present invention have been described with reference to the accompanying drawings, it will be understood by those of ordinary skill in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present invention as defined by the following claims.
Claims (5)
1. The radar target identification method based on the attention mechanism and the bidirectional stacking cyclic neural network is characterized by comprising the following steps of:
s1, collecting a data set, combining HRRP data sets collected by a radar according to the types of targets, respectively selecting a training sample and a test sample in different data segments by each type of samples, ensuring that the posture formed by the selected training set sample and the radar covers the posture formed by the test set sample and the radar in the process of selecting the training set sample and the test set sample, wherein the ratio of the numbers of the various target training sets to the test set sample is 8:2, and marking the selected data set asWherein x is i1 Represents the i1 st sample, y k1 Indicating that the samples belong to the k1 class, a c class target is collected altogether, and n1 indicates the total number of samples;
s2, preprocessing an original HRRP sample set, wherein the strength of the HRRP comprises radar transmitting power, target distance, radar antenna gain and radar receiver gain factors which are determined together, and before target identification is carried out by using the HRRP, the method comprises the steps of 2 The original HRRP echo is processed by an intensity normalization method, so that the intensity sensitivity problem of the HRRP is improved; HRRP is intercepted from radar echo data through a distance window, and the position of a distance image recorded in the intercepting process is not fixed in a distance wave gate, so that the translation sensitivity of the HRRP is caused, and the translation sensitivity problem of the HRRP is improved through a gravity center alignment method;
s3, because the amplitude difference of echoes in each distance unit in the HRRP is large, the data is directly sent into the convolution layer, so that the model is excessively focused on the distance units with large amplitude, the distance units with small amplitude contain some characteristics with strong separability, radar target identification is facilitated, a dynamic adjustment layer is added to carry out integral dynamic range adjustment on the HRRP before the HRRP is segmented, and the adjustment layer is trained through the model on the premise that the relative relation of the sizes of the distance units is not changed so as to determine how to adjust the integral dynamic of the HRRP, so that a better identification effect is achieved;
s4, selecting a sliding window with a fixed length to segment the HRRP sample processed by the method, wherein the segmented data format is the input format of a subsequent deep neural network;
s5, constructing an importance adjustment network to carry out channel adjustment on the processed data, automatically acquiring the importance degree of each characteristic channel by the importance network in a learning mode, and then improving useful characteristics according to the importance degree and inhibiting the characteristics with little use for the current task;
s6, deep nerve classification is built, parameter adjustment is carried out, optimization is carried out, a bidirectional cyclic neural network is adopted, HRRP data are respectively input into two independent RNN models in a positive and negative direction, and the obtained hidden layers are spliced;
the S6 further includes:
s601, the classification network is designed as a multi-layer stacked bidirectional RNN, assuming that the input is feature F RNN ,Wherein M is i2 Each time point dimension representing the i2 nd bi-directional RNN, N2 representing the input sequence length, assuming its output as F output ,Wherein H is the number of hidden units, and the vector corresponding to the kth time point in the sequence is expressed as:
wherein f (·) represents the activation function,representing hidden layer output matrix corresponding to forward RNNs included in the i2 nd bi-directional RNN,/v>Represents the kth 2 hidden layer state contained in the forward RNN contained in the ith 2 bidirectional RNN, and similarly,/th hidden layer state contained in the forward RNN>Represents a hidden layer output matrix corresponding to a backward RNN included in the i2 nd bi-directional RNN,/for>Represents the kth 2 hidden layer state, b, contained in the backward RNN contained in the ith 2 bidirectional RNN Fi2 Output layer bias representing the i2 nd bidirectional RNNPlacing;
s602, selecting hidden layers obtained by the last two-way RNNs at different moments for splicing according to an attention mechanism in a network, wherein the hidden layer state after the i-th layer is spliced is as follows:
finally, adding hidden layers after each layer is spliced to obtain hidden layer c after attention model processing ATT The method comprises the following steps:
wherein alpha is ik Represents the weight corresponding to the kth time point of the ith layer, M 1 Representing the number of hidden states, i.e., the point dimension, N, contained by either the forward RNN or the backward RNN of each layer in the bidirectional RNN model 1 Indicating the number of layers of the network stack, N 0 Representing taking hidden states in several layers stacked bidirectional RNNs for c-solving, starting from the last layer ATT ,α i3k3 The method of (2) is as follows:
wherein e i3k3 The added energy for the forward and backward hidden states in the i3 rd bi-directional RNN, expressed as:
e i3k3 =U ATT tanh(W ATT h i3k3 )
wherein the method comprises the steps ofThey are parameters for calculating the energy of the hidden units, l is the dimension of the hidden units, M 1 Is the point-in-time dimension;
s603, designing a loss function as cross entropy, learning parameters by calculating gradients of the loss function relative to the parameters by using training data, and fixing the learned parameters when the model converges, wherein the cost function based on the cross entropy is adopted and expressed as:
wherein N is 3 Representing the number of training samples in a batch,is one-hot vector representing the true label of the nth training sample, P (i 3 |x train ) Indicating that the training sample corresponds to the ith 3 Probability of individual targets;
s7, performing preprocessing operations of steps S2, S3 and S4 of a training stage on the test data acquired by the step S1;
s8, sending the sample processed in S7 into the model constructed in S6 for testing to obtain a result, namely, finally classifying the output through the attention mechanism by a softmax layer, i 13 Individual HRRP test samplesCorresponds to the kth in the target set 4 The probability of radar-like targets is calculated as:
wherein exp (·) represents the exponentiation, and c represents the number of categories.
