CN108694346A - A kind of Ship Radiated-Noise signal recognition method based on two-stage CNN - Google Patents
A kind of Ship Radiated-Noise signal recognition method based on two-stage CNN Download PDFInfo
- Publication number
- CN108694346A CN108694346A CN201710217744.2A CN201710217744A CN108694346A CN 108694346 A CN108694346 A CN 108694346A CN 201710217744 A CN201710217744 A CN 201710217744A CN 108694346 A CN108694346 A CN 108694346A
- Authority
- CN
- China
- Prior art keywords
- spectrum
- cnn
- target
- frequency spectrum
- test set
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Signal Processing (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN, and this method specifically includes:The radiated noise signals of the ship of known class are pre-processed, and build training set and test set;Determine the number of first order convolutional neural networks and second level convolutional neural networks;Build the training set and test set of first order CNN;Build the training set and test set of second level CNN;First order CNN and second level CNN is respectively trained;Ship Radiated-Noise signal to be identified is read, spectrum signature is extracted, then be input in the first order CNN, exports each lower target of frequency spectrum similarity, and obtain all kinds of higher targets of frequency spectrum similarity;The envelope characteristic and line spectrum feature of frequency spectrum are extracted respectively, then is separately input to the second level CNN, respectively obtain the higher target identification result of frequency spectrum similarity based on envelope and line spectrum feature;Decision fusion is carried out again, determines final recognition result.
Description
Technical field
The present invention relates to the technical field for propagating Recognition of radiated noises, more particularly to a kind of ship spokes based on two-stage CNN
Penetrate noise signal recognition methods.
Background technology
The identification of Ship Radiated-Noise signal belongs to complicated identification problem.Ship Radiated-Noise signal data is between various countries
It typically maintains secrecy, is not easy to obtain, therefore, the identification of Ship Radiated-Noise signal is the even small sample in finite sample
Under the conditions of complete.Even in addition, same ship, in different running working conditions, generated radiated noise is also to have
Certain difference, this also gives identification work to bring certain difficulty.In addition, often ship similar in tonnage, radiated noise frequency
Spectrum signature is also quite similar, is difficult to distinguish them one by one if only extraction spectrum signature.
Feature extraction is the initial characteristic data that will obtain by the method for mapping, the primitive character of higher-dimension is changing into low
The characteristic quantity of dimension.Currently, there are many clarification of objective extracting method, and they have very strong specific aim, not because of identification object
It is same and different.In modern signal processing, the characteristic quantities such as high-order statistic, wavelet coefficient, mel cepstrum coefficients by
It is widely used in the identification of Ship Radiated-Noise signal.But the extraction process of these features is all complex.
In the case where targeted species are less, spectrum signature is first extracted, reuses convolutional neural networks
(Convolutional Neural Network, abbreviation CNN) is identified as grader, can obtain identification effect well
Fruit.Meanwhile compared to traditional neural network, it does not have the problem of over-fitting, and training speed is quickly.However, this method
It is only applicable in the case that targeted species are less and target spectrum similarity-rough set is low, if the targeted species ratio identified in needs
It is more, and there are the frequency spectrum similarity-rough set height of some targets, after the spectrum signature for only extracting noise, single use CNN conducts
Grader is identified, then cannot reach more satisfactory recognition effect.
Invention content
It is an object of the present invention in order to solve above-mentioned the deficiencies in the prior art, the present invention provides one kind being based on two
The Ship Radiated-Noise signal recognition method of grade CNN.This method is effectively applied to a variety of Ship Radiated-Noise signals
In identification, the complexity of characteristic extraction procedure is reduced, improves the accuracy of identification.This method specifically includes:
Step 1, the pretreatment that the radiated noise signals of the ship of known class are normalized, removed with DC component,
Training set and test set are built to the Ship Radiated-Noise signal of pretreated known class;Obtain all kinds of frequency spectrum similarities compared with
High target and all kinds of lower targets of frequency spectrum similarity, determine first order convolutional neural networks respectively, i.e. first order CNN and the
Two level convolutional neural networks, the i.e. number of second level CNN;
In the step 1, training set and test set are built to the Ship Radiated-Noise signal of pretreated known class
Detailed process it is as follows:
Step 1-1), structure known class Ship Radiated-Noise signal training set and test set.
The Ship Radiated-Noise signal of known class after normalization is configured to training set, by the training set
10% data are taken out, and test set is configured to;
Step 1-2), respectively extraction step 1-1) constructed by training set and test set in the known class ship
The spectrum signature of radiated noise signals;Then specific extracting method is as follows:
Step 1-2-1), extraction step 1-1) in training set the known class Ship Radiated-Noise signal frequency
Spectrum signature, it is assumed that the sample frequency of the Ship Radiated-Noise signal data of known class be N hertz, to step 1-1) in training
The Ship Radiated-Noise signal data of the known class of concentration, carries out leaf transformation in N point discrete Fouriers, and calculation formula is as follows:
Wherein, x (j) is the data sequence of the training set of the Ship Radiated-Noise signal of known class, and X (k) is to carry out N points
The spectrum sequence of training set after discrete Fourier transform, the i.e. spectrum signature of training set;
Step 1-2-2), will be in step 1-2-1) in the spectrum sequence of gained indicated in the form of black white image, and it is described
The size of black white image is 32*32, wherein white indicates the spectrum sequence X (k), and black is background, by the black white image
Input sample as CNN;
Step 1-2-3), extraction step 1-1) in test set the known class Ship Radiated-Noise signal frequency
Spectrum signature;Its process and extraction step 1-1) in training set the known class Ship Radiated-Noise signal frequency spectrum it is special
Levy identical, detailed process is as follows:
Assuming that the sample frequency of the Ship Radiated-Noise signal data of known class be N hertz, to step 1-1) in survey
The Ship Radiated-Noise signal data of the known class of collection is tried, leaf transformation in N point discrete Fouriers is carried out, calculation formula is as follows:
Wherein, x (j) 'For the data sequence of the test set of the Ship Radiated-Noise signal of known class, X (k) 'To carry out N
The spectrum sequence of test set in point discrete Fourier after leaf transformation, i.e. test set spectrum signature;
Step 1-2-4), will be in step 1-2-3) in the spectrum sequence of gained indicated in the form of black white image, and it is described
The size of black white image is 32*32, wherein white indicates spectrum sequence X (the k) ', black is background;
Step 1-3), by step 1-2) in extraction the training set of spectrum signature input grader, iterations are set
It is 200 times, training grader, and complete to train;
The grader uses CNN, and concrete structure parameter is:Network is of five storeys altogether, and there are two convolutional layer and two ponds altogether
Layer, is connected as input layer (I)-convolutional layer (C1)-pond layer (S1)-convolutional layer (C2)-pond layer (S2)-output layer, wherein most
Two layers is full connection afterwards.Convolution kernel size is 5*5, and C1 layers have 6 convolution kernels, and C2 layers have 12 convolution kernels.Learning rate is 1, instruction
The quantity for practicing total sample is 200, activation primitive sigmoid;
Step 1-4), by step 1-2) in extracted the test set input step 1-3 of spectrum signature) in the classification
Device is presorted, and obtains recognition result;
Step 1-5), analytical procedure 1-4) in obtained recognition result, determine the frequency spectrum similarity between each target, specifically
Process is as follows:
Step 1-5-1), mutual misidentification number is classified as the higher target of frequency spectrum similarity more than 1 target, will be without mutual
Misidentification or mutual misidentification number is that 1 target is classified as frequency spectrum similarity compared with low target;
Wherein, mutual misidentification number refers to:If target A is identified as target B or target B is identified as target A, it is believed that mesh
A is marked, the mutual misidentification numbers of target B are 1;If target A, the mutual misidentification numbers of target B are 1, while target A, target C or target B,
The mutual misidentification numbers of target C are 1, it is believed that the mutual misidentification number of target A, target B, target C is 2.
