CN113325380A - Air micro-motion target classification online library building method based on Mondrian forest - Google Patents
Air micro-motion target classification online library building method based on Mondrian forest Download PDFInfo
- Publication number
- CN113325380A CN113325380A CN202110589218.5A CN202110589218A CN113325380A CN 113325380 A CN113325380 A CN 113325380A CN 202110589218 A CN202110589218 A CN 202110589218A CN 113325380 A CN113325380 A CN 113325380A
- Authority
- CN
- China
- Prior art keywords
- training sample
- node
- sample set
- feature
- frequency domain
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 49
- 238000012549 training Methods 0.000 claims abstract description 199
- 238000012360 testing method Methods 0.000 claims abstract description 84
- 239000011159 matrix material Substances 0.000 claims abstract description 49
- 239000013598 vector Substances 0.000 claims description 56
- 238000010606 normalization Methods 0.000 claims description 32
- 238000005070 sampling Methods 0.000 claims description 14
- 238000004422 calculation algorithm Methods 0.000 claims description 10
- 238000000605 extraction Methods 0.000 claims description 6
- 239000000463 material Substances 0.000 abstract description 2
- 238000004088 simulation Methods 0.000 description 10
- 230000000694 effects Effects 0.000 description 6
- 238000007637 random forest analysis Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 229920002160 Celluloid Polymers 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000002592 echocardiography Methods 0.000 description 1
- 230000004083 survival 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Remote Sensing (AREA)
- Astronomy & Astrophysics (AREA)
- Multimedia (AREA)
- Computer Networks & Wireless Communication (AREA)
- Radar, Positioning & Navigation (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses an air micro-motion target classification online library building method based on a Mongolian forest, and mainly solves the problems that the classification performance of an RF (radio frequency) model cannot be rapidly improved by using a new sample set and manpower and material resources are wasted in the prior art. The implementation scheme is as follows: generating an initial training sample set, a newly added training sample set and a testing sample set, and respectively extracting the characteristics of the initial training sample set, the newly added training sample set and the testing sample set to obtain a characteristic matrix of the initial training sample set, a characteristic matrix of the newly added training sample set and a characteristic matrix of the testing sample set; inputting the characteristic matrix of the initial training sample set into an MF classifier for training to obtain a pre-training model; updating the pre-training model by utilizing the feature matrix of the newly added training sample set to obtain a new model; and inputting the characteristic matrix of the test sample set into the new model to obtain a classification result of the test sample set. According to the method, the MF model is efficiently updated by using the new sample set, the classification accuracy of the aerial aircraft target is improved, and the method can be used for target identification.
Description
Technical Field
The invention belongs to the technical field of radars, and further relates to an air micro-motion target classification online library building method which can be used for automatically updating an original MF model for classifying air plane targets in real time and improving the automation and intelligence levels of radars.
Background
Helicopter, propeller plane and jet plane micro-motion components are different in physical structure and motion parameters, and can generate different micro-Doppler modulation on radar echoes. The characteristics of the micro-Doppler modulation differences of the three types of airplanes are extracted and input into a designed classifier model for training, so that the air airplane target classification task can be completed. First, in the operation of a narrow-band radar, hundreds of pulses are accumulated for each sample collected, and thus the total number of samples obtained in a short time is small. Secondly, the airplane moves due to the constant change of the attitude and the distance, so that the samples are distributed in a nonlinear space. The random forest RF model is a set formed by a plurality of classification decision trees, and can effectively solve the problems of small samples and nonlinear classification. Therefore, RF models are often used for airborne aircraft object classification tasks. However, in actual work, it is difficult to obtain a sufficient and complete training database in the early stage, and the classification performance of the air plane target is limited. The existing method for classifying the target of the airplane in the air based on the micro-Doppler effect needs to combine a newly acquired sample and an original sample set into a new training sample set periodically and retrain an RF (radio frequency) model again, so that the model is updated. However, this method requires frequent manual intervention and cannot update the model in time when a new sample arrives. A more feasible idea is to use a classifier model with online learning capabilities. When a new sample comes, the existing model is automatically updated, and the classification accuracy of the model to the aerial aircraft target is improved.
The Li-valuable celluloid learner introduces an airplane target classification method based on the micro-Doppler effect in published 'airplane target classification method research based on the JEM effect'. The method comprises the following specific steps: the method comprises the steps of firstly, acquiring sufficient and complete radar echo target micro-motion signals as a training sample set; secondly, extracting features of each training sample; thirdly, training an RF model by using a characteristic matrix of a training sample set; fourthly, acquiring radar echo target micro-motion signals as a test sample set; fifthly, extracting the characteristics of each test sample; and sixthly, inputting the characteristic matrix of the test sample set into the trained RF model to obtain a classification result.
On one hand, the method directly utilizes the newly added training sample set to train the RF model, and can input the characteristic matrix of the sample set into the RF model for training when the newly added training sample set is collected, so that the existing model is quickly updated, and the learning efficiency is improved. On the other hand, the initial training sample set and the newly added training sample set are combined into a new training set, the feature matrix of the new training set is input into the RF model for training, and the learned RF model has good classification performance because the training set has relatively complete distribution information of the airplane target samples in the feature space. However, this method has two disadvantages: firstly, because the newly added training sample set contains fewer samples, the real distribution condition of the airplane target in the feature space cannot be reflected, if the RF model is trained by directly utilizing the newly added training sample set, the classification performance of the obtained RF model is poor, and the classification performance of the existing model cannot be improved; secondly, when enough new samples are collected, workers need to combine new and old sample sets periodically and then retrain the classification model, so that the learning efficiency is low and human and material resources are wasted.
Disclosure of Invention
The invention aims to provide an air micro-motion target classification online library building method based on a Mongolian forest aiming at the defects of the prior art, so as to avoid frequent manual intervention, save time and space resources, quickly update the existing model and improve the classification performance of the model.
