CN104811276B - A kind of DL CNN demodulators of super Nyquist rate communication - Google Patents
A kind of DL CNN demodulators of super Nyquist rate communication Download PDFInfo
- Publication number
- CN104811276B CN104811276B CN201510220785.8A CN201510220785A CN104811276B CN 104811276 B CN104811276 B CN 104811276B CN 201510220785 A CN201510220785 A CN 201510220785A CN 104811276 B CN104811276 B CN 104811276B
- Authority
- CN
- China
- Prior art keywords
- cnn
- filter
- symbol
- nyquist rate
- demodulators
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L1/00—Arrangements for detecting or preventing errors in the information received
- H04L1/004—Arrangements for detecting or preventing errors in the information received by using forward error control
- H04L1/0056—Systems characterized by the type of code used
- H04L1/0059—Convolutional codes
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Digital Transmission Methods That Use Modulated Carrier Waves (AREA)
Abstract
The invention discloses a kind of DL CNN (deep learning convolutional neural networks) demodulator for super Nyquist rate communication.Digital modulation signals for being even more than the limitation of Nyquist rate bandwidth by strict bandwidth efficient channel, the demodulator is using convolutional neural networks (CNN) after the signal characteristic including receiving terminal is directly extracted comprising intersymbol interference from by shock filter or the filtered reception signal sample of normal bandpass filters, convolutional neural networks are trained using deep learning (DL), and use single element or multiple-symbol cascading judgement, DL CNN are classified under stronger intersymbol interference environment to reception signal sample, so as to realize the anti-ISI demodulation for super Nyquist rate modulation signal, and there are more excellent demodulation performance and very strong adaptability than the amplitude integration judgement demodulator of routine.
Description
Technical field
The present invention relates to the communications field, intersymbol interference modulation when the character rate particularly to communicate exceedes Nyquist rate
The judgement demodulation problem of signal.
Background technology
With the continuous development of information technology, wide-band mobile communication and smart mobile phone, not only resulting in radio-frequency spectrum turns into
Scarce resource, and the annual power consumption of communication base station to spread all over the country, also above over ten billion degree, therefore, advanced information society compels to be essential
Seek the information transfer system with high spectrum utilization and high-energy utilization rate.
1. super Nyquist speed
In digital communicating field, baud rate is modulation rate, refers to the speed of valid data signal modulation carrier wave, can manage
The number by transmitting symbol (also referred to as symbol) in the unit time is solved, unit is Symbol/s or baud (Baud or Bd).And how
Qwest (Nyquist) theorem points out, the chip rate of particular channel can not possibly exceed 2 times of low pass channel bandwidth, thus
2Bd/Hz is known as Nyquist rate.Existing communication system is intended to improve the rate of information throughput (the i.e. bit in per unit band
Rate, unit are bit/s or bps), the number that each communication symbol is quantized can only be increased, that is, introduce high order modulation.This is
The common knowledge of communication technical field and classical technology, it is associated and there are following two common knowledges：
1) Nyquist rate is the theoretical maximum transmission rate under noiseless state, and actual physics channel has unavoidably
Various interference, therefore channel capacity will be restricted by Shannon (Shannon) formula；
2) chip rate exceedes Nyquist rate and can not communicated, but intersymbol is bound to produce intersymbol interference
(ISI), existing optimal receiver model and correlation demodulation are theoretical, on the premise of being all built upon no intersymbol interference.
Therefore, the common knowledge based on this area, several conclusions are also obvious below：
1) communication that Nyquist rate modulated signal is carried out by the narrower channel of bandwidth, is exactly that super Nyquist speed is led to
Letter (chip rate that communicates exceedes Nyquist rate)；
2) super Nyquist rate communication can directly improve the availability of frequency spectrum；
3) apply the narrower bandpass filtering of bandwidth (referred to as " super bandlimiting filtering ") to Nyquist rate modulated signal, produce
Super Nyquist rate signal；
4) for same chip rate, transmission can use narrower channel with connecing with receiving super Nyquist rate signal
Receipts machine bandwidth, this helps to reduce receiver noise factor, improves receiver sensitivity, and be expected under same transmission power
Obtain higher received signal to noise ratio (SNR), or farther communication distance；
5) key of super Nyquist rate communication is realized, is correctly to demodulate intersymbol interference signal, and prior art
It is generally divided into two steps：Intersymbol interference is eliminated by technological means such as channel equalization, lifterings first, recovers normal modulation signal；
Demodulation is completed using conventional method again.