Step 1-5-2), the higher target of frequency spectrum similarity is classified as one kind respectively, the lower target of frequency spectrum similarity is respectively
It is classified as one kind;
Step 1-6), operation through the above steps, determine two-stage CNN identifying system structures, i.e. a first order CNN and
In step 1-5-2) in the corresponding second level CNN of each described higher target of frequency spectrum similarity;
Wherein, in step 1-6) in, the network architecture parameters of first order CNN and second level CNN are:Network is of five storeys altogether,
There are two convolutional layer and two pond layers altogether, are connected as input layer (I)-convolutional layer (C1)-pond layer (S1)-convolutional layer (C2)-
Pond layer (S2)-output layer, wherein last two layers is full connection.Convolution kernel size is 5*5, and C1 layers have 6 convolution kernels, C2 layers
There are 12 convolution kernels.Learning rate is 1, and the quantity of batch total sample of training is 200, activation primitive sigmoid.
The training set and test set of step 2, structure first order CNN;According to the construction method of step 1, first is extracted respectively
The grade training set of the CNN and spectrum signature of test set, obtain the first order CNN all kinds of higher targets of frequency spectrum similarity and
All kinds of lower targets of frequency spectrum similarity;
Step 3, the training set and test set for building second level CNN;In the training set and test set of the CNN of the extraction first order
On the basis of spectrum signature, according to the construction method of step 1, the spectrum envelope and line spectrum of the training set of second level CNN are extracted respectively
Feature extracts the spectrum envelope and line spectrum feature of the test set of second level CNN,
The training set for the first order CNN that spectrum signature is extracted in step 2 is inputted the first order CNN by step 4,
Iterations, training first order CNN are set, and complete to train;By the second level CNN after extraction spectrum signature in step 2
Training set input the second level CNN, iterations are set, and training second level CNN completes training;
Step 5 reads Ship Radiated-Noise signal to be identified by equipment, according to the method for step 1, is waited for described in extraction
The spectrum signature of the Ship Radiated-Noise signal of identification;
The Ship Radiated-Noise signal to be identified comprising spectrum signature is input to the first order CNN by step 6
In, classify, export all kinds of lower spectrum sequences of frequency spectrum similarity, and obtains the higher frequency spectrum sequence of all kinds of frequency spectrum similarities
Row;
Step 7, by all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extract frequency spectrum
Envelope characteristic, then it is input to the trained second level CNN, obtain the higher mesh of frequency spectrum similarity based on envelope characteristic
Mark recognition result;
Step 8, by all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extract frequency spectrum
Line spectrum feature, then it is input to the trained second level CNN, obtain the higher mesh of frequency spectrum similarity based on line spectrum feature
Mark recognition result;
Target identification in step 9, analytical procedure 7 and step 8 is as a result, find out step 7 and the obtained identification knot of step 8
The different target sample of fruit, and place it in together, as sample set K undetermined;Decision fusion is carried out, determines final identification knot
Fruit.
In the step 2, the training set of structure first order CNN and the detailed process of test set are as follows:
Step 2-1) radiated noise signals of the ship of known class are normalized, removed with the pre- place of DC component
Reason builds training set to the Ship Radiated-Noise signal of pretreated known class, the 10% of the data of training set is taken out,
It is configured to test set;
Step 2-2) according to step 1-2) extracting method, respectively extraction step 2-1) constructed by training set and test set
In the known class Ship Radiated-Noise signal spectrum signature.
In the step 3, the detailed process of the training set and test set that build second level CNN is as follows:
Step 3-1) radiated noise signals of the ship of known class are normalized, removed with the pre- place of DC component
Reason builds training set to the Ship Radiated-Noise signal of pretreated known class, the 10% of the data of training set is taken out,
It is configured to test set;
Step 3-2) on the basis of extracting the training set and test set spectrum signature of CNN of the first order, according to step 1-2)
Extracting method, respectively extraction step 3-1) constructed by training set and test set in the known class ship radiation make an uproar
The spectrum signature of acoustical signal.
In the step 6, the Ship Radiated-Noise signal to be identified comprising spectrum signature is input to described
In first order CNN, the detailed process classified is as follows:
Step 6-1) the higher target of all kinds of frequency spectrum similarities obtained in step 5 and all kinds of frequency spectrum similarities is lower
Target is as input sample, according to step 1-2) extracting method, extract respective spectrum sequence;
Step 6-2) by step 6-1) in the spectrum sequence that obtains be separately input to trained described first in step 4
Grade CNN in, classify, according to step 1-5) classification results, export each lower spectrum sequence of frequency spectrum similarity, and obtain
Obtain all kinds of higher spectrum sequences of frequency spectrum similarity.