The technical scheme for realizing the aim of the invention comprises the following steps:
establishing an initial training sample set containing micro-motion target radar echo signals, and performing feature extraction on the initial training sample set to obtain a feature matrix F of the initial training sample setO；
Establishing a newly added training sample set containing micro target radar echo signals, and performing feature extraction on the newly added training sample set to obtain a feature matrix F of the newly added training sample setI；
Establishing a test sample set containing a micro-motion target radar echo signal, and performing feature extraction on the test sample set to obtain a feature matrix F of the test sample setT；
A Mondrian trees are arranged in a Mondrian forest MF model, A is more than or equal to 10 and less than or equal to 50, and a feature matrix F of an initial training sample set is usedOInputting the information into the MF model, and training each Mondrian tree through a Mondrian tree generation algorithm until A Mondrian trees are trained to obtain a pre-training MF model;
feature matrix F using newly added training sample setIUpdating each trained Mondrian tree in the pre-trained MF model by a Mondrian tree expansion algorithm until A Mondrian trees are updated to obtain a new MF model;
feature matrix F of test sample setTAnd inputting the classification result into a new MF model to obtain a classification result of the test sample set.
Compared with the prior art, the method has the following advantages because the Mondrian forest algorithm is used:
first, the knowledge about new samples can be learned and stored by expanding the structure of each tree in the Mondriann forest, and the problem that the classification performance can not be improved by directly training a model by using a newly added sample set in the prior art is solved, so that the classification accuracy of the aerial aircraft target can be improved by updating the model.
And secondly, each Mondrian tree in the existing model can be updated and expanded by directly utilizing the newly added sample set, so that the method can automatically update the model base in real time, reduce the consumption of human resources and obtain the classification accuracy rate of the near-offline learning method on the aerial airplane target.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention;
FIG. 2 is a graph of the results of a simulation experiment of the present invention.
Detailed Description
Embodiments and effects of the present invention will be described in further detail below with reference to the accompanying drawings.
Referring to fig. 1, the specific implementation steps of this example include the following:
step 1, generating an initial training sample set, a newly added training sample set and a testing sample set.
1.1) P micro-motion target radar echo signals of D categories are obtained to serve as an initial training sample set, wherein D is larger than or equal to 3, and P is larger than or equal to 150;
1.2) acquiring P 'micro target radar echo signals of D categories as a newly added training sample set, wherein D is more than or equal to 3, and P' is more than or equal to 90;
1.3) obtaining Q micro-motion target radar echo signals of D categories as a test sample set, wherein D is larger than or equal to 3, and Q is larger than or equal to 1500.
Step 2, extracting the characteristics of the initial training sample set to obtain a characteristic matrix F of the P multiplied by 5 dimensional initial training sample setOAnd setting a label vector set L of the initial training sample setO。
2.1) carrying out fast Fourier transform on P echo signals in the initial training sample set respectively to obtain Doppler domain signals, U, of P initial training samplesnA doppler domain signal representing an nth initial training sample, n 1, 2.., P;
2.2) calculating the Doppler domain signal of the initial training sample after amplitude normalization according to the Doppler domain signal of the initial training sample according to the following formula:
wherein, Xn(k) Representing the Doppler domain signal X of the nth initial training sample after amplitude normalizationnK-th point, k1, 2, N represents the total number of points per doppler domain signal, Un(k) Doppler domain signal U representing nth initial training samplenThe kth point;
2.3) calculating the frequency domain waveform entropy characteristic of the initial training sample according to the Doppler domain signal of the initial training sample after amplitude normalization according to the following formula:
wherein E isnRepresenting the frequency-domain waveform entropy characteristics, X, of the nth initial training samplen(l) Representing the ith point in the doppler domain signal of the nth initial training sample after amplitude normalization, wherein l is 1, 2.
2.4) calculating the frequency domain p-order central moment characteristic of the initial training sample according to the Doppler domain signal of the initial training sample after amplitude normalization according to the following formula:
wherein MpnRepresenting the characteristic of the p-th central moment in frequency domain of the nth initial training sample,
representing the first-order origin moment of the Doppler domain signal of the nth initial training sample after amplitude normalization, and making p equal to 2,4,6 and 8 to obtain the frequency domain second-order center moment feature M2 of the nth initial training samplenFrequency domain fourth order central moment feature M4nFrequency domain sixth-order central moment feature M6nFrequency domain eighth-order central moment feature M8n；
2.5) frequency domain waveform entropy characteristic E according to the nth initial training samplenFrequency domain second order central moment feature M2nFrequency domain fourth order central moment feature M4nFrequency domain sixth-order central moment feature M6nFrequency domain eighth-order central moment feature M8nObtaining the feature vector F of the nth sample in the initial training sample setO,n＝[En,M2n,M4n,M6n,M8n]And combining the feature vectors of P samples in the initial training sample set to obtain a feature matrix of the initial training sample set:
FO＝[FO,1；...；FO,n；...；FO,P]；
2.6) respectively setting a class label value L for each sample in the initial training sample setO,nSetting the helicopter into a category 1, setting the propeller plane into a category 2, and setting the jet plane into a category 3 to obtain a label vector set of the P-dimensional initial training sample set:
LO＝[LO,1,...,LO,n,...LO,P]。
step 3, according to the class label set L of the initial training sample setOAnd a feature matrix F of the initial training sample setOAnd training the MF model of the Mondrian forest to obtain a pre-training MF model.