In a word, super Nyquist rate modulation signal is substantially because normal speed modulated signal is by narrow band logical
The filtering of (carrier (boc) modulated signals) or low pass (baseband signal) limits and loses the result of high fdrequency component, and existing reception processing skill
Art eliminates intersymbol interference using channel equalization or liftering first, and it is (right using an equivalent band resistance to be sought in frequency domain
In carrier (boc) modulated signals) or the next relative high fdrequency component for compensating reception signal of high pass (for baseband signal) wave filter, this can not
Also it can correspondingly lift out-of-band noise with avoiding, cause the deterioration of received signal to noise ratio before demodulation.Thus it is possible to directly demodulation it is super how
The high-performance demodulator of Nyquist rate modulated signal, it is that can lift the communication system availability of frequency spectrum and capacity usage ratio simultaneously
Key.
2. deep learning-convolutional neural networks (DL-CNN)
2006, University of Toronto professor Geoffrey Hinton existed《Science》A pass has been delivered on magazine
In the article of more hidden layer deep neural networks, deep learning (Deep learning, abbreviation DL) is opened in academia and industry
The research tide on boundary.
1) deep learning is a branch of machine learning, is mainly characterized by obtaining for original by multi-level study
The expression of beginning data difference level of abstraction, and then improve the accuracy of the task such as classification and prediction.Such as there is a pile input I (such as
The signal gathered under a pile varying environment), it is assumed that we devise the system S of a n-layer, by the parameter in adjustment system,
So that its output is still input I, then can automatically derives a series of input I level characteristics, i.e. S1 ..., Sn.
Therefore, the shallow-layer learning algorithms such as traditional SVM (SVM) are different from, DL need not be special by artificial experience sample drawn
Sign, but by building the training data (utilizing " big data ") of machine learning model and magnanimity with many hidden layers, come
Automatically more useful feature is learnt, so as to the accuracy that finally lifting is classified or predicted.
2) convolutional neural networks (Convolutional Neural Networks, abbreviation CNN) are artificial neural networks
One kind, its weights share network structure and are allowed to be more closely similar to biological neural network, reduce the complexity of network model, reduce
Weights numbers.CNN is first learning algorithm for really successfully training multitiered network structure, and its utilization space relation, which is reduced, to be needed
The parameter to be learnt, to improve the training performance of general forward direction BP (backpropagation) algorithm.In CNN, sub-fraction data (office
Portion's receptive field) inputted as the lowermost layer of hierarchical structure, information is transferred to different layers successively again, and every layer passes through a numeral filter
Ripple device obtain observation data most significant feature, can obtain accordingly to translate, scale and the observation data of invariable rotary it is notable
Feature, because the local receptor field of data allows neuron or processing unit to may have access to most basic feature.Therefore, CNN
It is mainly characterized by：Convolution (carrying out local connection to node between layers by local receptor field), weights share and pond
(down-sampling).And in traditional BP neural network, each node layer is a linear one dimensional arrangement state, the network of layer and layer
Connected entirely between node.
In recent years, the convolutional neural networks based on deep learning (DL-CNN) are in pattern classification and identification particularly language
Sound identifies obtains immense success with fields such as image recognitions.CNN has self study and adaptive ability, can be by pre-
Potential rule between the two is grasped in a collection of mutually corresponding input, the output data progress feature extraction first provided, analysis, this
Kind process is referred to as " training " of network.For new data, according to train before come rule can directly carry out detection and sentence
Certainly.Therefore, in voice or image recognition, speech samples or image pixel can avoid biography directly as the input of network
Complicated feature extraction and data reconstruction processes in recognizer of uniting.
In Fig. 1, above one layer be m layers, below one layer be m-1 layers.As can be seen that each node on m layers
Only it is connected with 3 nodes of m-1 layers corresponding region, this subrange is also referred to as receptive field, by local connection, significantly
Reduce weights connection number, the net input for one node of Far Left on m layers, be equal to all with this node be connected
The last layer neuron node value connect is cumulative with the product of corresponding weights, and such calculating process is referred to as convolution.Weights are shared
Then refer to that all neuron nodes on a characteristic pattern all with same convolution nuclear phase convolution, have extracted a kind of feature,
If necessary to extract various features, then every layer just has multiple characteristic patterns.By taking image recognition as an example, in theory by imagery exploitation
Different convolution kernels has obtained multiple images afterwards by convolution, is then directly classified using these images, but so
Amount of calculation is too big, can be reduced data volume using pond (down-sampling) operation, while retain original figure to a certain extent
As feature.Here, the region in pond is nonoverlapping, and the receptive field of convolution is overlapping.