In the step 7, by all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, divide
The envelope characteristic of respective frequency spectrum, detailed process is indescribably taken to be:
All kinds of higher spectrum sequences of frequency spectrum similarity obtained in the step 6 are divided into m sections, per segment length
For L/m, wherein m is the positive number less than L/10;Wherein, L is the total length of the spectrum sequence;;It is taken out respectively per band frequency sequence
In maximum value, the maximum value is connected, obtain spectrum envelope sequence;Again by the spectrum envelope sequence with artwork master
The form of picture indicates, and the size of the black white image is 32*32, wherein white indicates that spectrum envelope sequence, black are the back of the body
Scape, using the black white image as the input sample of second level CNN.
In step 8, pass through all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extraction frequency
The line spectrum feature of spectrum, detailed process are:
If amplitude threshold is M, the kth of all kinds of higher spectrum sequences of frequency spectrum similarity obtained in the step 6
A frequency point spoke value is Pf(k), k=1,2 ..., L, Fen Biejisuan |Pf(k)-Pf(k-1)|And |Pf(k)-Pf(k+1)|If simultaneously
Man Zu |Pf(k)-Pf(k-1)|> M and |Pf(k)-Pf(k+1)|> M, are disregarded, otherwise, by the k frequency points amplitude Pf(k) it sets
It is 0, obtains new Pf(k), k=1,2 ..., L, i.e. spectrum line spectral sequence.Again by the spectrum line spectral sequence with black white image
Form indicate, and the size of the black white image be 32*32, wherein white indicate spectrum line spectral sequence, black is background,
Using the black white image as the input sample of second level CNN.
In the step 9, Decision fusion is carried out, determines that the detailed process of final recognition result is as follows:
Step 9-1) the higher classification of frequency spectrum similarity that step 6 is obtained, the frequency spectrum line spectrum and packet that step 3 is obtained
The test set of network is inputted respectively in corresponding second level CNN, obtains setting respectively based on the test set recognition result of envelope and line spectrum
For m1, m2, the collection of all targets is combined into Ω in the test set of the second level CNN of step 71, Ω is determined respectively1In each target
Basic allocation probability m1(η1) and corresponding uncertainty probability m1(θ), η1It is Ω1In a certain target;The second level CNN of step 8
Test set in the collection of all targets be combined into Ω2, η2It is Ω2In a certain target, respectively determine Ω2In each target it is basic
Allocation probability m2(η2) and uncertainty probability m2(θ);
Step 9-2) target in analytical procedure 7 and step 8, the different target sample of wherein recognition result is found out, and will
It is put together, as sample set K undetermined;If spectrum envelope test set total sample number and frequency spectrum line spectrum test set total sample number are equal
For K;The final recognition result of sample K undetermined is determined according to the following rules:
(3) if m1(η1)>m2(η2), final recognition result is η1;
(4) if m1(η1)<m2(η2), final recognition result is η2;
(3) if m1(η1)=m2(η2), then compare m1(θ),m2(θ);If m1(θ) < m2(θ), final recognition result are η1, no
Then, final recognition result is η2。
In step 9-1) in, determine the basic allocation probability m of each target in spectrum envelope sequence1(η1) and it is uncertain general
Rate m1The detailed process of (θ) is as follows:
Step 9-1-1), set spectrum envelope test set total sample number as K, calculate Ω1In each target basic distribution
Probability m1(η1), calculation formula is as follows:
Wherein,For to target η1Identify correct sample size.
Step 9-1-2), determine the uncertainty probability m of spectrum envelope sequence1(θ), calculation formula is as follows:
In step 9-1) in, determine the basic allocation probability m of each target in spectrum line spectral sequence2(η2) and it is uncertain general
Rate m2The detailed process of (θ) is as follows:
Step 9-1-3), set frequency spectrum line spectrum test set total sample number as K, calculate Ω2In each target basic distribution
Probability m2(η2), calculation formula is as follows:
Wherein,For to target η2Identify correct sample size.
Step 9-1-4), determine the uncertainty probability m of spectrum line spectral sequence2(θ), calculation formula is as follows:
The envelope and line spectrum feature of the extraction frequency spectrum, refer to Ship Radiated-Noise data carrying out discrete Fourier transform
After obtaining frequency spectrum, the line of the envelope of the variation tendency of certain ship target noise and the main feature of reflection target noise will be represented
Spectrum is extracted as new identification feature.
The use two-stage convolutional neural networks (Convolutional Neural Network, abbreviation CNN) are to mesh
Target Ship Radiated-Noise is first classified to be identified afterwards, is referred to and is classified according to the spectrum signature of Ship Radiated-Noise by first order CNN
Afterwards, recognition result of the frequency spectrum similarity compared with low target is obtained, then for the target in the higher classification of frequency spectrum similarity, extraction is each
The envelope and line spectrum feature of class frequency spectrum are identified one by one, and final result is obtained through Decision fusion.
The advantage of the invention is that:The present invention is effectively applied in the identification of a variety of Ship Radiated-Noises, is reduced
The complexity of characteristic extraction procedure improves the accuracy of identification.
Description of the drawings
Fig. 1 is a kind of flow chart of Ship Radiated-Noise signal recognition method based on two-stage CNN of the present invention;
Fig. 2 is a kind of specific embodiment of Ship Radiated-Noise signal recognition method based on two-stage CNN of the present invention
Schematic diagram;
Fig. 3 is a kind of extraction known class of Ship Radiated-Noise signal recognition method based on two-stage CNN of the present invention
Ship Radiated-Noise signal training set spectrum signature spectrum sequence figure;
Fig. 4 is a kind of extraction first order CNN of Ship Radiated-Noise signal recognition method based on two-stage CNN of the present invention
Spectrum envelope feature sequence chart;
Fig. 5 is a kind of extraction first order CNN of Ship Radiated-Noise signal recognition method based on two-stage CNN of the present invention
Frequency spectrum line spectrum feature sequence chart;
Specific implementation mode
Below in conjunction with attached drawing, the present invention is described in further detail.
Embodiment 1.