The Mondrian forest MF is a novel existing random forest, has online updating capacity and is composed of a plurality of Mondrian trees. The specific training process is as follows:
3.1) setting A Mondrian trees in a Mondrian forest MF model, wherein A is more than or equal to 10 and less than or equal to 50, and performing initial training on a feature matrix F of a sample setOAnd a class label set L of the initial training sample setOInput to the MF model;
3.2) generating algorithm for each Mondrian tree T in turn by the Mondrian treetTraining is carried out, wherein t is epsilon [1, A ∈]：
3.2.1) initializing Tree TtTree TtThe method comprises the steps of including a root node epsilon, and setting a class label set L of an initial training sample setOAnd a feature matrix F of the initial training sample setOInput into ε and set tree TtThe survival time parameter λ of (a);
3.2.2) setting the Tree TtInitializing the iteration node j to a root node epsilon, and executing 3.2.3);
3.2.3) computing the feature matrix F of the sample set on the iteration node jjIn each feature dimension dm ∈ [1, d ]]Upper bound uj,dmAnd a lower bound lj,dmTo obtain the hyperspace B corresponding to the iteration node jjA range in each dimension, where d represents the dimension of the feature vector;
3.2.4) computing the Hyperspace BjSum of bound and bound distances in all dimensions
From thetajSampling in the index distribution of the rate index to obtain a time sampling value E;
3.2.5) determining τparent(j)Whether + E < λ holds:
if yes, then set τj＝τparent(j)+ E, execution 3.2.6), where τparent(j)Represents the splitting time limit corresponding to parent node parent (j) of iteration node j, and for root node epsilon, the corresponding tauparent(ε)＝0，τjRepresenting the split time limit corresponding to the iteration node j,
otherwise, set τjλ and node j as tree TtAnd ending the program;
3.2.6) according to the hyperspace BjIn each dimension dm ∈ [1, d ]]The distance between the upper boundary and the lower boundary, and the dimension delta of the splitting characteristic on the iteration node jjProbability of selecting as dimension dm:
according to the probability, the splitting characteristic dimension delta of the iteration node j is obtained by samplingj；
3.2.7) Hyperspaces B from iteration node jjIn the dimension deltajUpper and lower bound of interval, from interval
Upsampling, using the sampled value as the splitting threshold xi of the iteration node jjWherein
Representing a hyperspace BjIn the dimension deltajUpper bound value of，
Representing a hyperspace BjIn the dimension deltajUpper and lower bound values;
3.2.8) creating left child node left (j) and right child node right (j) of iteration node j, creating an empty feature vector set Fleft(j)And Fright(j)Creating an empty set L of category labelsleft(j)And Lright(j)Judging the dimension delta of each sample feature vector in the feature matrix F of the sample set in the iteration node jjWhether the characteristic value of (1) is less than or equal to the splitting condition xij：
If yes, respectively inputting the feature vector and the class label of the sample into a feature vector set F of the left child nodeleft(j)And a class label set L for the left child nodeleft(j)Performing the following steps;
otherwise, respectively inputting the feature vector and the class label of the sample into a feature vector set F of the right child noderight(j)And a class label set L for the right child noderight(j)Wherein F isleft(j)Represents the training sample feature vector, L, contained in node left (j)left(j)Indicates a sample class label contained in node left (j), Fright(j)Represents the training sample feature vector, L, contained in node right (j)right(j)A sample category label included in the representation node right (j);
3.2.9) adding Fleft(j)And Lleft(j)Inputting into left child node left (j), Fright(j)And Lright(j)Input into the right and right child nodes right (j);
3.2.10) updating the circulation node j to the left child node left (j), and returning to 3.2.3);
3.2.11) updating the circulation node j to the right child node right (j), and returning to 3.2.3);
3.3) obtaining the trained A Mondrian trees after the execution of the step 3.2), and integrating the trained A Mondrian trees to obtain the pre-training MF model.
Step 4, extracting the characteristics of the newly added training sample set to obtain a P' multiplied by 5-dimensional newly added training sampleFeature matrix F of a setIAnd setting the class label of each sample in the newly added training sample set to obtain a class label set L of the newly added training sample setI。
4.1) carrying out fast Fourier transform on P 'echo signals in the newly added training sample set respectively to obtain Doppler domain signals, U, of the P' newly added training samplesn'A doppler domain signal representing the nth new training sample, n '═ 1, 2.., P';
4.2) calculating the Doppler domain signal of the newly added training sample after the amplitude normalization according to the Doppler domain signal of the newly added training sample:
wherein, Xn'(k) Doppler domain signal X representing n' th newly added training sample after amplitude normalizationn'K-th point, k1, 2, N represents the total number of points per doppler domain signal, Un'(k) Doppler domain signal U representing nth' new added training samplen'The kth point;
4.3) calculating the frequency domain waveform entropy characteristics of the newly added training samples according to the Doppler domain signals of the newly added training samples after the amplitude normalization:
wherein E isn'Representing the frequency domain waveform entropy characteristics, X, of the nth' newly added training samplen'(l) Representing the ith point in the doppler domain signal of the nth' newly added training sample after the amplitude is normalized, wherein l is 1, 2.
4.4) calculating the frequency domain p-order central moment characteristic of the newly added training sample according to the Doppler domain signal of the newly added training sample after the amplitude normalization according to the following formula:
wherein Mpn'Representing the frequency domain p-order central moment characteristic of the nth' newly added training sample,
representing the first-order origin moment of the Doppler domain signal of the nth 'newly added training sample after amplitude normalization, and making p equal to 2,4,6 and 8 to obtain the frequency domain second-order center moment feature M2 of the nth' newly added training samplen'Frequency domain fourth order central moment feature M4n'Frequency domain sixth-order central moment feature M6n'Frequency domain eighth-order central moment feature M8n'；
4.5) according to the frequency domain waveform entropy characteristic E of the n' th newly added training samplen'Frequency domain second order central moment feature M2n'Frequency domain fourth order central moment feature M4n'Frequency domain sixth-order central moment feature M6n'Frequency domain eighth-order central moment feature M8n'And obtaining the feature vector of the nth' sample in the newly added training sample set: fI,n'＝[En',M2n',M4n',M6n',M8n']And combining the feature vectors of P' samples in the newly added training sample set to obtain a feature matrix of the newly added training sample set:
FI＝[FI,1；...；FI,n'；...；FI,P']；
4.6) respectively setting a class label value L for each sample in the newly added training sample setI,n'Setting the helicopter into a category 1, setting the propeller plane into a category 2, and setting the jet plane into a category 3 to obtain a label vector set of the P' dimension newly-added training sample set:
LI＝[LI,1,...,LI,n',...LI,P']。
5.1) adding the class label set L of the training sample setIWith additional training sample setsFeature matrix FIInputting the data into a pre-training MF model;
