CN108038471A - A kind of underwater sound communication signal type Identification method based on depth learning technology - Google Patents

A kind of underwater sound communication signal type Identification method based on depth learning technology Download PDF

Info

Publication number
CN108038471A
CN108038471A CN201711443626.XA CN201711443626A CN108038471A CN 108038471 A CN108038471 A CN 108038471A CN 201711443626 A CN201711443626 A CN 201711443626A CN 108038471 A CN108038471 A CN 108038471A
Authority
CN
China
Prior art keywords
layer
deep learning
data
neural networks
convolutional neural
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.)
Pending
Application number
CN201711443626.XA
Other languages
Chinese (zh)
Inventor
殷敬伟
邵梦琦
韩笑
周启明
李成
沈益冉
李理
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Engineering University
Original Assignee
Harbin Engineering University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Harbin Engineering University filed Critical Harbin Engineering University
Priority to CN201711443626.XA priority Critical patent/CN108038471A/en
Publication of CN108038471A publication Critical patent/CN108038471A/en
Pending legal-status Critical Current

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
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24133—Distances to prototypes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Signal Processing (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Image Analysis (AREA)

Abstract

The present invention relates to a kind of underwater sound communication signal type Identification method based on depth learning technology.This method includes：Build deep learning convolutional neural networks model；Default training set specimen discerning accuracy rate T, test set specimen discerning accuracy rate P；Obtain different modulating mode experimental data or emulation data；It will be pre-processed per N number of sample point data as a primary data sample；Data sample random division is training set and test set after pre-processing；With training sample set pair, it is trained；Whether training of judgement collection specimen discerning accuracy rate reaches preset value, when reached, input is switched to data sample test set, with data sample test set to testing；Otherwise continue to train；Judge whether test set specimen discerning accuracy rate reaches preset value, then complete model when reached；Otherwise, excessive data is obtained, with former data mixing, repeats this method.The present invention is solved since signal characteristic abstraction caused by the channel time-varying space-variant of ocean is difficult.