As shown in Figure 1, the present invention provides a kind of Ship Radiated-Noise signal recognition methods based on two-stage CNN.The party
Method is effectively applied in the identification of a variety of Ship Radiated-Noise signals, reduces the complexity of characteristic extraction procedure, is improved
The accuracy of identification.This method specifically includes:
Step 1, the pretreatment that the radiated noise signals of the ship of known class are normalized, removed with DC component,
Training set and test set are built to the Ship Radiated-Noise signal of pretreated known class;Obtain all kinds of frequency spectrum similarities compared with
High target and all kinds of lower targets of frequency spectrum similarity, determine first order convolutional neural networks respectively, i.e. first order CNN and the
Two level convolutional neural networks, the i.e. number of second level CNN;
In the step 1, training set and test set are built to the Ship Radiated-Noise signal of pretreated known class
Detailed process it is as follows:
Step 1-1), structure known class Ship Radiated-Noise signal training set and test set.
The Ship Radiated-Noise signal of known class after normalization is configured to training set, by the training set
10% data are taken out, and test set is configured to;
Step 1-2), respectively extraction step 1-1) constructed by training set and test set in the known class ship
The spectrum signature of radiated noise signals;Then specific extracting method is as follows:
Step 1-2-1), extraction step 1-1) in training set the known class Ship Radiated-Noise signal frequency
Spectrum signature, it is assumed that the sample frequency of the Ship Radiated-Noise signal data of known class be N hertz, to step 1-1) in training
The Ship Radiated-Noise signal data of the known class of concentration, carries out leaf transformation in N point discrete Fouriers, and calculation formula is as follows:
Wherein, x (j) is the data sequence of the training set of the Ship Radiated-Noise signal of known class, and X (k) is to carry out N points
The spectrum sequence of training set after discrete Fourier transform, the i.e. spectrum signature of training set;
Step 1-2-2), will be in step 1-2-1) in the spectrum sequence of gained indicated in the form of black white image, and it is described
The size of black white image is 32*32, as shown in Figure 3, wherein white indicates the spectrum sequence X (k), and black is background, by institute
State input sample of the black white image as CNN;
Step 1-2-3), extraction step 1-1) in test set the known class Ship Radiated-Noise signal frequency
Spectrum signature;Its process and extraction step 1-1) in training set the known class Ship Radiated-Noise signal frequency spectrum it is special
Levy identical, detailed process is as follows:
Assuming that the sample frequency of the Ship Radiated-Noise signal data of known class be N hertz, to step 1-1) in survey
The Ship Radiated-Noise signal data of the known class of collection is tried, leaf transformation in N point discrete Fouriers is carried out, calculation formula is as follows:
Wherein, x (j) 'For the data sequence of the test set of the Ship Radiated-Noise signal of known class, X (k) 'To carry out N
The spectrum sequence of test set in point discrete Fourier after leaf transformation, i.e. test set spectrum signature;
Step 1-2-4), will be in step 1-2-3) in the spectrum sequence of gained indicated in the form of black white image, and it is described
The size of black white image is 32*32, wherein white indicates the spectrum sequence X (k) ', black is background;
Step 1-3), by step 1-2) in extraction the training set of spectrum signature input grader, iterations are set
It is 200 times, training grader, and complete to train;
The grader uses CNN, and concrete structure parameter is:Network is of five storeys altogether, and there are two convolutional layer and two ponds altogether
Layer, is connected as input layer (I)-convolutional layer (C1)-pond layer (S1)-convolutional layer (C2)-pond layer (S2)-output layer, wherein most
Two layers is full connection afterwards.Convolution kernel size is 5*5, and C1 layers have 6 convolution kernels, and C2 layers have 12 convolution kernels.Learning rate is 1, instruction
The quantity for practicing total sample is 200, activation primitive sigmoid;
Step 1-4), by step 1-2) in extracted the test set input step 1-3 of spectrum signature) in the classification
Device is presorted, and obtains recognition result;
Step 1-5), analytical procedure 1-4) in obtained recognition result, determine the frequency spectrum similarity between each target, specifically
Process is as follows:
Step 1-5-1), mutual misidentification number is classified as the higher target of frequency spectrum similarity more than 1 target, will be without mutual
Misidentification or mutual misidentification number is that 1 target is classified as frequency spectrum similarity compared with low target;
Wherein, mutual misidentification number refers to:If target A is identified as target B or target B is identified as target A, it is believed that mesh
A is marked, the mutual misidentification numbers of target B are 1;If target A, the mutual misidentification numbers of target B are 1, while target A, target C or target B,
The mutual misidentification numbers of target C are 1, it is believed that the mutual misidentification number of target A, target B, target C is 2.
Step 1-5-2), the higher target of frequency spectrum similarity is classified as one kind respectively, the lower target of frequency spectrum similarity is respectively
It is classified as one kind;
Step 1-6), operation through the above steps, determine two-stage CNN identifying system structures, i.e. a first order CNN and
In step 1-5-2) in the corresponding second level CNN of each described higher target of frequency spectrum similarity;
Wherein, in step 1-6) in, the network architecture parameters of first order CNN and second level CNN are:Network is of five storeys altogether,
There are two convolutional layer and two pond layers altogether, are connected as input layer (I)-convolutional layer (C1)-pond layer (S1)-convolutional layer (C2)-
Pond layer (S2)-output layer, wherein last two layers is full connection.Convolution kernel size is 5*5, and C1 layers have 6 convolution kernels, C2 layers
There are 12 convolution kernels.Learning rate is 1, and the quantity of batch total sample of training is 200, activation primitive sigmoid.
The training set and test set of step 2, structure first order CNN;According to the construction method of step 1, first is extracted respectively
The grade training set of the CNN and spectrum signature of test set, obtain the first order CNN all kinds of higher targets of frequency spectrum similarity and
All kinds of lower targets of frequency spectrum similarity;
Step 3, the training set and test set for building second level CNN;In the training set and test set of the CNN of the extraction first order
On the basis of spectrum signature, according to the construction method of step 1, the spectrum envelope and line spectrum of the training set of second level CNN are extracted respectively
Feature extracts the spectrum envelope and line spectrum feature of the test set of second level CNN,
The training set for the first order CNN that spectrum signature is extracted in step 2 is inputted the first order CNN by step 4,
Iterations, training first order CNN are set, and complete to train;By the second level CNN after extraction spectrum signature in step 2
Training set input the second level CNN, iterations are set, and training second level CNN completes training;
Step 5 reads Ship Radiated-Noise signal to be identified by equipment, according to the method for step 1, is waited for described in extraction
The spectrum signature of the Ship Radiated-Noise signal of identification;
The Ship Radiated-Noise signal to be identified comprising spectrum signature is input to the first order CNN by step 6
In, classify, export all kinds of lower spectrum sequences of frequency spectrum similarity, and obtains the higher frequency spectrum sequence of all kinds of frequency spectrum similarities
Row;
Step 7, by all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extract frequency spectrum
Envelope characteristic, then it is input to the trained second level CNN, obtain the higher mesh of frequency spectrum similarity based on envelope characteristic
Mark recognition result;
Step 8, by all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extract frequency spectrum
Line spectrum feature, then it is input to the trained second level CNN, obtain the higher mesh of frequency spectrum similarity based on line spectrum feature
Mark recognition result;
Step 9, analytical procedure 7 and the target identification in step 8 are as a result, find out the different target sample of wherein recognition result
This, and place it in together, as sample set K undetermined;Decision fusion is carried out, determines final recognition result.