5.2) updating each Mondrian tree T in turn by the Mondrian tree updating algorithmtUpdating is carried out, wherein t is epsilon [1, A]：
5.2.1) sequentially taking a feature matrix F of a newly added training sample setICharacteristic vector x of each sample and from LITakes out the corresponding sample category labels y and inputs them into the tree TtIn the root node epsilon;
5.2.2) setting Tree TtInitializing the iteration node j to a root node epsilon, and executing 5.2.3);
5.2.3) computing the hyperspace B of the new sample feature vector x and the iteration node j in each dimensionjUpper boundary ujDifference e ofuAnd x is in each dimension in relation to the hyperspace BjLower boundary ljDifference e ofl：
eu＝max(x-uj,0)，el＝max(lj-x,0)；
5.2.4) calculating elAnd euIn each dimension dm ∈ [1, d ]]Sum of elements of
From thetaeFor sampling in an exponential distribution of the rate index, obtaining a time sample value E, wherein
Denotes elThe elements in the dimension dm are such that,
denotes euAn element in the dimension dm, d representing the dimension of the feature vector;
5.2.5) determination of τparent(j)+E＜τjWhether or not: if so, perform 5.2.6) to 5.2.10), otherwise, perform 5.2.11) to 5.2.13);
5.2.6) according to elAnd euEach dimension dm ∈ [1, d ]]The splitting feature dimension is selected as the probability of each dimension:
sampling to obtain a splitting characteristic dimension delta according to the probability;
5.2.7) judging the characteristic value x of the new sample characteristic vector x in the dimension deltaδWhether the value is larger than the upper boundary value of the hyperspace of the iteration node j on the dimension delta
Namely, it is
If yes, the slave value taking interval
Up-sampling, using the sampled value as splitting threshold xi, otherwise, from interval
Up-sampling uniformly to obtain a splitting threshold xi, wherein,
representing the lower boundary value of the hyperspace of the node j in the dimension delta;
5.2.8) create a new node j', let δj'＝δ，ξj'＝ξ，τj'＝τparent(j)+E，lj'＝min(lj,x)，uj'＝max(ujX) and replacing the position of the iteration node j in the tree with a new node j', where δj'The dimension, ξ, of the splitting feature representing node jj'Denotes the splitting threshold, τ, of node jj'Denotes the splitting time limit of node jj'Lower bound value, u, representing each dimension of the hyperspace of node jj'Representing the upper bound of the hyperspace of node j';
5.2.9) creating a new leaf node j ", inputting the new sample feature vector x and the corresponding sample class label y into the leaf node j";
5.2.10) determining the new sample feature vector x in the dimension deltaj'Characteristic value x ofδj'Whether greater than xij': if so, taking the iteration node j as a right child node of the new node j ', and taking the new leaf node j ' as a left child node of the new node j '; otherwise, the iteration node j is used as a left child node of the new node j ', and the new leaf node j ' is used as a left child node of the node j ';
5.2.11) updating the superspace upper boundary u of the iteration node j according to the new sample feature vector xjAnd a lower boundary lj：
lj＝min(lj,x)，uj＝max(uj,x)；
5.2.12) judging whether the iteration node j is a leaf node, if so, ending the program, otherwise, executing 5.2.13);
5.2.13) determining the x dimension delta of the new sample feature vectorjCharacteristic value of
Whether or not less than or equal to splitting threshold xi of iteration node jj：
If yes, inputting the new sample feature vector x and the corresponding sample type label y into a left child node left (j) of the iteration node j, updating the iteration node j into the left child node left (j), and returning to 5.2.3);
otherwise, inputting the new sample feature vector x and the corresponding sample category label y into a right child node right (j) of the iteration node j, updating the iteration node j into the right child node right (j), and returning to 5.2.3);
5.3) after the execution of the step 3.2), obtaining updated A Mondrian trees, and integrating the updated A Mondrian trees to obtain a new MF model.
Step 6, extracting the characteristics of the test sample set to obtain a characteristic matrix F of the Qx 5-dimensional test sample setTAnd setting the class label of each sample in the test sample set to obtain a class label set L of the test sample setT。
6.1) respectively carrying out fast Fourier transform on Q echo signals in the test sample set to obtain Doppler domain signals, U, of the Q test sample setsmDoppler domain representing the mth test sampleA signal, m ═ 1,2,.., Q;
6.2) calculating the Doppler domain signal of the test sample after the amplitude normalization according to the following formula:
wherein, Xm(k) Showing the Doppler domain signal X of the m test sample after the amplitude normalizationmK-th point, k1, 2, N represents the total number of points per doppler domain signal, Um(k) Doppler domain signal U representing the mth test samplenThe kth point;
6.3) calculating the frequency domain waveform entropy characteristics of the test sample according to the Doppler domain signal of the test sample after the amplitude normalization:
wherein E ismRepresenting the frequency domain waveform entropy characteristics, X, of the mth test samplem(l) Denotes the ith point in the doppler domain signal of the mth test sample after the amplitude normalization, i.e. 1, 2.
6.4) calculating the frequency domain p-order central moment characteristic of the test sample according to the Doppler domain signal of the test sample after the amplitude normalization according to the following formula:
wherein MpmRepresenting the p-th central moment characteristic of the m-th test sample in the frequency domain,
the first-order origin moment of the Doppler domain signal of the mth test sample after the amplitude normalization is represented, p is made to be 2,4,6 and 8, and the frequency domain second-order center moment feature M2 of the mth test sample is obtainedmFrequency domain fourth order central moment feature M4mFrequency, frequencyCentral moment feature M6 of sixth order of domainmFrequency domain eighth-order central moment feature M8m；
6.5) frequency domain waveform entropy characteristics E according to the m-th test samplemFrequency domain second order central moment feature M2mFrequency domain fourth order central moment feature M4mFrequency domain sixth-order central moment feature M6mFrequency domain eighth-order central moment feature M8mObtaining the characteristic vector F of the mth sample in the test sample setT,m＝[Em,M2m,M4m,M6m,M8m]And combining the feature vectors of Q samples in the test sample set to obtain a feature matrix of the test sample set:
FT＝[FT,1；...；FT,m；...；FT,Q]；
6.6) setting a class label value L for each sample in the test sample set respectivelyT,mSetting the helicopter into a category 1, setting the propeller plane into a category 2, and setting the jet plane into a category 3, and obtaining a label vector set of the Q-dimensional test sample set:
LT＝[LT,1,...,LT,m,...LT,Q]。
step 7, a class label set L of the test sample setTAnd feature matrix F of the test sample setTAnd inputting the result into a new MF model to obtain the classification accuracy of the test sample set.
7.1) sequentially testing the feature matrix F of the sample setTThe sample feature vector s in (1) is input into a new MF model to obtain the classification result of each Mondrian tree
Wherein t is 1, 2.. a, a:
classification results using A Mondrian trees
Voting, and taking the label value with the maximum number of votes as a prediction label of the sample feature vector s;
7.2) obtaining a feature matrix F of the test sample setTAll ofA set of predicted labels for the sample;
7.3) prediction tag set and class tag set L from test sample setTAnd obtaining the classification accuracy of the test sample set.