Description

A kind of underwater sound communication signal type Identification method based on deep learning technology
Technical field
The present invention relates to a kind of underwater sound communication signal type Identification method, particularly a kind of water based on deep learning technology Sound communication signal type Identification method, belongs to underwater sound communication and area of pattern recognition.
Background technology
Non- cooperation water sound communication signal signal type Identification studies the important component as hydroacoustic electronic warfare field, increasingly As important research topic.Effectively to realize Classification and Identification, it is necessary to which Modulation recognition difference another characteristic can most be reflected by obtaining.
Model hypo (model hypo, Yang Zhijun, Cao Zhigang《The automatic identification of modulating mode is often used in satellite communication》, communicate journal, 2004,25(1)：A kind of signal of communication modulating mode automatic identifying method based on spectrum signature 140-149) is proposed, from signal Extraction is used as feature vector without the characteristic parameter of modulation parameter in power spectrum, square spectrum, compared with low signal-to-noise ratio still with very Still, underwater acoustic channel has the complex characteristics such as time-varying space-variant to good recognition accuracy so that the characteristic parameter without priori Often show serious unstability, randomness.
From 2012 rise deep learning be a kind of Automatic Feature Extraction recognizer based on data-driven, compare with Toward the feature extraction algorithm based on engineer, the effect that deep learning obtains is more outstanding, the language of Microsoft Research and Google Sound Study of recognition personnel successively reduce speech recognition errors rate 20%~30% using deep learning, are field of speech recognition more than 10 Maximum breakthrough over year.
Especially in image recognition application, the algorithm based on the convolutional neural networks model in deep learning embodies other Non- deep learning method not available for advantage.Document《ImageNet classification with deep convolutional neural networks》(neural information processing systems 2012 can, 1097-1105 pages) exist Error rate is reduced to 15% from 26% in ImageNet image recognitions evaluation and test, 10% or so is higher by than second place.
The content of the invention
To overcome above-mentioned technological deficiency, the technical problem to be solved in the present invention is to provide a kind of more robust based on depth The underwater sound communication signal type Identification method of learning art.
In order to solve the above technical problems, the technical scheme is that：
Step 1：Build a deep learning convolutional neural networks model；
Step 2：Default training set specimen discerning accuracy rate T, presets test set specimen discerning accuracy rate P；
Step 3：Obtain the experimental data or emulation data of the underwater sound communication modulated signal of different modulating mode；
Step 4：Using every N number of sample point data of similar modulation system signal as a primary data sample, each Primary data sample is the column vector of N × 1, to the modulation system of primary data sample mark signal；
Step 5：Each data sample is pre-processed, all primary data samples are converted into deep learning convolution The input form of neutral net, including bandpass filtering is carried out to primary data sample, the column vector of N × 1 is obtained, it is then right The column vector of N × 1 obtained after filtering carries out Hilbert transform, the column vector of another N × 1 is obtained, by the row of 2 N × 1 Vector is merged into the matrix of 1 N × 2；
Step 6：It is training set and test set by pretreated data sample matrix random division；
Step 7：With the parameter of training sample set training deep learning convolutional neural networks model；
Step 8：Whether training of judgement collection specimen discerning accuracy rate reaches preset value, when training set specimen discerning accuracy rate During more than or equal to T, step 9 is jumped to；Otherwise, step 7 is jumped to；
Step 9：The input of deep learning convolutional neural networks is switched to data sample test set, is surveyed with data sample Examination set pair deep learning convolutional neural networks model is tested；
Step 10：Judge whether test set specimen discerning accuracy rate reaches preset value, when test set specimen discerning accuracy rate During more than or equal to P, then step 11 is performed；Otherwise, the former data mixing in excessive data, with the step three is obtained, is expanded Big primary data sample, then jumps to step 4；
Step 11：The setting of deep learning convolutional neural networks parameter is completed, obtains network model parameter, completes mould Type.
Present invention additionally comprises：
1. the output of the deep learning convolutional neural networks model of step 1 is that the data sample tune of acquisition is judged through model Mode processed, to being exported after the advanced row standardization processing of each layer of output result.
2. it is that backpropagation is calculated to be trained method with training sample set pair deep learning convolutional neural networks in step 7 Method.
3. the primary data sample of step 4 meets：
(N-1) × t × B ＞ 40
Wherein, t is the sampling period, and B is the signal band width of the experimental data and emulation data described in step 3.
4. the deep learning convolutional neural networks model of step 1 includes convolutional layer, pond layer and full articulamentum, described Roll up the basic unit number of plies and be more than or equal to 3 layers, the port number of each convolutional layer should be all higher than being equal to 30, each convolutional layer it is every The neuron number of a passage successively successively decreases with convolutional layer, and channel number is successively incremented by；Pond is added behind each layer of convolutional layer Layer, pond layer are maximum pond layer, and the port number of pond layer is equal with the port number of convolutional layer before；The full articulamentum After being connected to last pond layer, the full articulamentum number of plies is more than or equal to 2 layers and is single channel, the full articulamentum of first layer The number of neuron is more than or equal to 100, and the number of the neuron of last layer of full articulamentum is more than or equal to 30, the full connection The neuron number of layer successively successively decreases.
Beneficial effect of the present invention：Since the present invention uses the data-driven learning ability of deep learning, can be very good to solve Certainly since signal characteristic abstraction caused by the channel time-varying space-variant of ocean is difficult, and model can be moved to by transfer learning Embedded development platform, this method are suitable for resisting field for underwater sound communication.Deep learning convolutional neural networks model can be certainly Dynamic extraction feature simultaneously identifies.The present invention saves characteristic of human nature's extraction step, saves a large amount of manpowers, and pattern recognition result is more Robust, the present invention propose a kind of underwater sound communication signal type Identification technology based on deep learning convolutional neural networks algorithm, and A series of signal pretreatment is carried out, such as bandpass filtering, Hilbert transform so that the Feature Extraction Technology of convolutional neural networks obtains With applied in underwater sound Modulation Signals Recognition, since deep learning convolutional neural networks are by data self character Practise, therefore this method has stronger robustness.The present invention is by the use of a part of known signal as training set to well-established Deep learning network is trained；The characteristics of make use of the convolutional neural networks Automatic Feature Extraction in deep learning, makes to letter Number identification process in eliminate the link of artificial extraction feature；The characteristics such as the time-varying space-variant due to ocean channel, artificial spy Sign extraction does not often have robustness, and is a kind of data-driven learning algorithm by deep learning, can be according to available data Automatically extract out the feature of more robust so that discrimination can reach higher level.
Brief description of the drawings
Fig. 1 is inventive algorithm block diagram；
Fig. 2 (a) is the time domain ripple of the underwater sound communication modulated signal sample of binary phase shift keying (BPSK) identification to be sorted Shape；
Fig. 2 (b) is the time domain ripple of the underwater sound communication modulated signal sample of quaternary phase-shift keying (PSK) (QPSK) identification to be sorted Shape；
Fig. 2 (c) is the time domain ripple of the underwater sound communication modulated signal sample of octal system phase-shift keying (PSK) (8PSK) identification to be sorted Shape；
Fig. 2 (d) is the time domain waveform of the underwater sound communication modulated signal sample of direct spreading sequence (DSSS) identification to be sorted；
Fig. 2 (e) is the time domain waveform of the underwater sound communication modulated signal sample of Orthogonal Frequency Division Multiplexing (OFDM) identification to be sorted；
Fig. 3 is the basic computational ele- ment neuron calculating process in deep learning convolutional neural networks；
Fig. 4 is the full binder couse schematic diagram of neutral net
Fig. 5 is the calculating process of a convolution kernel in neutral net convolutional layer