In the step 2, the training set of structure first order CNN and the detailed process of test set are as follows:
Step 2-1) radiated noise signals of the ship of known class are normalized, removed with the pre- place of DC component
Reason builds training set to the Ship Radiated-Noise signal of pretreated known class, the 10% of the data of training set is taken out,
It is configured to test set;
Step 2-2) according to step 1-2) extracting method, respectively extraction step 2-1) constructed by training set and test set
In the known class Ship Radiated-Noise signal spectrum signature.
In the step 3, the detailed process of the training set and test set that build second level CNN is as follows:
Step 3-1) radiated noise signals of the ship of known class are normalized, removed with the pre- place of DC component
Reason builds training set to the Ship Radiated-Noise signal of pretreated known class, the 10% of the data of training set is taken out,
It is configured to test set;
Step 3-2) on the basis of extracting the training set and test set spectrum signature of CNN of the first order, according to step 1-2)
Extracting method, respectively extraction step 3-1) constructed by training set and test set in the known class ship radiation make an uproar
The spectrum signature of acoustical signal.
In the step 6, the Ship Radiated-Noise signal to be identified comprising spectrum signature is input to described
In first order CNN, the detailed process classified is as follows:
Step 6-1) the higher target of all kinds of frequency spectrum similarities obtained in step 5 and all kinds of frequency spectrum similarities is lower
Target is as input sample, according to step 1-2) extracting method, extract respective spectrum sequence;
Step 6-2) by step 6-1) in the spectrum sequence that obtains be separately input to trained described first in step 4
Grade CNN in, classify, according to step 1-5) classification results, export each lower spectrum sequence of frequency spectrum similarity, obtain
All kinds of higher spectrum sequences of frequency spectrum similarity.
In the step 7, by all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, divide
The envelope characteristic of respective frequency spectrum, detailed process is indescribably taken to be:
All kinds of higher spectrum sequences of frequency spectrum similarity obtained in the step 6 are divided into m sections, per segment length
For L/m, wherein m is the positive number less than L/10;Wherein, L is the total length of the spectrum sequence;It is taken out respectively per band frequency sequence
In maximum value, the maximum value is connected, obtain spectrum envelope sequence;Again by the spectrum envelope sequence with artwork master
The form of picture indicates, and the size of the black white image is 32*32, as shown in Figure 4, wherein white indicates spectrum envelope sequence,
Black is background, using the black white image as the input sample of second level CNN.
In step 8, pass through all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extraction frequency
The line spectrum feature of spectrum, detailed process are:
If amplitude threshold is M, the kth of all kinds of higher spectrum sequences of frequency spectrum similarity obtained in the step 6
A frequency point spoke value is Pf(k), k=1,2 ..., L, Fen Biejisuan |Pf(k)-Pf(k-1)|And |Pf(k)-Pf(k+1)|If simultaneously
Man Zu |Pf(k)-Pf(k-1)|> M and |Pf(k)-Pf(k+1)|> M, are disregarded, otherwise, by the k frequency points amplitude Pf(k) it sets
It is 0, obtains new Pf(k), k=1,2 ..., L, i.e. spectrum line spectral sequence.Again by the spectrum line spectral sequence with black white image
Form indicate, and the size of the black white image be 32*32, as shown in Figure 5, wherein white indicate spectrum line spectral sequence, it is black
Color is background, using the black white image as the input sample of second level CNN.
In the step 9, Decision fusion is carried out, determines that the detailed process of final recognition result is as follows:
Step 9-1) the higher classification of frequency spectrum similarity that step 6 is obtained, the frequency spectrum line spectrum and packet that step 3 is obtained
The test set of network is inputted respectively in corresponding second level CNN, obtains setting respectively based on the test set recognition result of envelope and line spectrum
For m1, m2, the collection of all targets is combined into Ω in the test set of the second level CNN of step 71, Ω is determined respectively1In each target
Basic allocation probability m1(η1) and corresponding uncertainty probability m1(θ), η1It is Ω1In a certain target;The second level CNN of step 8
Test set in the collection of all targets be combined into Ω2, η2It is Ω2In a certain target, respectively determine Ω2In each target it is basic
Allocation probability m2(η2) and uncertainty probability m2(θ);
Step 9-2) target in analytical procedure 7 and step 8, the different target sample of wherein recognition result is found out, and will
It is put together, as sample set K undetermined;If spectrum envelope test set total sample number and frequency spectrum line spectrum test set total sample number are equal
For K;The final recognition result of sample K undetermined is determined according to the following rules:
(5) if m1(η1)>m2(η2), final recognition result is η1;
(6) if m1(η1)<m2(η2), final recognition result is η2;
(3) if m1(η1)=m2(η2), then compare m1(θ),m2(θ);If m1(θ) < m2(θ), final recognition result are η1, no
Then, final recognition result is η2。
In step 9-1) in, determine the basic allocation probability m of each target in spectrum envelope sequence1(η1) and it is uncertain general
Rate m1The detailed process of (θ) is as follows:
Step 9-1-1), set spectrum envelope test set total sample number as K, calculate Ω1In each target basic distribution
Probability m1(η1), calculation formula is as follows:
Wherein,For to target η1Identify correct sample size.
Step 9-1-2), determine the uncertainty probability m of spectrum envelope sequence1(θ), calculation formula is as follows:
In step 9-1) in, determine the basic allocation probability m of each target in spectrum line spectral sequence2(η2) and it is uncertain general
Rate m2The detailed process of (θ) is as follows:
Step 9-1-3), set frequency spectrum line spectrum test set total sample number as K, calculate Ω2In each target basic distribution
Probability m2(η2), calculation formula is as follows:
Wherein,For to target η2Identify correct sample size.