Using New York Univ USA Yann Le professors Cun propose based on CNN character identification system LeNet-5 as
Example, its network structure such as Fig. 2, a total of 8 layers including input layer and output layer.In convolution and sub-sampling procedures, convolution
Process includes：With a trainable wave filter fxDeconvolute an image inputted (first stage is original input picture,
Stage below is exactly convolution characteristic pattern), then plus one biases bx, obtain convolutional layer Cx；Sub-sampling procedures include：Per neighborhood 4
Individual pixel summation is changed into 1 pixel, then passes through scalar Wx+1Weighting, it is further added by biasing bx+1, then swashed by a Sigmoid
Function living, produce a Feature Mapping figure S for about reducing 4 timesx+1.Each feature extraction layer (C layers) followed by one in CNN
Individual to be used for seeking the computation layer of local average and second extraction (S layers), this distinctive structure of feature extraction twice is knowing network
There is higher distortion tolerance when other to input sample.
The content of the invention
Goal of the invention：In order to overcome the deficiencies in the prior art, the present invention provide one kind can directly demodulate it is super how
The high-performance demodulator of Nyquist rate modulated signal, for overcoming the existing communication system availability of frequency spectrum and capacity usage ratio
Low technical problem.
It was noted that：
1) demodulation of M systems symbol signal of communication is generally only the classification of most M symbols, thus is known in theory than image
The situation such as not is simply too much；
2) super Nyquist rate communication or bandwidth efficient channel are exactly convolution for the mechanism of transmission signal, thus CNN
This situation should be also suitable for；
3) influence of limitation and the intersymbol interference of channel width so that the front and rear symbol of super Nyquist rate modulation signal
Local correlation is provided between sample, the local correlations and CNN local receptor field there should be certain corresponding relation；
4) by selecting different signal to noise ratio and bandwidth (or super Nyquist speed), modulated pattern can form enough
" big data " carries out deep learning for CNN.
Therefore, convolutional neural networks are trained based on deep learning method, allow DL-CNN to learn and remember it is super how Kui
In this special rate modulation signal symbol feature and intersymbol interference pattern after, to the super Nyquist rate modulation of its input
Signal sampling value sequence carries out pattern classification and non-linear judgement, you can hopes and realizes for super Nyquist rate modulation signal
Correct demodulation.
Technical scheme：Based on above-mentioned thinking, the technical solution adopted by the present invention is：
A kind of DN-CNN demodulators of super Nyquist rate communication, including receiving filter and DL-CNN classification judgements
Device, said two devices cascade；The DL-CNN classification decision device is made up of the convolutional neural networks grader of multilayer, the convolution god
Depth is carried out in the communication sample under with different channel width, signal to noise ratio and intersymbol interference environment through network classifier
Practise, the wave character and internal association of the modulated signal filter response of extraction and memory with intersymbol interference are so as to completing to train.
The super Nyquist rate modulation device supporting with above-mentioned DN-CNN demodulators is in structure by modulator and super band limit
Wave filter is formed, and wherein modulator can be any system (such as phase-shift keying (PSK), frequency shift keying, quadrature amplitude modulation), (two enter system
System, multi-system) and speed conventional modulator, and super band limiting filter is alternatively any type of bandpass filter, simply
Modulated signal speed can super Nyquist speed controlled by its filtering bandwidth.Therefore, super Nyquist rate modulation device with it is general
Logical modulator has no difference in system architecture.
The present invention does not specialize in CNN in itself, but the CNN network structures for being successfully used for image recognition are carried out
It is suitably modified, it is allowed to be applied to the sample sequence of training super Nyquist rate signal, proposes the modulated signal based on DL-CNN
The judgement detection method of anti-ISI.Here CNN is trained for grader, the whole symbol received is identified.From
Deep learning is carried out in " big data " communication sample under different channel width, signal to noise ratio and intersymbol interference environment, is carried
Take and remember the wave character and internal association of the modulated signal filter response with intersymbol interference.CNN inputs for each
Symbol exports a numeral for indicating its classification, such as represents " 0 " symbol with 0, represents " 1 " symbol (until representing " M codes with M with 1
Member ", if for M ary modulation signals), then go control to export corresponding local standard symbol waveform with this numeral, i.e.,
It can complete to adjudicate.The CNN networks that the present invention uses are of five storeys altogether except input, output layer, train signal to noise ratio and training iteration time
It is several that different numerical value is set according to different environment.After training successfully, it is possible to the network succeeded in school to new input
Signal sample sequence make decisions.