Fig. 6 is the convolutional layer schematic diagram of neutral net multichannel
Fig. 7 is the pond layer calculating process in deep learning convolutional neural networks；
Fig. 8 (a) is the schematic diagram without the full articulamentums of dropout；
Fig. 8 (b) is the schematic diagram by the full articulamentums of dropout；
Fig. 9 is deep learning convolutional neural networks model schematic；
Figure 10 is the convolutional layer containing standardization processing, the combined calculation schematic diagram of pond layer
Figure 11 (a) is the recognition result figure of various modulation systems；
Figure 11 (b) is the discrimination figure of various modulation systems.
Embodiment
Illustrate below in conjunction with the accompanying drawings and the present invention is described in more detail：
The method of the present invention implementation is as follows：
(1) output is built as the deep learning convolutional neural networks model of the judgement to sample modulation system, such as Fig. 9 Shown in Figure 10, each to adding pond layer behind convolutional layer, pond layer is maximum pond layer, and the output to each layer As a result advanced row standardization processing exports again；
(2) experimental data or emulation data of the water sound communication signal of different modulating mode are obtained, then by uniformity signal Data per N number of sampled point as a primary data sample；
(3) all primary data samples are converted to the input form of deep learning convolutional neural networks, i.e., first to original Data sample carries out bandpass filtering, obtains column vector of the filtered result for a N-dimensional, then filtered result is carried out Hilbert transform, the result of conversion is also the column vector of a N-dimensional.Then by Hilbert transform result and bandpass filtering knot Fruit is merged into the data sample matrix of N × 2；
(4) all data sample matrixes are randomly divided into two class of training sample set and test sample collection；
(5) it is trained with training sample set pair deep learning convolutional neural networks, method therefor is to pass through backpropagation Algorithm algorithm.Judge whether the accuracy rate of the recognition result of deep learning convolutional neural networks reaches default index, if reached To index, retain the parameter of network, the input of deep learning convolutional neural networks is switched to test data sample set, if do not had Touch the mark, on the basis of original network model parameter, continue with training sample set pair deep learning convolutional neural networks into Row training, so as to constantly update the parameter of network.Recognition accuracy is judged again, is moved in circles, until recognition accuracy reaches finger Mark；
(6) tested with test sample set pair deep learning convolutional neural networks, if, deep learning convolution during test The recognition accuracy of neutral net also touches the mark, then it is assumed that has completed to be used to identify different water sound communication signal modulation systems Deep learning convolutional neural networks parameter setting, available for other application or test, if being not reaching to index, obtain Excessive data, with former data mixing, expands primary data sample, returns to process flow (2) step, move in circles, until depth Untill study convolutional neural networks touch the mark the discrimination of test sample collection.
The output of deep learning convolutional neural networks model described in step 1 is that the data sample of acquisition is judged through model Modulation system.The sample data length N of this model should ensure that time span (N-1) T of samplesWith multiplying for signal band width B Product is more than 40, i.e. (N-1) TsB ＞ 40, TsFor the sampling period.The convolution number of plies of this model should be greater than being equal to three layers, each convolutional layer Port number should be greater than being equal to 30, the number of neuron can successively successively decrease in each passage in general every layer of convolutional layer, so logical Road number can successively be incremented by.The advanced row standardization processing of output result of convolutional layer exports again.Added behind each layer of convolutional layer Pond layer, pond layer are maximum pond layer, and the port number of pond layer is corresponding with the port number of convolutional layer before.And the volume of model Lamination and the latter linked full articulamentum of pond layer should be greater than being equal to two layers, and be single channel.The neuron of the full articulamentum of first layer Number be more than or equal to 100, the number of the neuron of last layer of full articulamentum should be greater than equal to 30, middle full articulamentum The number of neuron can successively successively decrease.
1st, the detailed elaboration of the related notion in relation to convolutional neural networks and calculating
A, Hilbert transform
Equipped with a time-domain signal f (x), its Hilbert transform is defined as follows：
After conversionStill it is a time-domain signal, due toFrequency spectrum be
SoFrequency spectrum occur relative to the frequency spectrum of f (x) in positive frequencyPhase shift, and occur on negative frequency spectrumPhase shift, i.e.,
For discrete digital signal, discrete Fourier transform can be first carried out, carries out phase-shift operations on positive and negative frequency spectrum, then into Row inverse discrete Fourier transform is tried to achieve.
B, neuron and its calculating
As shown in figure 3, neuron is the basic operation unit in neutral net, it is some of mimic biology nerve cell Characteristic and the mathematicization model built, the input of neuron can be multiple variables but between have point of weight, in neuron Linear combination is carried out, and then itself and needs can just encourage compared with a certain threshold value as neuron self attributes Neuron response output, and generally non-linear relation between the linear combination of response output and the input of neuron, so The calculating of the mathematical model of neuron is as follows
If
Wherein xiFor each input of neuron, wiFor corresponding weight,For nonlinear activation functional vector, i.e.,
Nonlinear activation functionIt is generally following several：
Nonlinear activation function in this model uses the third.
C, full articulamentum neutral net
As shown in figure 4, full articulamentum neutral net is on the basis of single neuron behavior, multiple neurons are arranged It is classified as one layer, each neuron is attached with each input, but the connection weight of each neuron and input and different, Input can be original input data, or the output of preceding layer neuron, but between layer neuron and being not connected to, The connection between two layers of neuron is realized, so that network consisting.So calculating and the single nerve of full articulamentum neutral net The consistent but more subscript of a dimension of calculating of member：
Wherein wijRepresent the i-th connection weight for being input to jth neuron, bjAnd yjIt is threshold value and the output of jth neuron, If with the form of matrix, formula can be rewritten as follows：
The weights and threshold value of neutral net are generally obtained by back-propagation algorithm, the calculating process in relation to back-propagation algorithm It is introduced later.
D, convolutional layer
As shown in Figure 5, Figure 6, the convolutional layer in neural network model has some difference with full articulamentum, first, in convolution Each neuron is connected with all inputs in layer, and is simply connected with the continuum of a part of input data Connect, which is also referred to as the local receptor field of the neuron, and the local receptor fields of all neurons is evenly distributed in whole A input data, and each the size for the local receptor field that neuron is connected is consistent.Second, all receptive fields with it is each From the connection between neuron using same weight coefficient vector, i.e., weights are shared.This shared weights is referred to as convolution kernel. There is the size in size due to convolution kernel, can not be calculated in data edges, output data then can be smaller than the size of input data.Such as Fruit require output with input size it is consistent, can on the periphery of input data with 0 cover so that convolution kernel is at the edge of data It can be calculated.If the element of data is overturn by subscript, the computing of the above is substantially a convolution algorithm, and defeated Enter data and convolution kernel is generally matrix form, then two-dimensional convolution computing is as follows：
Wherein table Mu′Show the computable u ' maximum occurrences of convolution, Mv′It is similar therewith.
Then convolutional layer carries out Nonlinear Mapping equivalent to the image filtering that original two dimensional data are first carried out with some groups, by Obtained in the parameter of wave filter by training of the mass data sample to model, there is certain adaptivity, make filter result more Close to feature of 2-D data itself, and information of the 2-D data on space structure is make use of, the extraction meeting to feature It is more beneficial.
Outputting and inputting for convolutional layer can be multiple passages, but each channel shape is in the same size.When input data is During the composition of multiple channel datas, the connection weight of neuron is shared in same passage, and threshold value is in all channels shares, and The argument of neuron nonlinear activation function is the sum of the convolution at the neuron position of each passage, i.e.,