Step 9-1-4), determine the uncertainty probability m of spectrum line spectral sequence2(θ), calculation formula is as follows:
The envelope and line spectrum feature of the extraction frequency spectrum, refer to Ship Radiated-Noise data carrying out discrete Fourier transform
After obtaining frequency spectrum, the line of the envelope of the variation tendency of certain ship target noise and the main feature of reflection target noise will be represented
Spectrum is extracted as new identification feature.
The use two-stage convolutional neural networks (Convolutional Neural Network, abbreviation CNN) are to mesh
Target Ship Radiated-Noise is first classified to be identified afterwards, is referred to and is classified according to the spectrum signature of Ship Radiated-Noise by first order CNN
Afterwards, recognition result of the frequency spectrum similarity compared with low target is obtained, then for the target in the higher classification of frequency spectrum similarity, extraction is each
The envelope and line spectrum feature of class frequency spectrum are identified one by one, and final result is obtained through Decision fusion.
Embodiment 2.
As shown in Fig. 2, the present invention provides a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN, the party
Method specifically includes:
Step 1,12 kinds of targets to the Ship Radiated-Noise signal of known class, i.e. target A1, A2, A3, A4, A5, B1,
B2, C1, C2, C3, C4, C5 are normalized, remove the pretreatment of DC component, to the ship spoke of pretreated known class
Penetrate noise signal structure training set and test set;Wherein, the test set comes from 10% data of the training set, carries respectively
Respective spectrum signature is taken, acquisition target A1, A2, A3, A4, A5 and B1, the frequency spectrum similarity of B2 are higher, and are classified as respectively
Two class of target A, B;Target C1, C2, C3, C4, C5 frequency spectrum similarity is relatively low, it is respectively classified as one kind;Determine that first order CNN is
One, second level CNN is 2;
The training set and test set of step 2, structure first order CNN;According to the method for step 1, first order CNN is extracted respectively
Training set and test set spectrum signature;
Step 3, the training set and test set for building second level CNN;In the training set and test set of the CNN of the extraction first order
On the basis of spectrum signature, the spectrum envelope and line spectrum feature of target A1, A2, A3, A4, A5 and target B1, B2 are extracted respectively;
The training set for the first order CNN that spectrum signature is extracted in step 2 is inputted the first order CNN by step 4,
Iterations are set 50 times, training first order CNN, and complete to train;By the second level after extraction spectrum signature in step 2
The training set of CNN inputs the second level CNN, setting iterations 200 times, and training second level CNN completes training;
Step 5 reads Ship Radiated-Noise signal to be identified, i.e. target A, B, C1, C2, C3, C4, C5 by equipment,
Totally 7 noise-like signal;According to the method for step 1, all kinds of higher targets of frequency spectrum similarity, i.e. two class of A and B are obtained;And it is each
The lower target of class frequency spectrum similarity, i.e. target C1, C2, C3, C4, C5;
Step 6, according to step 1-2) extracting method, 7 classification target spectrum sequences in extraction step 5 input respectively
Into the trained first order CNN, classify, the lower recognition result of output spectrum similarity, i.e. target C1,
C2,C3,C4,C5;And obtain the higher spectrum sequence of frequency spectrum similarity, i.e. two class of target A, B;
Step 7, by target A, B to being obtained in step 6, extract the envelope characteristic of frequency spectrum, then be input to by training
The second level CNN, obtain the higher target identification result of frequency spectrum similarity based on envelope characteristic;
Step 8, by target A, B to being obtained in step 6, extract the line spectrum feature of frequency spectrum, then be input to by training
The second level CNN, obtain the higher target identification result of frequency spectrum similarity based on line spectrum feature;
Step 9, analytical procedure 7 and the target identification in step 8 as a result, find out the different target sample of recognition result, and
It places it in together, as sample set K undetermined;Carry out Decision fusion, determine final recognition result, i.e. target A1, A2, A3, A4,
A5 and B1, B2.
It should be noted last that the above examples are only used to illustrate the technical scheme of the present invention and are not limiting.Although ginseng
It is described the invention in detail according to embodiment, it will be understood by those of ordinary skill in the art that, to the technical side of the present invention
Case is modified or replaced equivalently, and without departure from the spirit and scope of technical solution of the present invention, should all be covered in the present invention
Right in.
Claims (6)
1. a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN, which is characterized in that this method specifically includes:
Step 1 pre-processes the radiated noise signals of the ship of known class, and builds training set and test set;It determines
First order convolutional neural networks, i.e. first order CNN and second level convolutional neural networks, the i.e. number of second level CNN;
The training set and test set of step 2, structure first order CNN;
Step 3, the training set and test set for building second level CNN;
The first order CNN and the second level CNN is respectively trained in step 4;
Step 5 reads Ship Radiated-Noise signal to be identified by equipment;Extract the Ship Radiated-Noise letter to be identified
Number spectrum signature;
The Ship Radiated-Noise signal to be identified comprising spectrum signature is input in the first order CNN by step 6,
And classify, all kinds of lower spectrum sequences of frequency spectrum similarity are exported, and obtain the higher frequency spectrum sequence of all kinds of frequency spectrum similarities
Row;
Step 7, all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extract the envelope characteristic of frequency spectrum,
It is input to the second level CNN again, obtains the higher target identification result of frequency spectrum similarity based on envelope characteristic;
Step 8, all kinds of higher spectrum sequences of frequency spectrum similarity to being obtained in step 6, extract the line spectrum feature of frequency spectrum,
It is input to the second level CNN again, obtains the higher target identification result of frequency spectrum similarity based on line spectrum feature;
Target in step 9, analytical procedure 7 and step 8 is found out the different target sample of wherein recognition result, and is placed it in
Together, as sample set K undetermined;Decision fusion is carried out, determines final recognition result.