The present invention is further described below in conjunction with simulation experiments.
1. Simulation conditions are as follows:
the hardware platform of the simulation experiment of the invention is as follows: the processor is Intel (R) core (TM) i7-10750H, the main frequency is 2.60GHz, and the memory is 16 GB.
The software platform of the simulation experiment of the invention is as follows: windows 10 operating system and python 3.6.
The data set used in the simulation experiment of the invention is simulation data, and the physical parameters of the rotors of the helicopter, the propeller aircraft and the jet aircraft are as follows:
the table comprises three types of airplanes including a helicopter, a propeller airplane and a jet airplane, each type of airplane comprises 5 types of airplanes, 15 types of airplanes and L1Distance, L, from the proximal end of the rotor blade to the center of rotation2Indicating the distance of the rotor blade tip to the center of rotation, helicopter data were simulated with reference to actual rotor structure and motion parameters of BK117, m 17, AS350, bell 212, TH-28, proprotor data were simulated with reference to rotor structure and motion parameters of SAAB2000, L-420, L-610G, F406, C-295, and jet data were simulated with reference to 5 typical micro-motion feature structures and motion parameters, respectively designated A, B, C, D, E.
The parameters of the simulated radar are as follows: carrier frequency f05GHz, pulse repetition frequency PRF 4kHz, dwell time tres50 ms. Distance R of target rotor center from radar010km, target azimuth
The signal-to-noise ratio SNR is 10dB and the blade angle modulation is taken into account.
The initial training sample set contains 10 samples of 15 models, and the total number is 150.
The newly added training sample set comprises 90 samples of 15 models, and the total number is 1350.
The test sample set contained 100 samples of 15 models each for a total of 1500.
2. Simulation content and result analysis:
the simulation experiment of the invention is respectively carried out by adopting the invention and a traditional airplane target classification method. The simulation experiment result is the average result of 20 groups of data.
The traditional method for classifying the target of the airplane is as follows: the Leveprofessor introduces the micro-Doppler effect-based airplane target classification method in published 'JEM effect-based airplane target classification method research'.
And creating 15 empty newly-added sample subsets, sequentially taking 30 samples of the three types of airplanes from the newly-added training sample set according to the order of the airplane types, and putting 90 samples into different subsets.
The method is utilized to carry out online learning, firstly, an initial training sample set is used for training the MF model to obtain a pre-training MF model, then, different newly-added training sample subsets are input for 15 times to obtain the classification accuracy of the pre-training MF model and 15 sets of classifier models obtained by online learning on a test sample set.
The method comprises the steps of conducting off-line learning by using a traditional method, firstly training an RF model by using an initial training sample set to obtain a pre-training RF model, then inputting different newly-added training sample subsets into the initial training sample set for 15 times, and retraining the RF model by using an expanded training sample set to obtain the classification accuracy of the pre-training RF model and 15 groups of RF models obtained through off-line learning on a test sample set.
The method comprises the steps of conducting forgetting learning by using a traditional method, firstly training an RF model by using an initial training sample set to obtain a pre-training RF model, then training a new RF model by using different newly-added training sample subsets for 15 times to obtain classification accuracy of the pre-training RF model and 15 groups of RF models obtained by forgetting learning on a test sample set.
The model obtained by the online learning method is compared with the classification performance experimental results of the model obtained by the offline learning and the forgetting learning of the traditional method, namely the classification accuracy rate results of the model obtained by the learning of 15 times in the three methods under the test sample set are drawn into a curve, as shown in fig. 2. In fig. 2, the abscissa represents the number of learning times, and the ordinate represents the classification accuracy of the model on the test sample set. In fig. 2, the solid line marked by circles represents a change curve of the classification accuracy of the model obtained by the method of the present invention on the test sample set increasing with the learning frequency, the solid line marked by asterisks represents a change curve of the classification accuracy of the model obtained by the offline learning method on the test sample set increasing with the learning frequency, and the solid line marked by plus signs represents a change curve of the classification accuracy of the model obtained by directly using each newly added training sample set to train on the test sample set increasing with the learning frequency.
As can be seen from fig. 2, the classification performance of the model obtained by the online learning of the method of the present invention is better than that of the model obtained by the forgetting of the traditional method, and with the continuous input of new samples, the classification performance of the method of the present invention is continuously enhanced, and the classification performance of the finally obtained model is improved by 4.00% compared with that of the initial model; compared with the model obtained by off-line learning in the traditional method, the model obtained by on-line learning in the method has equivalent classification performance.
In conclusion, the method can rapidly and automatically update the original model and continuously improve the classification performance of the model on the aerial airplane target, and the model is updated only by utilizing the newly added training sample set during online learning, so that the method has less memory consumption and good practicability.
Claims (7)
1. An aerial micro-motion target classification online library building method based on a Mongolian forest is characterized by comprising the following steps:
establishing an initial training sample set containing micro-motion target radar echo signals, and performing feature extraction on the initial training sample set to obtain a feature matrix F of the initial training sample setO；
Establishing a newly added training sample set containing micro target radar echo signals, and extracting characteristics of the newly added training sample set to obtain a new training sample setFeature matrix F for increasing training sample setI；
Establishing a test sample set containing a micro-motion target radar echo signal, and performing feature extraction on the test sample set to obtain a feature matrix F of the test sample setT；
A Mondrian trees are arranged in a Mondrian forest MF model, A is more than or equal to 10 and less than or equal to 50, and a feature matrix F of an initial training sample set is usedOInputting the information into the MF model, and training each Mondrian tree through a Mondrian tree generation algorithm until A Mondrian trees are trained to obtain a pre-training MF model;
feature matrix F using newly added training sample setIUpdating each trained Mondrian tree in the pre-trained MF model by a Mondrian tree expansion algorithm until A Mondrian trees are updated to obtain a new MF model;
feature matrix F of test sample setTAnd inputting the classification result into a new MF model to obtain a classification result of the test sample set.
2. The method of claim 1, wherein:
the initial training sample set comprises P micro-motion target radar echo signals of D categories, wherein D is more than or equal to 3, and P is more than or equal to 150;
the newly added training sample set comprises P 'micro target radar echo signals D of D categories, wherein D is more than or equal to 3, and P' is more than or equal to 90;
the test sample set comprises Q micro-motion target radar echo signals of D categories, wherein D is larger than or equal to 3, and Q is larger than or equal to 1500.