y(u,v)j=f (∑six(u,v)i*w(u,v)ij+bj)
U, v represent the element position in the 2-D data of input and output, and subscript i represents the i-th passage of input, subscript j tables Show the jth passage of output.
E, standardization processing
Before each neuron carries out Nonlinear Mapping, by the argument of the nonlinear function of all neurons of the passage a(u,v)j=f (∑six(u,v)i*w(u,v)ij+bj) standardization processing is carried out, due to the number one of the neuron of each passage Sample, m represent the number of an output channel neuron, and processing procedure is as follows：
Wherein ε represents the number of some very little more than 0, it is acted on simply prevents a (u, v) all same in Computing So as to the situation generation that denominator is 0.γjAnd βjThe parameter adaptively adjusted for neutral net, is determined by back-propagation algorithm.
F, pond layer
As shown in fig. 7, due to when the number of the neuron when neutral net is many, it is necessary to adjustment parameter also can be very much, Pondization operation can be carried out to one layer of neutral net, i.e., it is down-sampled to input data progress, so both reduce calculation amount, also increase Strong generalization ability.As shown in the figure, more numbers in the scope of pond are synthesized into a number, it is most common to have maximum pond, be averaged Chi Hua, summation pond etc..The pond layer of this model takes the maximum of pond scope using maximum pond layer, the i.e. result in pond Number.If due to the shape of input data and the shape of pond scope, pond scope impartial cannot divide, can be appropriate With 0 cover, pondization operation is being carried out.
G, dropout processing
As shown in Fig. 8 (a) and Fig. 8 (b), to further improve the generalization ability of model, Hinton proposes once training During some neurons can be hidden according to certain probability, i.e., there is related parameter to be trained specifically with the neuron Updated in journey, keep the numerical value of last training, and in the last test stage, all neurons can all participate in calculating.
H, back-propagation algorithm
The schematic diagram of deep learning convolutional neural networks is as follows：
As shown in figure 9, by taking the output that convolutional layer only has 4 or 3 passages as an example, each passage of convolutional layer only with it is opposite The pond layer answered is attached, and pond layer is connected entirely with each passage of next layer of convolutional layer.Exemplified by above figure due to Input layer is single channel, and the convolutional layer being attached thereto has four passages, so the wave filter of convolutional layer has four, but each filter Ripple device only has a convolution kernel.And the convolutional layer of the second layer is three passages, the pond layer of preceding layer has four passages, so should Layer convolutional layer have three wave filters, each wave filter has four convolution kernels, respectively with each passage of the pond layer of preceding layer Into convolution.
Last layer of model is the output layer of model, and the result of output is the classification to signal modulation mode, such as at this In Modulation Mode Recognition model, preferable classification results, are represented with five dimension unit vectors respectively, such as use respectively (1,0,0,0, 0), (0,1,0,0,0) ..., (0,0,0,0,1) represent BPSK, QPSK ..., DSSS etc., but not limited to this.Herein in connection with square Battle array, the subscript of vectorial element are started counting up from 1, consistent to the way of reference of matrix-vector element with MATLAB.If output ForThe result of output layer is directly obtained by the output of the full articulamentum of preceding layer, if last layer is complete Articulamentum is L layers, and the output of jth neuron is usedRepresent.Since the initial value of model parameter is random, outputCan be withThere is deviation, therefore bias vector of classifying isThe norm of defining classification deviation is error in classification
Assuming that using two norms, then
It is the function of each layer parameter of neutral net, therefore E is also the function of each layer parameter, to make category of model optimal, E should be made minimum.Optimized parameter is scanned for along negative gradient direction on the Θ of parameter state space, step-size in search ξ, generally It is expressed as
For this deep learning convolutional neural networks model, the partial derivative that last layer exports conveniently is tried to achieve,
A) the connection weight matrix W and threshold value of full articulamentum
OrderN, i is the number and sequence number of neuron in L-1 layers of neutral net,Represent L- The i-th neuron is connected to the connection weight of jth neuron in L layers in 1 layer, then can be obtained by the chain rule of partial derivative
When obtainingFormula that can be more than is tried to achieveAnd so on, successively successively decrease, try to achieve full articulamentum Parameter.
When the input of full articulamentum is the output of convolutional layer, the multichannel two dimension output of convolutional layer will be rearranged for Input of the one-dimensional vector as full articulamentum.
B) the connection weight matrix W and threshold value of convolutional layer
As shown in Figure 10, in convolutional layer, due to introducing, we by pond, standardize and the calculating of convolutional layer combine Consider to seek the partial derivative of each parameter, also therefore, in derivation formula when representing the number of plies of neutral net, convolutional layer and thereafter After the ordinal number of pond layer all represented with l.
Assuming that the partial derivative of l layers of outputIt has been obtained that, be listed below and other variables are asked by chain type Rule for derivation The process of partial derivative
Due to pond process the output variable of convolutional layer has been carried out it is down-sampled, so the number of neuron can be reduced, because This is when the parameter of recursive convolution layer is propagated in direction, it is necessary to by the method for interpolation to carrying out a liter sampling.It can be adopted to simplify calculating With the simplest liter of method of sampling, i.e. the partial derivative of the output of all neurons in same pond scope is identical, is pond Partial derivative after change, i.e.,
Wherein k represents kth passage, jpThe output sequence number for being jth neuron in the layer of pond.Then convolutional layer can be asked to join Number,
For the output of standardization processing, the partial derivative of the parameter in standardization processing is then sought
Wherein J is the number of the neuron of each passage in convolutional layer.Since in convolutional layer, the output on upper strata is not with working as All neurons of front layer are all connected with, and the weights of convolution kernel are shared,Represent the kth wave filter of l layers of neutral net P-channel convolution kernel weights.The output of l-1 layers of p-channelPartial derivativeIt only have impact on and transported with convolution Related partial output resultsG be convolution kernel in weights sequence number, MiRepresenting can be with the in convolution algorithm The weights that i-th neuron of l-1 layers of pth passage is connected, λ represent the element caused by the shape, zero padding situation of convolution kernel Subscript deviation.
So far, can be by for the partial derivative of the parameter of l layers of convolutional layer neutral netDeduce, and can be released The output partial derivative of l-1 layers of each passage
The computational methods of the partial derivative of the parameters of full articulamentum and convolutional layer with reference to more than, and so on, according to 9 schematic diagram of attached drawing, the partial derivative of each parameter of preceding layer is deduced by later layer, so as to try to achieve entire depth study convolutional neural networks The partial derivative of model, finally, by the negative gradient direction undated parameter of parameter space, this is back-propagation algorithm.
2nd, different waters experimental data is verified
In order to verify the validity of system and robustness, underwater sound identification is carried out using different waters communication system data and has tested Confirmation is tested.Data are 2kHz or 4kHz from ground, signal bandwidths such as Bayuquan, Dalian and Song Hua River.For enhancing recognition methods Generalization ability, the signal of communication of each modulating mode under different channels, parameter is mixed to deep learning convolutional Neural Network is trained, identifies that the design parameter of Fig. 2 (a) each signals of communication into Fig. 2 (e) is as shown in table 1：
The modulation parameter of 1 experimental signal of table
Training result and experimental results such as Figure 11 (a) and Figure 11 (b) are shown, utilize deep learning convolutional neural networks Method identification BPSK, QPSK and spread-spectrum signal can reach 100% discrimination.Based on we it can be seen from test result The signal modulate of method, can reach for binary phase shift keying (BPSK) signal, Direct Sequence Spread Spectrum (DSSS) signal To 100% identification, the recognition correct rate of octal system phase-shift keying (8PSK) signal and Orthogonal Frequency Division Multiplexing (OFDM) modulated signal Also above 90%.Therefore, result above exists demonstrating the deep learning convolutional neural networks structure that designs herein to a certain degree Solve the problems, such as the validity in the Modulation Identification of water sound communication signal.