2. a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN according to claim 1, feature exist
In in the step 7, to all kinds of higher spectrum sequences of frequency spectrum similarity obtained in step 6, extraction is respective respectively
Frequency spectrum envelope characteristic, detailed process is:
All kinds of higher spectrum sequences of frequency spectrum similarity obtained in the step 6 are divided into m sections, are L/ per segment length
M, wherein m is the positive number less than L/10;Wherein, L is the total length of the spectrum sequence;It is taken out respectively per in band frequency sequence
Maximum value connects the maximum value, obtains spectrum envelope sequence;Again by the spectrum envelope sequence with black white image
Form indicates, and the size of the black white image is 32*32, wherein white indicates spectrum envelope sequence, and black is background, will
Input sample of the black white image as second level CNN.
3. a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN according to claim 1, feature exist
In in step 8, to all kinds of higher spectrum sequences of frequency spectrum similarity obtained in step 6, the line spectrum for extracting frequency spectrum is special
Sign, detailed process are:
If amplitude threshold is M, k-th of frequency of all kinds of higher spectrum sequences of frequency spectrum similarity obtained in the step 6
Point spoke value is Pf(k), k=1,2 ..., L, Fen Biejisuan |Pf(k)-Pf(k-1)|And |Pf(k)-Pf(k+1)|If Man Zu | simultaneously;
Pf(k)-Pf(k-1)|> M and |Pf(k)-Pf(k+1)|> M, are disregarded, otherwise, by the k frequency points amplitude Pf(k) it is set to 0,
Obtain new Pf(k), k=1,2 ..., L, i.e. spectrum line spectral sequence;Again by the spectrum line spectral sequence with the shape of black white image
Formula indicates, and the size of the black white image is 32*32, wherein white indicates spectrum line spectral sequence, and black is background, by institute
State input sample of the black white image as second level CNN.
4. a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN according to claim 1, feature exist
In in the step 9, progress Decision fusion determines that the detailed process of final recognition result is as follows:
Step 9-1) the higher classification of frequency spectrum similarity that step 6 is obtained, the frequency spectrum line spectrum that step 3 is obtained and envelope
Test set is inputted respectively in corresponding second level CNN, obtains being set to m based on the test set recognition result of envelope and line spectrum1,
m2, the collection of all targets is combined into Ω in the test set of the second level CNN of step 71, Ω is determined respectively1In basic point of each target
With probability m1(η1) and corresponding uncertainty probability m1(θ), wherein η1It is Ω1In a certain target;The second level CNN's of step 8
The collection of all targets is combined into Ω in test set2, wherein η2It is Ω2In a certain target;Ω is determined respectively2In each target base
This allocation probability m2(η2) and uncertainty probability m2(θ);
Step 9-2) target in analytical procedure 7 and step 8, the different target sample of wherein recognition result is found out, and put
Together, as sample set K undetermined;If spectrum envelope test set total sample number and frequency spectrum line spectrum test set total sample number are K;
The final recognition result of sample K undetermined is determined according to the following rules:
(1) if m1(η1)>m2(η2), final recognition result is η1;
(2) if m1(η1)<m2(η2), final recognition result is η2;
(3) if m1(η1)=m2(η2), then compare m1(θ),m2(θ);If m1(θ) < m2(θ), final recognition result are η1, otherwise,
Final recognition result is η2。
5. a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN according to claim 4, feature exist
In in step 9-1) in, determine the basic allocation probability m of each target in spectrum envelope sequence1(η1) and uncertainty probability m1
The detailed process of (θ) is as follows:
Step 9-1-1), set spectrum envelope test set total sample number as K, calculate Ω1In each target basic allocation probability m1
(η1), calculation formula is as follows:
Wherein,For to target η1Identify correct sample size;
Step 9-1-2), determine the uncertainty probability m of spectrum envelope sequence1(θ), calculation formula is as follows:
6. a kind of Ship Radiated-Noise signal recognition method based on two-stage CNN according to claim 4, feature exist
In in step 9-1) in, determine the basic allocation probability m of each target in spectrum line spectral sequence2(η2) and uncertainty probability m2
The detailed process of (θ) is as follows:
Step 9-1-3), set frequency spectrum line spectrum test set total sample number as K, calculate Ω2In each target basic allocation probability m2
(η2), calculation formula is as follows:
Wherein,For to target η2Identify correct sample size;
Step 9-1-4), determine the uncertainty probability m of spectrum line spectral sequence2(θ), calculation formula is as follows:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710217744.2A CN108694346B (en) | 2017-04-05 | 2017-04-05 | Ship radiation noise signal identification method based on two-stage CNN |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710217744.2A CN108694346B (en) | 2017-04-05 | 2017-04-05 | Ship radiation noise signal identification method based on two-stage CNN |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108694346A true CN108694346A (en) | 2018-10-23 |
CN108694346B CN108694346B (en) | 2021-12-07 |
Family
ID=63842658
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710217744.2A Active CN108694346B (en) | 2017-04-05 | 2017-04-05 | Ship radiation noise signal identification method based on two-stage CNN |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108694346B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110135316A (en) * | 2019-05-07 | 2019-08-16 | 中国人民解放军海军潜艇学院 | The automatic detection and extracting method of low frequency spectrum lines in a kind of ship-radiated noise |
CN111354373A (en) * | 2018-12-21 | 2020-06-30 | 中国科学院声学研究所 | Audio signal classification method based on neural network intermediate layer characteristic filtering |
CN111354372A (en) * | 2018-12-21 | 2020-06-30 | 中国科学院声学研究所 | Audio scene classification method and system based on front-end and back-end joint training |
CN112257521A (en) * | 2020-09-30 | 2021-01-22 | 中国人民解放军军事科学院国防科技创新研究院 | CNN underwater acoustic signal target identification method based on data enhancement and time-frequency separation |
CN112364779A (en) * | 2020-11-12 | 2021-02-12 | 中国电子科技集团公司第五十四研究所 | Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion |
CN112885362A (en) * | 2021-01-14 | 2021-06-01 | 珠海市岭南大数据研究院 | Target identification method, system, device and medium based on radiation noise |