3. The method of claim 1, wherein: performing feature extraction on the initial training sample set to obtain a feature matrix F of the initial training sample setOThe implementation is as follows:
firstly, respectively carrying out fast Fourier transform on P echo signals in an initial training sample set to obtain Doppler domain signals, U, of P initial training samplesnA doppler domain signal representing an nth initial training sample, n 1, 2.., P;
then, according to the following formula, calculating the doppler domain signal of the initial training sample after amplitude normalization according to the doppler domain signal of the initial training sample:
wherein, Xn(k) Representing the Doppler domain signal X of the nth initial training sample after amplitude normalizationnK-th point, k1, 2, N represents the total number of points per doppler domain signal, Un(k) Doppler domain signal U representing nth initial training samplenThe kth point;
according to the following formula, calculating the frequency domain waveform entropy characteristics of the initial training sample according to the Doppler domain signal of the initial training sample after the amplitude normalization:
wherein E isnRepresenting the frequency-domain waveform entropy characteristics, X, of the nth initial training samplen(l) Representing the ith point in the doppler domain signal of the nth initial training sample after amplitude normalization, wherein l is 1, 2.
Then, according to the following formula, calculating the frequency domain p-order central moment characteristic of the initial training sample according to the Doppler domain signal of the initial training sample after the amplitude normalization:
wherein MpnRepresenting the characteristic of the p-th central moment in frequency domain of the nth initial training sample,
the first-order origin moment of the Doppler domain signal of the nth initial training sample after the amplitude normalization is represented, and p is made to be 2,4,6 and 8, so that the nth initial training is obtainedFrequency domain second-order central moment feature M2 of training samplenFrequency domain fourth order central moment feature M4nFrequency domain sixth-order central moment feature M6nFrequency domain eighth-order central moment feature M8n；
Finally, according to the frequency domain waveform entropy characteristic E of the nth initial training samplenFrequency domain second order central moment feature M2nFrequency domain fourth order central moment feature M4nFrequency domain sixth-order central moment feature M6nFrequency domain eighth-order central moment feature M8nObtaining the feature vector F of the nth sample in the initial training sample setO,n＝[En,M2n,M4n,M6n,M8n]And combining the feature vectors of P samples in the initial training sample set to obtain a feature matrix of the initial training sample set:
FO＝[FO,1；...；FO,n；...；FO,P]。
4. the method of claim 1, wherein: extracting the characteristics of the newly added training sample set to obtain a characteristic matrix F of the newly added training sample setIThe implementation is as follows:
respectively carrying out fast Fourier transform on P 'echo signals in the newly added training sample set to obtain Doppler domain signals, U, of the P' newly added training samplesn'A doppler domain signal representing the nth ' new training sample, n ' ═ 1, 2.., P ';
according to the following formula, calculating the Doppler domain signal of the newly added training sample after the amplitude normalization according to the Doppler domain signal of the newly added training sample:
wherein, Xn'(k) Doppler domain signal X representing n' th newly added training sample after amplitude normalizationn'K-th point, k1, 2, N represents the total number of points per doppler domain signal, Un'(k) Doppler domain signal U representing nth' new added training samplen'To middlek points;
calculating the frequency domain waveform entropy characteristics of the newly added training sample according to the Doppler domain signal of the newly added training sample after the amplitude normalization:
wherein E isn'Representing the frequency domain waveform entropy characteristics, X, of the nth' newly added training samplen'(l) Representing the ith point in the doppler domain signal of the nth' newly added training sample after the amplitude is normalized, wherein l is 1, 2.
Then, according to the following formula, calculating the frequency domain p-order central moment characteristic of the newly added training sample according to the Doppler domain signal of the newly added training sample after the amplitude normalization:
wherein Mpn'Representing the frequency domain p-order central moment characteristic of the nth' newly added training sample,
representing the first-order origin moment of the Doppler domain signal of the nth 'newly added training sample after amplitude normalization, and making p equal to 2,4,6 and 8 to obtain the frequency domain second-order center moment feature M2 of the nth' newly added training samplen'Frequency domain fourth order central moment feature M4n'Frequency domain sixth-order central moment feature M6n'Frequency domain eighth-order central moment feature M8n'；
According to the frequency domain waveform entropy characteristic E of the n' th newly added training samplen'Frequency domain second order central moment feature M2n'Frequency domain fourth order central moment feature M4n'Frequency domain sixth-order central moment feature M6n'Frequency domain eighth-order central moment feature M8n'And obtaining the feature vector of the nth' sample in the newly added training sample set: fI,n'＝[En',M2n',M4n',M6n',M8n']From newly added training sample setsCombining the feature vectors of P' samples to obtain a feature matrix of the newly added training sample set:
FI＝[FI,1；...；FI,n'；...；FI,P']。
5. the method of claim 1, wherein: extracting the characteristics of the test sample set to obtain a characteristic matrix F of the test sample setTThe implementation is as follows:
respectively carrying out fast Fourier transform on Q echo signals in the test sample set to obtain Doppler domain signals, U, of the Q test sample setsmA doppler domain signal representing the mth test sample, m 1, 2.., Q;
then, according to the following formula, calculating the Doppler domain signal of the test sample after the amplitude normalization according to the Doppler domain signal of the test sample:
wherein, Xm(k) Showing the Doppler domain signal X of the m test sample after the amplitude normalizationmK-th point, k1, 2, N represents the total number of points per doppler domain signal, Um(k) Doppler domain signal U representing the mth test samplenThe kth point;
calculating the frequency domain waveform entropy characteristics of the test sample according to the Doppler domain signal of the test sample after the amplitude normalization:
wherein E ismRepresenting the frequency domain waveform entropy characteristics, X, of the mth test samplem(l) Representing the ith point in the doppler domain signal of the mth test sample after the amplitude value is normalized, wherein l is 1, 2.