Claims (5)

  1. A kind of 1. underwater sound communication signal type Identification method based on deep learning technology, it is characterised in that：Comprise the following steps：
    Step 1：Build a deep learning convolutional neural networks model；
    Step 2：Default training set specimen discerning accuracy rate T, presets test set specimen discerning accuracy rate P；
    Step 3：Obtain the experimental data or emulation data of the underwater sound communication modulated signal of different modulating mode；
    Step 4：Using every N number of sample point data of similar modulation system signal as a primary data sample, each is original Data sample is the column vector of N × 1, to the modulation system of primary data sample mark signal；
    Step 5：Each data sample is pre-processed, all primary data samples are converted into deep learning convolutional Neural The input form of network, including bandpass filtering is carried out to primary data sample, the column vector of N × 1 is obtained, then to filtering The column vector of N × 1 obtained afterwards carries out Hilbert transform, the column vector of another N × 1 is obtained, by the column vector of 2 N × 1 It is merged into the matrix of 1 N × 2；
    Step 6：It is training set and test set by pretreated data sample matrix random division；
    Step 7：With the parameter of training sample set training deep learning convolutional neural networks model；
    Step 8：Whether training of judgement collection specimen discerning accuracy rate reaches preset value, when training set specimen discerning accuracy rate is more than During equal to T, step 9 is jumped to；Otherwise, step 7 is jumped to；
    Step 9：The input of deep learning convolutional neural networks is switched to data sample test set, with data sample test set Test to deep learning convolutional neural networks model；
    Step 10：Judge whether test set specimen discerning accuracy rate reaches preset value, when test set specimen discerning accuracy rate is more than During equal to P, then step 11 is performed；Otherwise, the former data mixing in excessive data, with the step three is obtained, is expanded former Beginning data sample, then jumps to step 4；
    Step 11：The setting of deep learning convolutional neural networks parameter is completed, obtains network model parameter, completes model.
  2. 2. a kind of underwater sound communication signal type Identification method based on deep learning technology according to claim 1, it is special Sign is：The output of deep learning convolutional neural networks model described in step 1 is that the data sample tune of acquisition is judged through model Mode processed, to being exported after the advanced row standardization processing of each layer of output result.
  3. 3. a kind of underwater sound communication signal type Identification method based on deep learning technology according to claim 1, it is special Sign is：It is back-propagation algorithm to be trained method with training sample set pair deep learning convolutional neural networks in step 7.
  4. 4. a kind of underwater sound communication signal type Identification method based on deep learning technology according to claim 1, it is special Sign is：Primary data sample described in step 4 meets：
    (N-1) × t × B ＞ 40
    Wherein, t is the sampling period, and B is the signal band width of the experimental data and emulation data described in step 3.
  5. 5. a kind of underwater sound communication signal type Identification method based on deep learning technology according to claim 1, it is special Sign is：Deep learning convolutional neural networks model described in step 1 includes convolutional layer, pond layer and full articulamentum, described Roll up the basic unit number of plies and be more than or equal to 3 layers, the port number of each convolutional layer should be all higher than being equal to 30, each convolutional layer it is every The neuron number of a passage successively successively decreases with convolutional layer, and channel number is successively incremented by；Pond is added behind each layer of convolutional layer Layer, pond layer are maximum pond layer, and the port number of pond layer is equal with the port number of convolutional layer before；The full articulamentum After being connected to last pond layer, the full articulamentum number of plies is more than or equal to 2 layers and is single channel, the full articulamentum of first layer The number of neuron is more than or equal to 100, and the number of the neuron of last layer of full articulamentum is more than or equal to 30, the full connection The neuron number of layer successively successively decreases.
CN201711443626.XA 2017-12-27 2017-12-27 A kind of underwater sound communication signal type Identification method based on depth learning technology Pending CN108038471A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711443626.XA CN108038471A (en) 2017-12-27 2017-12-27 A kind of underwater sound communication signal type Identification method based on depth learning technology

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711443626.XA CN108038471A (en) 2017-12-27 2017-12-27 A kind of underwater sound communication signal type Identification method based on depth learning technology

Publications (1)

Publication Number Publication Date
CN108038471A true CN108038471A (en) 2018-05-15

Family

ID=62097943

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711443626.XA Pending CN108038471A (en) 2017-12-27 2017-12-27 A kind of underwater sound communication signal type Identification method based on depth learning technology

Country Status (1)

Country Link
CN (1) CN108038471A (en)