2. The method for radar target identification based on an attention mechanism and a bi-directional stacked recurrent neural network as claimed in claim 1, wherein said S2 further comprises the steps of:
s201, intensity normalization, assuming that the original HRRP is denoted as x raw =[x 1 ,x 2 ,…,x i ,…,x L ]Where L represents the total number of distance units contained within the HRRP, then the HRRP after intensity normalization is expressed as:
s202, aligning samples, translating the HRRP so that the center of gravity g of the HRRP moves to the vicinity of L/2, and thus, the distance units containing information in the HRRP are distributed in the vicinity of the center, wherein the calculation method of the center of gravity g of the HRRP is as follows:
3. The method for radar target identification based on an attention mechanism and a bi-directional stacked recurrent neural network of claim 1, wherein S3 further comprises: the HRRP sample is dynamically adjusted, namely the sample is subjected to multiple power processing, the data is subjected to power processing, so that diversity of target class differences is reflected from multiple angles, information contained in the radar HRRP is reflected in multiple different forms from multiple angles, characteristics are conveniently extracted from multiple angles by a subsequent network for identification, and the output of a dynamic adjustment layer is expressed as follows:
wherein M is 5 Is the channel number of the dynamic adjustment layer, ith 5 Dynamic adjustment channelsExpressed as:
4. The method for radar target identification based on an attention mechanism and a bi-directional stacked recurrent neural network as claimed in claim 3, wherein said S4 further comprises:
s401, sliding window segmentation is carried out on the HRRP sample after dynamic adjustment, and the length of the sliding window is set to be N 4 The sliding distance is d, wherein d < N 4 I.e. adjacent two signals after cutting have a length N 4 The overlapping part of d is overlapped and segmented to be larger, the sequence characteristics in the HRRP sample are reserved, the subsequent deep neural network can learn the characteristics which are more useful for classification in the sample to be larger, wherein the number of the segments corresponds to the dimension of the time point in the input format of the subsequent deep neural network, and the length N of the sliding window 4 Corresponding to the input signal dimension at each time point;
s402, the output after the sliding window segmentation is expressed as:
wherein M is 6 The number of sequences after segmentation is that the t-th segmentation sequence is
Wherein d is window sliding distance, N 4 Is the sliding window length.
5. The method for radar target identification based on an attention mechanism and a bi-directional stacked recurrent neural network of claim 4, wherein S5 further comprises:
s501, importance network carries on importance adjustment to the HRRP after segmentation, through learning the overall information of the convolution channel to emphasize the input sequence of some time points with more separable information selectively and restrain some other less important input sequences of time points, after the importance network adjustment, the model becomes more balanced, make more important, more useful characteristic can be highlighted, has improved model and represented HRRP ability, the importance adjustment is divided into compression characteristic and excitation characteristic two parts;
s502, compressing the feature part: the sample after sliding window cutting isThe characteristic is composed of M 6 Each sequence is formed by N 6 Vector of dimensions, each sequence of which is compressed by the full-join layer and the activation function into a real weight x representing the importance of the sequence sq ,x slide The output through the full connection is calculated from the following formula:
x sq =f(Wx slide +b)
s503, feature excitation section: selectively adjusting the extracted features through an expression formula to obtain adjusted features F E :
F E =x slide ⊙x sq
Wherein x is sq =[x sq (1),x sq (2),…,x sq (M 7 )]It is an M 7 The dimension vector, ++indicates that X will be slide Each element in each channel is multiplied by x sq Numbers on corresponding dimensions in this vector, e.g. feature F E The mth channel of (a) is adjusted to:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010256158.0A CN111736125B (en) | 2020-04-02 | 2020-04-02 | Radar target identification method based on attention mechanism and bidirectional stacking cyclic neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010256158.0A CN111736125B (en) | 2020-04-02 | 2020-04-02 | Radar target identification method based on attention mechanism and bidirectional stacking cyclic neural network |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111736125A CN111736125A (en) | 2020-10-02 |
CN111736125B true CN111736125B (en) | 2023-07-07 |
Family
ID=72646547
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010256158.0A Active CN111736125B (en) | 2020-04-02 | 2020-04-02 | Radar target identification method based on attention mechanism and bidirectional stacking cyclic neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111736125B (en) |