According to the following formula, calculating the frequency domain p-order central moment characteristic of the test sample according to the Doppler domain signal of the test sample after the amplitude is normalized:
wherein MpmRepresenting the p-th central moment characteristic of the m-th test sample in the frequency domain,
the first-order origin moment of the Doppler domain signal of the mth test sample after the amplitude normalization is represented, p is made to be 2,4,6 and 8, and the frequency domain second-order center moment feature M2 of the mth test sample is obtainedmFrequency domain fourth order central moment feature M4mFrequency domain sixth-order central moment feature M6mFrequency domain eighth-order central moment feature M8m；
According to the frequency domain waveform entropy characteristic E of the mth test samplemFrequency domain second order central moment feature M2mFrequency domain fourth order central moment feature M4mFrequency domain sixth-order central moment feature M6mFrequency domain eighth-order central moment feature M8mObtaining the characteristic vector F of the mth sample in the test sample setT,m＝[Em,M2m,M4m,M6m,M8m]And combining the feature vectors of Q samples in the test sample set to obtain a feature matrix of the test sample set:
FT＝[FT,1；...；FT,m；...；FT,Q]。
6. the method of claim 1, wherein: each Mondrian tree is trained through a Mondrian tree generation algorithm, and the following is realized:
6a) initializing tree TtTree TtThe method comprises the steps of including a root node epsilon, and using a feature matrix F of an initial training sample setOInput into ε and set tree TtWith a lifetime parameter of [ lambda ], where t ∈ [1, A ]]；
6b) Setting tree TtInitializing the iteration node j to a root node epsilon, and executing 6 c);
6c) calculating a feature matrix F of a sample set on an iteration node jjIn each feature dimension dm ∈ [1, d ]]Upper bound uj,dmAnd a lower bound lj,dmTo obtain the hyperspace B corresponding to the iteration node jjA range in each dimension, where d represents the dimension of the feature vector;
6d) computing the hyperspace BjSum of bound and bound distances in all dimensions
From thetajSampling in the index distribution of the rate index to obtain a time sampling value E;
6e) determination of tauparent(j)Whether + E < λ holds:
if yes, then set τj＝τparent(j)+ E, execution 6f), where τparent(j)Represents the splitting time limit corresponding to parent node parent (j) of iteration node j, and for root node epsilon, the corresponding tauparent(ε)＝0，τjRepresenting the split time limit corresponding to the iteration node j,
otherwise, set τjLet iteration node j be tree TtAnd ending the program;
6f) according to the hyperspace BjIn each dimension dm ∈ [1, d ]]The distance between the upper boundary and the lower boundary, and the dimension delta of the splitting characteristic on the iteration node jjProbability of selecting as dimension dm:
according to the probability, the splitting characteristic dimension delta of the iteration node j is obtained by samplingj；
6g) Hyperspace B from iteration node jjIn the dimension deltajUpper and lower bound of interval, from interval
Upsampling, using the sampled value as the splitting threshold xi of the iteration node jjWherein
Representing a hyperspace BjIn the dimension deltajThe upper bound value of (a) is,
representing a hyperspace BjIn the dimension deltajUpper and lower bound values;
6h) creating a left child node left (j) and a right child node right (j) of the iteration node j, and creating an empty feature vector set Fleft(j)And Fright(j)Judging a feature matrix F of the sample set in the iteration node jjFeature vector of each sample in the dimension deltajWhether the characteristic value of (1) is less than or equal to the splitting threshold xij: if yes, inputting the sample feature vector into a feature vector set F of the left child nodeleft(j)Otherwise, inputting the sample feature vector into the feature vector set F of the right child noderight(j)Wherein, Fleft(j)Represents the training sample feature vector contained in node left (j), Fright(j)Represents the training sample feature vector contained in node right (j);
6i) f is to beleft(j)And Fright(j)Respectively input into a left sub-node left (j) and a right sub-node right (j);
6j) updating the iteration node j to be a left child node left (j), and executing 6 c);
6k) update iteration node j to right child node right (j), perform 6 c).
7. The method of claim 1, wherein: updating each Mongolian Reed-Solomon tree in the pre-trained MF model by a Mongolian Reed-Solomon tree expansion algorithm to realize the following
7a) Sequentially taking feature matrix F of newly added training sample setIThe feature vector x of each sample is input into the tree TtIn the root node ε, wherein t ∈ [1, A ]]；
7b) Setting tree TtInitializing the iteration node j to a root node epsilon, and executing 7 c);
7c) calculating new sample feature vector x from iteration node j in each dimensionUltra space BjUpper boundary ujDifference e ofuAnd x is in each dimension in relation to the hyperspace BjLower boundary ljDifference e ofl：
eu＝max(x-uj,0)，el＝max(lj-x,0)；
7d) Calculating elAnd euIn each dimension dm ∈ [1, d ]]Sum of elements of
From thetaeFor sampling in an exponential distribution of the rate index, obtaining a time sample value E, wherein
Denotes elThe elements in the dimension dm are such that,
denotes euAn element in the dimension dm, d representing the dimension of the feature vector;
7e) determination of tauparent(j)+E＜τjWhether or not: if so, executing 7f) to 7j), otherwise, executing 7k) to 7 m);
7f) according to elAnd euEach dimension dm ∈ [1, d ]]Calculating the probability of selecting the splitting characteristic dimension as each dimension:
sampling to obtain a splitting characteristic dimension delta according to the probability;
7g) judging the characteristic value x of the new sample characteristic vector x on the dimension deltaδWhether the value is larger than the upper boundary value of the hyperspace of the iteration node j on the dimension delta
Namely, it is
If yes, the slave value taking interval
Up-sampling, using the sampled value as splitting threshold xi, otherwise, from interval
Up-sampling uniformly to obtain a splitting threshold xi, wherein,
representing the lower boundary value of the hyperspace of the node j in the dimension delta;
7h) create a new node j', let deltaj'＝δ，ξj'＝ξ，τj'＝τparent(j)+E，lj'＝min(lj,x)，uj'＝max(ujX) and replacing the position of the iteration node j in the tree with a new node j', where δj'The dimension, ξ, of the splitting feature representing node jj'Denotes the splitting threshold, τ, of node jj'Denotes the splitting time limit of node jj'Lower bound value, u, representing each dimension of the hyperspace of node jj'Representing the upper bound of the hyperspace of node j';
7i) creating a new leaf node j ', and inputting a new sample feature vector x into the leaf node j';