Cited By (37)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109299697A (en) * 2018-09-30 2019-02-01 泰山学院 Deep neural network system and method based on underwater sound communication Modulation Mode Recognition
CN109361635A (en) * 2018-11-23 2019-02-19 泰山学院 Subsurface communication Modulation Mode Recognition method and system based on depth residual error network
CN109409442A (en) * 2018-11-21 2019-03-01 电子科技大学 Convolutional neural networks model selection method in transfer learning
CN109462564A (en) * 2018-11-16 2019-03-12 泰山学院 Subsurface communication Modulation Mode Recognition method and system based on deep neural network
CN109474352A (en) * 2018-12-24 2019-03-15 哈尔滨工程大学 A kind of underwater sound orthogonal frequency division multiplexing communication method based on deep learning
CN109495214A (en) * 2018-11-26 2019-03-19 电子科技大学 Channel coding type recognition methods based on one-dimensional Inception structure
CN109523023A (en) * 2018-11-16 2019-03-26 泰山学院 A kind of deep learning network and system for subsurface communication Modulation Identification
CN109547374A (en) * 2018-11-23 2019-03-29 泰山学院 A kind of depth residual error network and system for subsurface communication Modulation Identification
CN109660475A (en) * 2018-12-19 2019-04-19 南京世海声学科技有限公司 A kind of non-cooperation phase code water sound communication signal autonomous identifying method
CN109787929A (en) * 2019-02-20 2019-05-21 深圳市宝链人工智能科技有限公司 Signal modulate method, electronic device and computer readable storage medium
CN109886287A (en) * 2019-01-04 2019-06-14 平安科技（深圳）有限公司 Insecticide determines method, apparatus, computer equipment and storage medium
CN109951554A (en) * 2019-03-25 2019-06-28 北京理工大学 Information security technology contest anti-cheat method in real time
CN110162995A (en) * 2019-04-22 2019-08-23 阿里巴巴集团控股有限公司 Assess the method and device thereof of contribution data degree
CN110163282A (en) * 2019-05-22 2019-08-23 西安电子科技大学 Modulation Mode Recognition method based on deep learning
CN110489661A (en) * 2019-07-24 2019-11-22 武汉大学 A kind of social networks prediction technique based on generation confrontation network and transfer learning
CN110555341A (en) * 2018-05-31 2019-12-10 北京深鉴智能科技有限公司 Pooling method and apparatus, detection method and apparatus, electronic device, storage medium
CN110567558A (en) * 2019-08-28 2019-12-13 华南理工大学 Ultrasonic guided wave detection method based on deep convolution characteristics
CN111131108A (en) * 2019-12-30 2020-05-08 哈尔滨工程大学 Non-cooperative underwater acoustic OFDM subcarrier modulation mode identification method
CN111242235A (en) * 2020-01-19 2020-06-05 中国科学院计算技术研究所厦门数据智能研究院 Similar characteristic test data set generation method
CN111259861A (en) * 2020-02-18 2020-06-09 西北工业大学 Underwater acoustic communication signal modulation mode identification method under data set unbalanced condition
WO2020113464A1 (en) * 2018-12-05 2020-06-11 唐山照澜海洋科技有限公司 5g multi-carrier underwater acoustic communication method
CN111460932A (en) * 2020-03-17 2020-07-28 哈尔滨工程大学 Underwater sound signal classification and identification method based on self-adaptive convolution
CN111476311A (en) * 2020-04-20 2020-07-31 江苏科技大学 Anchor chain flash welding quality online detection method based on incremental learning
CN112259119A (en) * 2020-10-19 2021-01-22 成都明杰科技有限公司 Music source separation method based on stacked hourglass network
CN112464837A (en) * 2020-12-03 2021-03-09 中国人民解放军战略支援部队信息工程大学 Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples
CN112488092A (en) * 2021-02-05 2021-03-12 中国人民解放军国防科技大学 Navigation frequency band signal type identification method and system based on deep neural network
CN112615804A (en) * 2020-12-12 2021-04-06 中国人民解放军战略支援部队信息工程大学 Short burst underwater acoustic communication signal modulation identification method based on deep learning
CN112733811A (en) * 2020-09-23 2021-04-30 青岛科技大学 Underwater sound signal modulation mode inter-class identification method based on improved dense neural network
CN113259288A (en) * 2021-05-05 2021-08-13 青岛科技大学 Underwater acoustic communication modulation mode identification method based on feature fusion and lightweight hybrid neural network
CN113408401A (en) * 2021-06-16 2021-09-17 中国科学院南海海洋研究所 Method and device for quickly and automatically identifying ship traveling wave based on machine learning
CN113453153A (en) * 2020-03-27 2021-09-28 威海北洋电气集团股份有限公司 Wireless positioning fusion method and system based on deep learning
CN113709075A (en) * 2021-07-27 2021-11-26 天津大学 Method for realizing underwater acoustic communication receiver by using underwater acoustic channel multipath effect
CN113780468A (en) * 2021-09-28 2021-12-10 中国人民解放军国防科技大学 Robust model training method based on small number of neuron connections
CN113959071A (en) * 2021-07-21 2022-01-21 北京金茂绿建科技有限公司 Machine learning assistance-based operation control optimization method for centralized water chilling unit air conditioning system
CN114065822A (en) * 2021-11-30 2022-02-18 中国海洋大学 Electromagnetic identification method and system for ocean tide fluctuation
CN114422311A (en) * 2022-01-26 2022-04-29 中国人民解放军战略支援部队信息工程大学 Signal modulation identification method and system combining deep neural network and expert prior characteristics
CN117219125A (en) * 2023-11-07 2023-12-12 青岛科技大学 Marine mammal sound signal imitation hidden scoring method based on audio fingerprint

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104143115A (en) * 2014-06-25 2014-11-12 中国矿业大学(北京) Technological method for achieving soil water content classified identification through geological radar technology
CN106096616A (en) * 2016-06-08 2016-11-09 四川大学华西医院 A kind of nuclear magnetic resonance image feature extraction based on degree of depth study and sorting technique

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104143115A (en) * 2014-06-25 2014-11-12 中国矿业大学(北京) Technological method for achieving soil water content classified identification through geological radar technology
CN106096616A (en) * 2016-06-08 2016-11-09 四川大学华西医院 A kind of nuclear magnetic resonance image feature extraction based on degree of depth study and sorting technique