Families Citing this family (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112782660B (en) * | 2020-12-29 | 2023-06-30 | 杭州电子科技大学 | Radar target recognition method based on Bert |
CN113238197B (en) * | 2020-12-29 | 2023-07-04 | 杭州电子科技大学 | Radar target identification and judgment method based on Bert and BiLSTM |
CN112764024B (en) * | 2020-12-29 | 2023-06-16 | 杭州电子科技大学 | Radar target identification method based on convolutional neural network and Bert |
CN112731309B (en) * | 2021-01-06 | 2022-09-02 | 哈尔滨工程大学 | Active interference identification method based on bilinear efficient neural network |
CN112986941B (en) * | 2021-02-08 | 2022-03-04 | 天津大学 | Radar target micro-motion feature extraction method |
CN113486917B (en) * | 2021-05-17 | 2023-06-02 | 西安电子科技大学 | Radar HRRP small sample target recognition method based on metric learning |
CN114509736B (en) * | 2022-01-19 | 2023-08-15 | 电子科技大学 | Radar target identification method based on ultra-wide band electromagnetic scattering characteristics |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20170262996A1 (en) * | 2016-03-11 | 2017-09-14 | Qualcomm Incorporated | Action localization in sequential data with attention proposals from a recurrent network |
CN109086700B (en) * | 2018-07-20 | 2021-08-13 | 杭州电子科技大学 | Radar one-dimensional range profile target identification method based on deep convolutional neural network |
CN109214452B (en) * | 2018-08-29 | 2020-06-23 | 杭州电子科技大学 | HRRP target identification method based on attention depth bidirectional cyclic neural network |
CN110334741B (en) * | 2019-06-06 | 2023-03-31 | 西安电子科技大学 | Radar one-dimensional range profile identification method based on cyclic neural network |
CN110418210B (en) * | 2019-07-12 | 2021-09-10 | 东南大学 | Video description generation method based on bidirectional cyclic neural network and depth output |
-
2020
- 2020-04-02 CN CN202010256158.0A patent/CN111736125B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN111736125A (en) | 2020-10-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111736125B (en) | Radar target identification method based on attention mechanism and bidirectional stacking cyclic neural network | |
CN109214452B (en) | HRRP target identification method based on attention depth bidirectional cyclic neural network | |
CN112364779B (en) | Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion | |
CN110443143B (en) | Multi-branch convolutional neural network fused remote sensing image scene classification method | |
CN110334741B (en) | Radar one-dimensional range profile identification method based on cyclic neural network | |
CN108922560B (en) | Urban noise identification method based on hybrid deep neural network model | |
CN110045015B (en) | Concrete structure internal defect detection method based on deep learning | |
CN111596276B (en) | Radar HRRP target identification method based on spectrogram transformation and attention mechanism circulating neural network | |
CN106951915B (en) | One-dimensional range profile multi-classifier fusion recognition method based on category confidence | |
EP3690741A2 (en) | Method for automatically evaluating labeling reliability of training images for use in deep learning network to analyze images, and reliability-evaluating device using the same | |
CN114332649B (en) | Cross-scene remote sensing image depth countermeasure migration method based on double-channel attention | |
CN112764024B (en) | Radar target identification method based on convolutional neural network and Bert | |
Lin et al. | Detection of gravitational waves using Bayesian neural networks | |
CN110751044A (en) | Urban noise identification method based on deep network migration characteristics and augmented self-coding | |
CN111596292B (en) | Radar target identification method based on importance network and bidirectional stacking cyclic neural network | |
CN110766084B (en) | Small sample SAR target identification method based on CAE and HL-CNN | |
CN113378796B (en) | Cervical cell full-section classification method based on context modeling | |
CN111580097A (en) | Radar target identification method based on single-layer bidirectional cyclic neural network | |
CN109239670B (en) | Radar HRRP (high resolution ratio) identification method based on structure embedding and deep neural network | |
CN109948722B (en) | Method for identifying space target | |
CN111580058A (en) | Radar HRRP target identification method based on multi-scale convolution neural network | |
CN115438708A (en) | Classification and identification method based on convolutional neural network and multi-mode fusion | |
CN117131436A (en) | Radiation source individual identification method oriented to open environment | |
CN116106880B (en) | Underwater sound source ranging method and device based on attention mechanism and multi-scale fusion | |
CN116486183A (en) | SAR image building area classification method based on multiple attention weight fusion characteristics |
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 | ||
GR01 | Patent grant |