CN113077813A (en) * | 2021-03-22 | 2021-07-06 | 自然资源部第一海洋研究所 | Ship noise identification method based on holographic spectrum and deep learning |
CN115173971A (en) * | 2022-07-08 | 2022-10-11 | 电信科学技术第五研究所有限公司 | Broadband signal real-time detection method based on frequency spectrum data |
GB2607290A (en) * | 2021-05-28 | 2022-12-07 | Bae Systems Plc | Apparatus and method of classification |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8111581B1 (en) * | 1985-08-09 | 2012-02-07 | Qinetiq Limited | Monitoring system for a ship'S radiated noise |
CN104732970A (en) * | 2013-12-20 | 2015-06-24 | 中国科学院声学研究所 | Ship radiation noise recognition method based on comprehensive features |
CN106529428A (en) * | 2016-10-31 | 2017-03-22 | 西北工业大学 | Underwater target recognition method based on deep learning |
-
2017
- 2017-04-05 CN CN201710217744.2A patent/CN108694346B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8111581B1 (en) * | 1985-08-09 | 2012-02-07 | Qinetiq Limited | Monitoring system for a ship'S radiated noise |
CN104732970A (en) * | 2013-12-20 | 2015-06-24 | 中国科学院声学研究所 | Ship radiation noise recognition method based on comprehensive features |
CN106529428A (en) * | 2016-10-31 | 2017-03-22 | 西北工业大学 | Underwater target recognition method based on deep learning |
Non-Patent Citations (3)
Title |
---|
V.M.KUMBHAR ET AL.: "Underwater ship radiated noise model", 《INTERNATIONAL JOURNAL OF CURRENT ENGINEERING AND SCIENTIFIC RESEARCH》 * |
安良 等: "支持向量机在舰船辐射噪声识别中的应用", 《声学技术》 * |
施建礼 等: "基于子波变换和神经网络的舰船目标识别", 《系统工程与电子技术》 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111354373A (en) * | 2018-12-21 | 2020-06-30 | 中国科学院声学研究所 | Audio signal classification method based on neural network intermediate layer characteristic filtering |
CN111354372A (en) * | 2018-12-21 | 2020-06-30 | 中国科学院声学研究所 | Audio scene classification method and system based on front-end and back-end joint training |
CN110135316B (en) * | 2019-05-07 | 2019-12-31 | 中国人民解放军海军潜艇学院 | Automatic detection and extraction method for low-frequency line spectrum in ship radiation noise |
CN110135316A (en) * | 2019-05-07 | 2019-08-16 | 中国人民解放军海军潜艇学院 | The automatic detection and extracting method of low frequency spectrum lines in a kind of ship-radiated noise |
CN112257521A (en) * | 2020-09-30 | 2021-01-22 | 中国人民解放军军事科学院国防科技创新研究院 | CNN underwater acoustic signal target identification method based on data enhancement and time-frequency separation |
CN112364779B (en) * | 2020-11-12 | 2022-10-21 | 中国电子科技集团公司第五十四研究所 | Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion |
CN112364779A (en) * | 2020-11-12 | 2021-02-12 | 中国电子科技集团公司第五十四研究所 | Underwater sound target identification method based on signal processing and deep-shallow network multi-model fusion |
CN112885362A (en) * | 2021-01-14 | 2021-06-01 | 珠海市岭南大数据研究院 | Target identification method, system, device and medium based on radiation noise |
CN112885362B (en) * | 2021-01-14 | 2024-04-09 | 珠海市岭南大数据研究院 | Target identification method, system, device and medium based on radiation noise |
CN113077813B (en) * | 2021-03-22 | 2022-03-01 | 自然资源部第一海洋研究所 | Ship noise identification method based on holographic spectrum and deep learning |
CN113077813A (en) * | 2021-03-22 | 2021-07-06 | 自然资源部第一海洋研究所 | Ship noise identification method based on holographic spectrum and deep learning |
GB2607290A (en) * | 2021-05-28 | 2022-12-07 | Bae Systems Plc | Apparatus and method of classification |
CN115173971A (en) * | 2022-07-08 | 2022-10-11 | 电信科学技术第五研究所有限公司 | Broadband signal real-time detection method based on frequency spectrum data |
CN115173971B (en) * | 2022-07-08 | 2023-10-03 | 电信科学技术第五研究所有限公司 | Broadband signal real-time detection method based on frequency spectrum data |
Also Published As
Publication number | Publication date |
---|---|
CN108694346B (en) | 2021-12-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108694346A (en) | A kind of Ship Radiated-Noise signal recognition method based on two-stage CNN | |
CN112308158A (en) | Multi-source field self-adaptive model and method based on partial feature alignment | |
CN1258733C (en) | Method for recognizing human its using Daubechies wavelet transform | |
CN105718889A (en) | Human face identity recognition method based on GB(2D)2PCANet depth convolution model | |
CN110414371A (en) | A kind of real-time face expression recognition method based on multiple dimensioned nuclear convolution neural network | |
CN110837768B (en) | Online detection and identification method for rare animal protection | |
CN109828251A (en) | Radar target identification method based on feature pyramid light weight convolutional neural networks | |
CN109635784A (en) | Traffic sign recognition method based on improved convolutional neural networks | |
WO2018166316A1 (en) | Speaker's flu symptoms recognition method fused with multiple end-to-end neural network structures | |
CN110148425A (en) | A kind of camouflage speech detection method based on complete local binary pattern | |
CN104795064A (en) | Recognition method for sound event under scene of low signal to noise ratio | |
CN110175615A (en) | The adaptive visual position recognition methods in model training method, domain and device | |
CN108478216A (en) | A kind of epileptic seizure intelligent Forecasting early period based on convolutional neural networks | |
CN110659649A (en) | Image processing and character recognition algorithm based on near infrared light imaging | |
CN113488060B (en) | Voiceprint recognition method and system based on variation information bottleneck | |
CN110176250A (en) | It is a kind of based on the robust acoustics scene recognition method locally learnt | |
CN113111786B (en) | Underwater target identification method based on small sample training diagram convolutional network | |
CN110866547B (en) | Automatic classification system and method for traditional Chinese medicine decoction pieces based on multiple features and random forests | |
CN104036296A (en) | Method and device for representing and processing image | |
CN114157539A (en) | Data-aware dual-drive modulation intelligent identification method | |
CN106778897B (en) | Plant species recognition methods twice based on COS distance and center profile distance | |
CN109995690B (en) | Neural network self-optimization method for MFSK digital signal subclass modulation recognition | |
CN206781702U (en) | A kind of speech recognition automotive theft proof system based on quantum nerve network | |
CN110472540B (en) | LMD-ICA-PNN-based phi-OTDR vibration signal classification algorithm | |
CN115656958A (en) | Detection method and detection device for real-time track initiation and track classification |
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 |