Further, in the present invention, according to input signal, selection receives the type of wave filter, including shock filter
With two kinds of conventional band pass filter.
It is less than " asymmetric " modulated signal of code-element period for the modulates information period, selects receiving filter to be filtered for impact
Ripple device, this is a kind of bandpass filter of particular design, amplifies the waveform of asymmetric modulated signal using precipitous phase-frequency characteristic
Difference simultaneously lifts output signal-to-noise ratio, can protrude the modulation signature of signal.
For the situation of shock filter failure, it is conventional band pass filter to receive wave filter selection, suitable for appointing
The universal demodulation of what modulated signal, can preferably filter out out-of-band noise, lift the demodulation performance under strong intersymbol interference environment.
For example, a kind of extended binary phase shift keying of simplification (Extended Binary Phase Shift Keying, letter
Claiming EBPSK) modulated signal is defined as follows：
s0(t)=sin ωcT, 0≤t ＜ T
Wherein, s0And s (t)1(t) modulation waveform of symbol " 0 " and " 1 ", ω are represented respectivelycFor carrier angular frequencies；Symbol week
Phase T=2 π N/ ωcN >=1 carrier cycle is continue for, the π K/ ω of the modulation time span τ of " 1 " symbol=2cIt continue for the N number of loads of K ＜
Wave period, K and N are integer to ensure that complete cycle modulates.When τ=T be symbol " 0 " and " 1 " modulation duration it is equal when, (1)
Formula deteriorates to binary phase shift keying (BPSK) modulation of classics, it is seen that symmetric modulation BPSK is asymmetric modulation EBPSK spy
Example, thus this specification is discussed by taking (1) formula as an example and does not lose representativeness.
According to the definition of (1) formula, the number of winning the confidence carrier frequency is 30kHz and modulation parameter N=28, K=5, then by transmitting terminal
The time domain waveform of EBPSK signals after 2kHz bandpass filterings is as shown in Figure 4, it is seen that for information sequence " 11100010
10 ", intersymbol interference has caused EBPSK modulated signals waveform to produce more serious distortion and has spread to adjacent symbol.Filter front signal
Power spectrum such as Fig. 5 (a) shown in, its main lobe width is about 12kHz；Then such as Fig. 5 (b) is shown for the power spectrum of filtered signal, this
When 2kHz sending filter cut most of main lobe of modulated signal power spectrum.
Further, in the present invention, the DL-CNN classification decision device uses single element decision method or multiple-symbol
Cascading judgement method, the multiple-symbol cascading judgement method are that whole sampled values in multiple code-element periods combine sentencing
Certainly.
Illustrate for multiple-symbol cascading judgement method still by taking (1) formula as an example.Adjusted for the EBPSK after accepting filter
Signal processed, it was noted that：When channel width constriction, stronger intersymbol interference causes its adjacent symbol especially continuous
Impact envelope corresponding to " 1 " extends, is overlapping or even connected.Statistics between the symbol of (or hypothesis) required by classical communication is only
Vertical property condition no longer meets.Accordingly, the present invention is considered as multiple-symbol joint-detection mode and designs CNN judging modules, i.e. CNN divides
The input of class is the EBPSK impact filterings output waveform sampling corresponding to n symbol, and its output also just corresponds to this n symbol
The code character (i.e. continuous n positions) of composition：
1. it is traditional single element independent detection as n=1, its input is the waveform sampling corresponding to 1 symbol, and
Export then non-" 1 " i.e. " 0 "；
2. as n=2, it inputs the waveform sampling corresponding to adjacent 2 symbols, and it is then following 4 kind of 2 symbol to export
One of code character (2)：" 11 ", " 10 ", " 01 ", " 00 "；
3. as n=3, it inputs the waveform sampling corresponding to successive 3 symbols, and it is then following 8 kind of 3 symbol to export
One of code character (3)：" 000 ", " 001 ", " 010 ", " 011 ", " 100 ", " 101 ", " 110 ", " 11
1”.Remaining can the like.