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
周启明等: "深度学习技术在水声通信体制识别中的应用", 《声学技术》 *
王青龙: "数字通信信号调制方式自动识别研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 *

Cited By (62)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110555341A (en) * 2018-05-31 2019-12-10 北京深鉴智能科技有限公司 Pooling method and apparatus, detection method and apparatus, electronic device, storage medium
CN109299697A (en) * 2018-09-30 2019-02-01 泰山学院 Deep neural network system and method based on underwater sound communication Modulation Mode Recognition
CN109462564A (en) * 2018-11-16 2019-03-12 泰山学院 Subsurface communication Modulation Mode Recognition method and system based on deep neural network
CN109462564B (en) * 2018-11-16 2021-08-03 泰山学院 Underwater communication modulation mode identification method and system based on deep neural network
CN109523023A (en) * 2018-11-16 2019-03-26 泰山学院 A kind of deep learning network and system for subsurface communication Modulation Identification
CN109409442A (en) * 2018-11-21 2019-03-01 电子科技大学 Convolutional neural networks model selection method in transfer learning
CN109361635A (en) * 2018-11-23 2019-02-19 泰山学院 Subsurface communication Modulation Mode Recognition method and system based on depth residual error network
CN109361635B (en) * 2018-11-23 2021-06-04 泰山学院 Underwater communication modulation mode identification method and system based on depth residual error network
CN109547374A (en) * 2018-11-23 2019-03-29 泰山学院 A kind of depth residual error network and system for subsurface communication Modulation Identification
CN109547374B (en) * 2018-11-23 2021-11-23 泰山学院 Depth residual error network and system for underwater communication modulation recognition
CN109495214A (en) * 2018-11-26 2019-03-19 电子科技大学 Channel coding type recognition methods based on one-dimensional Inception structure
WO2020113464A1 (en) * 2018-12-05 2020-06-11 唐山照澜海洋科技有限公司 5g multi-carrier underwater acoustic communication method
CN109660475A (en) * 2018-12-19 2019-04-19 南京世海声学科技有限公司 A kind of non-cooperation phase code water sound communication signal autonomous identifying method
CN109474352A (en) * 2018-12-24 2019-03-15 哈尔滨工程大学 A kind of underwater sound orthogonal frequency division multiplexing communication method based on deep learning
CN109474352B (en) * 2018-12-24 2021-07-06 三亚哈尔滨工程大学南海创新发展基地 Underwater sound orthogonal frequency division multiplexing communication method based on deep learning
CN109886287A (en) * 2019-01-04 2019-06-14 平安科技（深圳）有限公司 Insecticide determines method, apparatus, computer equipment and storage medium
CN109886287B (en) * 2019-01-04 2024-05-14 平安科技（深圳）有限公司 Pesticide determination method, device, computer equipment and storage medium
CN109787929A (en) * 2019-02-20 2019-05-21 深圳市宝链人工智能科技有限公司 Signal modulate method, electronic device and computer readable storage medium
CN109951554A (en) * 2019-03-25 2019-06-28 北京理工大学 Information security technology contest anti-cheat method in real time
CN110162995A (en) * 2019-04-22 2019-08-23 阿里巴巴集团控股有限公司 Assess the method and device thereof of contribution data degree
CN110162995B (en) * 2019-04-22 2023-01-10 创新先进技术有限公司 Method and device for evaluating data contribution degree
CN110163282A (en) * 2019-05-22 2019-08-23 西安电子科技大学 Modulation Mode Recognition method based on deep learning
CN110163282B (en) * 2019-05-22 2022-12-06 西安电子科技大学 Modulation mode identification method based on deep learning
CN110489661A (en) * 2019-07-24 2019-11-22 武汉大学 A kind of social networks prediction technique based on generation confrontation network and transfer learning
CN110489661B (en) * 2019-07-24 2022-04-26 武汉大学 Social relationship prediction method based on generation of confrontation network and transfer learning
CN110567558A (en) * 2019-08-28 2019-12-13 华南理工大学 Ultrasonic guided wave detection method based on deep convolution characteristics
CN110567558B (en) * 2019-08-28 2021-08-10 华南理工大学 Ultrasonic guided wave detection method based on deep convolution characteristics
CN111131108B (en) * 2019-12-30 2022-05-20 哈尔滨工程大学 Non-cooperative underwater sound OFDM subcarrier modulation mode identification method
CN111131108A (en) * 2019-12-30 2020-05-08 哈尔滨工程大学 Non-cooperative underwater acoustic OFDM subcarrier modulation mode identification method
CN111242235A (en) * 2020-01-19 2020-06-05 中国科学院计算技术研究所厦门数据智能研究院 Similar characteristic test data set generation method
CN111242235B (en) * 2020-01-19 2023-04-07 中科(厦门)数据智能研究院 Similar characteristic test data set generation method
CN111259861A (en) * 2020-02-18 2020-06-09 西北工业大学 Underwater acoustic communication signal modulation mode identification method under data set unbalanced condition
CN111460932B (en) * 2020-03-17 2022-06-21 哈尔滨工程大学 Underwater sound signal classification and identification method based on self-adaptive convolution
CN111460932A (en) * 2020-03-17 2020-07-28 哈尔滨工程大学 Underwater sound signal classification and identification method based on self-adaptive convolution
CN113453153B (en) * 2020-03-27 2022-05-17 威海北洋电气集团股份有限公司 Wireless positioning fusion method and system based on deep learning
CN113453153A (en) * 2020-03-27 2021-09-28 威海北洋电气集团股份有限公司 Wireless positioning fusion method and system based on deep learning
CN111476311A (en) * 2020-04-20 2020-07-31 江苏科技大学 Anchor chain flash welding quality online detection method based on incremental learning
CN111476311B (en) * 2020-04-20 2023-05-26 江苏科技大学 Anchor chain flash welding quality online detection method based on increment learning
CN112733811A (en) * 2020-09-23 2021-04-30 青岛科技大学 Underwater sound signal modulation mode inter-class identification method based on improved dense neural network
CN112259119B (en) * 2020-10-19 2021-11-16 深圳市策慧科技有限公司 Music source separation method based on stacked hourglass network
CN112259119A (en) * 2020-10-19 2021-01-22 成都明杰科技有限公司 Music source separation method based on stacked hourglass network
CN112464837B (en) * 2020-12-03 2023-04-07 中国人民解放军战略支援部队信息工程大学 Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples
CN112464837A (en) * 2020-12-03 2021-03-09 中国人民解放军战略支援部队信息工程大学 Shallow sea underwater acoustic communication signal modulation identification method and system based on small data samples
CN112615804A (en) * 2020-12-12 2021-04-06 中国人民解放军战略支援部队信息工程大学 Short burst underwater acoustic communication signal modulation identification method based on deep learning
CN112488092B (en) * 2021-02-05 2021-08-24 中国人民解放军国防科技大学 Navigation frequency band signal type identification method and system based on deep neural network
CN112488092A (en) * 2021-02-05 2021-03-12 中国人民解放军国防科技大学 Navigation frequency band signal type identification method and system based on deep neural network
CN113259288A (en) * 2021-05-05 2021-08-13 青岛科技大学 Underwater acoustic communication modulation mode identification method based on feature fusion and lightweight hybrid neural network
CN113259288B (en) * 2021-05-05 2023-08-08 青岛科技大学 Underwater sound modulation mode identification method based on feature fusion and lightweight hybrid model
CN113408401B (en) * 2021-06-16 2022-02-22 中国科学院南海海洋研究所 Method and device for quickly and automatically identifying ship traveling wave based on machine learning
CN113408401A (en) * 2021-06-16 2021-09-17 中国科学院南海海洋研究所 Method and device for quickly and automatically identifying ship traveling wave based on machine learning
CN113959071A (en) * 2021-07-21 2022-01-21 北京金茂绿建科技有限公司 Machine learning assistance-based operation control optimization method for centralized water chilling unit air conditioning system
CN113959071B (en) * 2021-07-21 2023-05-26 北京金茂绿建科技有限公司 Centralized water chilling unit air conditioning system operation control optimization method based on machine learning assistance
CN113709075B (en) * 2021-07-27 2022-08-05 天津大学 Method for realizing underwater acoustic communication receiver by using underwater acoustic channel multipath effect
CN113709075A (en) * 2021-07-27 2021-11-26 天津大学 Method for realizing underwater acoustic communication receiver by using underwater acoustic channel multipath effect
CN113780468B (en) * 2021-09-28 2022-08-09 中国人民解放军国防科技大学 Robust image classification model training method based on small number of neuron connections
CN113780468A (en) * 2021-09-28 2021-12-10 中国人民解放军国防科技大学 Robust model training method based on small number of neuron connections
CN114065822B (en) * 2021-11-30 2024-04-26 中国海洋大学 Electromagnetic identification method and system for ocean tide fluctuation
CN114065822A (en) * 2021-11-30 2022-02-18 中国海洋大学 Electromagnetic identification method and system for ocean tide fluctuation
CN114422311A (en) * 2022-01-26 2022-04-29 中国人民解放军战略支援部队信息工程大学 Signal modulation identification method and system combining deep neural network and expert prior characteristics
CN114422311B (en) * 2022-01-26 2023-08-18 中国人民解放军战略支援部队信息工程大学 Signal modulation recognition method and system combining deep neural network and expert priori features
CN117219125A (en) * 2023-11-07 2023-12-12 青岛科技大学 Marine mammal sound signal imitation hidden scoring method based on audio fingerprint
CN117219125B (en) * 2023-11-07 2024-01-30 青岛科技大学 Marine mammal sound signal imitation hidden scoring method based on audio fingerprint