7j) judging the dimension delta of the new sample characteristic vector xj'Characteristic value x ofδj'Whether greater than xij': if so, taking the iteration node j as a right child node of the new node j ', and taking the new leaf node j ' as a left child node of the new node j '; otherwise, the iteration node j is used as a left child node of the new node j ', and the new leaf node j ' is used as a left child node of the node j ';
7k) updating the upper boundary u of the hyperspace of the iteration node j according to the new sample feature vector xjAnd a lower boundary lj：
lj＝min(lj,x)，uj＝max(uj,x)；
7l) judging whether the iteration node j is a leaf node, if so, ending the program, otherwise, executing 7 m);
7m) determining the x dimension delta of the new sample feature vectorjCharacteristic value of
Whether or not less than or equal to splitting threshold xi of iteration node jj：
If yes, inputting the new sample feature vector x into the left child node left (j) of the iteration node j, updating the iteration node j into the left child node left (j), returning to 7c),
otherwise, the new sample feature vector x is input into the right child node right (j) of the iteration node j, the iteration node j is updated to the right child node right (j), and 7c) is returned.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110589218.5A CN113325380B (en) | 2021-05-28 | 2021-05-28 | Air micro-motion target classification on-line library building method based on Mondrian forest |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110589218.5A CN113325380B (en) | 2021-05-28 | 2021-05-28 | Air micro-motion target classification on-line library building method based on Mondrian forest |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113325380A true CN113325380A (en) | 2021-08-31 |
| CN113325380B CN113325380B (en) | 2024-04-02 |
Family
ID=77422135
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110589218.5A Active CN113325380B (en) | 2021-05-28 | 2021-05-28 | Air micro-motion target classification on-line library building method based on Mondrian forest |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113325380B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114841201A (en) * | 2022-04-23 | 2022-08-02 | 中国人民解放军32802部队 | Design method and device of dynamic knowledge base for intelligent radar countermeasure |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109190673A (en) * | 2018-08-03 | 2019-01-11 | 西安电子科技大学 | The ground target classification method sentenced is refused based on random forest and data |
| WO2019153388A1 (en) * | 2018-02-12 | 2019-08-15 | 大连理工大学 | Power spectral entropy random forest-based aeroengine rolling bearing fault diagnosis method |
| US20200271720A1 (en) * | 2020-05-09 | 2020-08-27 | Hefei University Of Technology | Method for diagnosing analog circuit fault based on vector-valued regularized kernel function approximation |
| CN112686871A (en) * | 2020-12-31 | 2021-04-20 | 杭州电子科技大学 | SAR image change detection method based on improved logarithmic ratio operator and Gabor _ ELM |
| CN112835003A (en) * | 2020-12-31 | 2021-05-25 | 西安电子科技大学 | Radar repetition frequency change steady target identification method based on resampling preprocessing |
-
2021
- 2021-05-28 CN CN202110589218.5A patent/CN113325380B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019153388A1 (en) * | 2018-02-12 | 2019-08-15 | 大连理工大学 | Power spectral entropy random forest-based aeroengine rolling bearing fault diagnosis method |
| CN109190673A (en) * | 2018-08-03 | 2019-01-11 | 西安电子科技大学 | The ground target classification method sentenced is refused based on random forest and data |
| US20200271720A1 (en) * | 2020-05-09 | 2020-08-27 | Hefei University Of Technology | Method for diagnosing analog circuit fault based on vector-valued regularized kernel function approximation |
| CN112686871A (en) * | 2020-12-31 | 2021-04-20 | 杭州电子科技大学 | SAR image change detection method based on improved logarithmic ratio operator and Gabor _ ELM |
| CN112835003A (en) * | 2020-12-31 | 2021-05-25 | 西安电子科技大学 | Radar repetition frequency change steady target identification method based on resampling preprocessing |
Non-Patent Citations (1)
| Title |
|---|
| 邱一卉;米红;: "基于随机森林和转导推理的特征提取方法", 厦门大学学报(自然科学版), no. 03 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114841201A (en) * | 2022-04-23 | 2022-08-02 | 中国人民解放军32802部队 | Design method and device of dynamic knowledge base for intelligent radar countermeasure |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113325380B (en) | 2024-04-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109766583A (en) | Based on no label, unbalanced, initial value uncertain data aero-engine service life prediction technique | |
| CN110133599B (en) | Intelligent radar radiation source signal classification method based on long-time and short-time memory model | |
| CN110018453B (en) | Intelligent model identification method based on aircraft track characteristics | |
| CN111444790B (en) | Pulse-level intelligent identification method for multifunctional radar working mode sequence | |
| CN111273288B (en) | Radar unknown target identification method based on long-term and short-term memory network | |
| CN110082738B (en) | Radar target identification method based on Gaussian mixture and tensor recurrent neural network | |
| CN113341919B (en) | Computing system fault prediction method based on time sequence data length optimization | |
| CN111897353A (en) | Aircraft maneuvering trajectory prediction method based on GRU | |
| Zhou et al. | A novel radar signal recognition method based on a deep restricted Boltzmann machine | |
| CN113325380A (en) | Air micro-motion target classification online library building method based on Mondrian forest | |
| CN114818853B (en) | Intention recognition method based on bidirectional gating circulating unit and conditional random field | |
| CN112666533B (en) | Repetition frequency change steady target identification method based on spatial pyramid pooling network | |
| CN108983187B (en) | Online radar target identification method based on EWC | |
| CN115705393A (en) | Radar radiation source grading identification method based on continuous learning | |
| CN113406623A (en) | Target identification method, device and medium based on radar high-resolution range profile | |
| CN113311407B (en) | Air plane target classification online learning method based on support vector machine | |
| CN115964923A (en) | Modeling method for forecasting 80-100km atmospheric wind speed in adjacent space based on VMD-PSO-LSTM | |
| CN116304966A (en) | Track association method based on multi-source data fusion | |
| CN114296067A (en) | Pulse Doppler radar low-slow small target identification method based on LSTM model | |
| CN111160464B (en) | Industrial high-order dynamic process soft measurement method based on multi-hidden-layer weighted dynamic model | |
| Liu et al. | Recognition and classification of rotorcraft by micro-Doppler signatures using deep learning | |
| Kouemou et al. | Hidden Markov models in radar target classification | |
| Guo et al. | Deep Transfer Learning Based Method for Radar Automatic Recognition with Small Data Size | |
| CN114298183B (en) | Intelligent recognition method for flight actions | |
| Qiang et al. | Deinterleaving of radar pulse based on implicit feature |
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 |