Beneficial effect：
A kind of DN-CNN demodulators of super Nyquist rate communication provided by the invention have the beneficial of following several respects
Effect：
1st, capacity usage ratio is improved.
Advantage is from two aspects：
1) DL-CNN judgements detection fully learns and make use of global feature and the inherence of intersymbol interference modulated signal waveform
Information, thus than the simply simple threshold judgement using amplitude information, greatly improve the demodulation performance of receiver, and intersymbol
Interference is more serious, and the advantage that judgement demodulator is integrated compared to traditional amplitude is bigger；
2) for same chip rate, transmission can use narrower channel with connecing with receiving super Nyquist rate signal
Receipts machine bandwidth, this helps to reduce receiver noise factor, improves receiver sensitivity, and be expected under same transmission power
Obtain higher received signal to noise ratio.
And the lifting of receiver demodulation performance and sensitivity is all equivalent to the increase of transmission power, can extend communication away from
From；If keeping, former communication distance index is constant, and transmission power can reduce.This for reduce communication system energy resource consumption and
Electromagnetic pollution, extend battery life, realize " green communications ", there is practical significance.
2nd, the availability of frequency spectrum is improved.
The one-sided constriction signal spectrum bandwidth in modulator end no doubt can directly improve the availability of frequency spectrum, but may not be feasible,
Because intersymbol interference may make system performance degradation to losing use value；Unless demodulator end is in modulated signal bandwidth constriction
When, remain to make demodulation performance be maintained at acceptable level.And DL-CNN demodulators are using non-possessed by such multitiered network
Linearly, adaptability and deep learning ability, by classifying under stronger intersymbol interference to reception signal sample, so as to realize
Demodulated for the anti-ISI of modulated signal " speed-raising " modulated signal of starting so that lifting the availability of frequency spectrum has feasibility
And practical significance.
3rd, adaptability is enhanced.
1) deep learning ability of the DL-CNN demodulators under " big data " support, it is allow to remember more signals special
Seek peace the characteristic of channel, thus improve the generalization ability of general nonlinearity demodulator, there is stronger robust in real work
Property, it is wider to adapt to occasion；
2) using classical Threshold detection or amplitude integration judgement, usually taken near impact filtering output peak value some
Threshold judgement is carried out again after the point-by-point summation of sampling, and CNN judgements are the batch processings to all being sampled in a code-element period, even
It is the disposable cascading judgement to all being sampled in n code-element period, thus its required precision for sample-synchronous is far below
The former, the requirement to bit synchronization is also lower；
3) existing communication receiver for compensation or is eliminated because band limit is (even if not yet reach Nyquist rate or super Nai Kuisi
The band limit of special speed) caused by intersymbol interference, first have to carry out channel estimation, channel equalization, the technology such as liftering before demodulation
Processing, and the DL-CNN demodulators of the present invention can be saved or the step for " merging " by advance study, it also avoid because
The deterioration of signal to noise ratio before being demodulated caused by out-of-band noise lifting；
4) DL-CNN demodulators can be with on-line study, thus has with " constant " of demodulator network topology or hardware configuration
To adapt to the potentiality of " ten thousand become " of modulation system and the characteristic of channel, be advantageous to the normalization of demodulator and receiver, versatility, can
Customization and software radio are realized.
Brief description of the drawings
Fig. 1 is the local receptor field connection diagram in convolutional neural networks between layers.
Fig. 2 is the CNN network structures of a character identification system LeNet-5 based on CNN, including input layer output layer
A total of 8 layers inside.Including 3 convolutional layers, 2 down-sampling layers, 1 full articulamentum and input, output layer.
Fig. 3 is the super Nyquist rate communication system block diagram based on DL-CNN demodulation.
Fig. 4 is that carrier frequency is 30kHz, and when transmitting terminal bandwidth is limited to 2kHz, N=28, K=5 are anti-phase during 10 sampling rate
Time domain waveform of the EBPSK signals after transmitting terminal bandpass filter compares figure, and sequence of symhols is " 111000101
0”。
Fig. 5 is that carrier frequency is 30kHz, modulation parameter N=28, K=5, simplifies the work(of EBPSK modulated signals during 10 sampling rate
Rate is composed, and the information symbol sequence modulated is " 1110001010 ", and wherein Fig. 5 (a) is original modulated signal work(
Rate is composed；Fig. 5 (b) is that Fig. 5 (a) signals pass through the power spectrum after the transmitting terminal bandpass filter with a width of 2kHz.