Similar Documents

Publication Publication Date Title
CN108038471A (en) A kind of underwater sound communication signal type Identification method based on depth learning technology
CN107979554B (en) Radio signal Modulation Identification method based on multiple dimensioned convolutional neural networks
Nourani et al. Daily and monthly suspended sediment load predictions using wavelet based artificial intelligence approaches
CN108416755A (en) A kind of image de-noising method and system based on deep learning
CN109299697A (en) Deep neural network system and method based on underwater sound communication Modulation Mode Recognition
CN106446942A (en) Crop disease identification method based on incremental learning
CN106127217A (en) The method and system that neutral net detects is goed deep into for anatomical object for approximation
CN109785344A (en) The remote sensing image segmentation method of binary channel residual error network based on feature recalibration
CN110456332A (en) A kind of underwater sound signal Enhancement Method based on autocoder
CN106600595A (en) Human body characteristic dimension automatic measuring method based on artificial intelligence algorithm
CN110532932A (en) A kind of multi -components radar emitter signal intra-pulse modulation mode recognition methods
CN111523713A (en) Method and device for predicting residual oil saturation distribution in oil field
CN105913117A (en) Intelligent related neural network computer identification method
CN110048978A (en) A kind of signal modulate method
Shoaib et al. A wavelet based approach for combining the outputs of different rainfall–runoff models
Roushangar et al. Identifying complexity of annual precipitation variation in Iran during 1960–2010 based on information theory and discrete wavelet transform
CN113673312A (en) Radar signal intra-pulse modulation identification method based on deep learning
Han et al. Spatial steganalysis based on non-local block and multi-channel convolutional networks
Gan Signal processing and image processing for acoustical imaging
CN113743317A (en) License plate picture generation method, medium and equipment based on generation countermeasure network
Pandhiani et al. Time series forecasting by using hybrid models for monthly streamflow data
Niu et al. Deep learning for ocean acoustic source localization using one sensor
Al Iqbal Eclectic rule extraction from Neural Networks using aggregated Decision Trees
CN116430317A (en) Radiation source modulation pattern and individual identification method and system
CN115470818A (en) High-order modulation signal pattern recognition method and system based on variational modal decomposition

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20180515