Signal when Fig. 6 is noiseless in Fig. 5 (b) is used by the time domain waveform after receiving filter, wherein Fig. 6 (a)
Shock filter is as receiving filter, and Fig. 6 (b) is using FIR bandpass filters as receiving filter.
Fig. 7 is that single element, double is respectively adopted using shock filter processing, DL-CNN in transmitting terminal limit band 4kHz, receiving terminal
Bit error rate performance when symbol and 3 bit decision, and the bit error rate performance of classical amplitude integration judgement.
Fig. 8 is that transmitting terminal limit is handled with 2kHz, receiving terminal using common FIR bandpass filters, and list is respectively adopted in DL-CNN
Bit error rate performance when symbol, dicode member and 3 bit decision, and the bit error rate performance of classical amplitude integration judgement.
Fig. 9 is that transmitting terminal limit is respectively adopted at shock filter and common FIR bandpass filters with 2kHz, receiving terminal
Bit error rate performance when single element, dicode member and 3 bit decision is respectively adopted in reason, DL-CNN, and classical amplitude integration is sentenced
Bit error rate performance certainly.
Figure 10 is that transmitting terminal limit is handled with 1kHz, receiving terminal using common FIR bandpass filters, and list is respectively adopted in DL-CNN
Bit error rate performance when symbol, dicode member and 3 bit decision, and the bit error rate performance of classical amplitude integration judgement.
Embodiment
The present invention is further described below in conjunction with the accompanying drawings.
1. System describe
The specific embodiment of the present invention is to limit band EBPSK modulation letters for simplifying under additive white Gaussian noise (AWGN) channel
Number the demodulation of DL-CNN anti-ISIs, for its system block diagram referring to Fig. 3, the left side of channel is super Nyquist rate modulation device,
The right of channel is the DL-CNN demodulators of anti-ISI.
Here super Nyquist rate modulation device is made up of in structure modulator 1 and super band limiting filter 2, wherein modulating
Device 1 can be any system (such as phase-shift keying (PSK), frequency shift keying, quadrature amplitude modulation), system (binary system, multi-system) and speed is normal
Advise modulator, and super band limiting filter 2 be alternatively any type of bandpass filter, simply modulated signal speed can surpass how
Nyquist rate is controlled by its filtering bandwidth.Therefore, super Nyquist rate modulation device and generalised modulator are in system architecture
Have no difference.
DL-CNN demodulators include receiving filter and DL-CNN classification decision devices 6, said two devices cascade；The DL-CNN
Classification decision device 6 is made up of the convolutional neural networks grader of multilayer.
Transmitting terminal sequence of symhols first passes around modulator 1, realizes according to simplified EBPSK modulation defined in 1 formula；Then pass through
Super band limiting filter 2 is crossed, the band communication channel of Nyquist rate can be exceeded with analog bandwidth limitation；Finally it is superimposed by channel 3
Upper white Gaussian noise.
Receiving terminal provides two kinds of alternative lifting demodulation performances first：First, amplified " 0 " " 1 " by shock filter 4
Difference between symbol, and filter out part out-of-band noise；Second, out-of-band noise is filtered out by the FIR bandpass filters 5 of routine.So
The EBPSK reception signals after after filtering are sent into afterwards in the DL-CNN classification decision devices 6 trained, demodulated corresponding
Information symbol.Because EBPSK is binary modulated, thus information code current is directly obtained herein.
The specific implementation of description demodulator developed below.
2. receiving filter designs
Receiving filter both can with Digital Implementation, can also simulated implementation, only to facilitate simulation comparison, the present embodiment
Realized using digital filter.
1) digital shock filter (Digital Impacting Filters)
The digital shock filter 4 in Fig. 3 is that a kind of infinite-duration impulse response (IIR) bandpass filter (can also be adopted
With FIR filter), by a pair of conjugation zero points and at least two pairs of conjugate poles are formed, and zero point resonant frequency is less than input signal
Carrier frequency, pole frequency is all higher than the carrier frequency of input signal, and zero frequency can not be between pole frequency, zero point
Frequency and the close degree of pole frequency are not inferior to the 10 of input signal carrier frequency-3The order of magnitude, to ensure in filter passband
An extremely narrow trap-selecting frequency characteristic (or utilizing any precipitous phase-frequency characteristic) is showed at interior centre frequency so that
The filtering output waveform of asymmetric binary modulating signal produces parastic modulation impact at modulates information, and signal to noise ratio is lifted.
Principle and design on digital shock filter refer to " being used for the impact filtering method for strengthening asymmetric binary modulating signal "
(patent of invention number：ZL200910029875.3), this specification is intended merely to contrast using same digital shock filter conduct
The amplitude integration judgement demodulator of receiving filter and the performance of DL-CNN demodulators, so, the present embodiment is directly continued to use described
The numeral impact filter scheme of a limit of simple zero -3 in patent.The transmission function of the digital shock filter is as follows：
Filter coefficient in formula is：
b0=1, b1=-1.630, b2=1；
a1=-4.5781931992746454, a2=9.6546659241157258, a3=-
11.692079480819313,
a4=8.5756341567768217, a5=-3.6121554794765309, a6=
0.70084076007371199。
2) digital band-pass filter
For the bandpass filter 5 in Fig. 3, the present embodiment is realized using digital FIR filter.Needed for need to only giving
Performance of filter require (its bandwidth of the present embodiment major requirement is consistent with sending filter), you can imitated in Matlab softwares
Corresponding filter parameter is obtained in true platform, this is known for the art.
3.DL-CNN network trainings and design
In order to carry out Classification and Identification to the sequence of symhols newly inputted using CNN, first have to mutually interconnect its intrinsic nerve member
The weight coefficient connect is trained, that is, to be allowed CNN to learn and be remembered object to be classified or pattern.Convolutional neural networks
Be trained for Training model, the present embodiment directly carries out Matlab training and emulation to DL-CNN.
1) rule
Input signal during due to DL-CNN demodulator real works be containing interchannel noise, thus training when also need
Very important person is adds certain noise or disturbance, first, in order to meet the real work situation of demodulator in future, second, it can be direct
Influence the final generalization ability of CNN demodulators.In the present invention, the specific size of noise is trained according to the varying strength of intersymbol interference
Depending on, but there is no theory to follow at present, the present embodiment can only obtain some Experience norms and specific practice according to substantial amounts of emulation：
1. channel circumstance is more severe, intersymbol interference is more serious, and training noise should be smaller.This point is it can be appreciated that because right
Generally all it is that reception signal and reception environment are more severe in any communication system, receptivity is also poorer (or even can not work),
Thus now also just add bigger noise without extra again when CNN is trained；
2. the method for many experiments can be taken, to obtain the relatively best network of generalization ability.
2) simulation parameter
All kinds of simulation parameters in the present embodiment are as follows：
Still using EBPSK modulation, carrier frequency 30kHz, N=28, K=5 is simplified defined in (1) formula, sample rate is letter for modulation
10 times of number carrier frequency, sending filter bandwidth takes 4kHz, 2kHz and 1kHz respectively, and (2) formula is respectively adopted in receiving filter
Digital shock filter or common FIR bandpass filters.
3) determine that CNN trains noise
Selected through many experiments, during transmitting terminal band a width of 4kHz, 2kHz, 1kHz, training noise is respectively：
1. receiving terminal uses shock filter：[-5:0] dB, [- 5:0]dB.
2. receiving terminal uses common FIR bandpass filters：[-5:0] dB, [- 5:0] dB, [- 3:3]dB.
4) CNN structures and iterations are selected and carries out CNN network trainings
1. the CNN that the present embodiment uses is 6 Rotating fields, including 2 convolutional layers, 2 down-sampling layers, 1 input layer and 1
Output layer, convolution kernel size are 5.
It is balance quality and training generation 2. increase iterations can improve CNN learning ability, but will take considerable time
Valency, the present embodiment employ 10 iteration and CNN are trained.
4. performance simulation
The performance comparision of the DL-CNN demodulators designed as described above and existing amplitude judgement demodulator is shown in Fig. 7~figure
10.It can be seen that when intersymbol interference be present：
1) demodulation performance of most the superior has larger lifting (to increase than amplitude judgement demodulator in all DL-CNN demodulators
Benefit is at least in more than 2dB), and intersymbol interference is bigger, performance boost amount is also bigger；
2) Fig. 7 shows, when transmitting terminal bandwidth is limited to 4kHz, impact filtering+DL-CNN dicode member cascading judgement performances
It is optimal, than the performance boost about 2.5dB of classical integral judgement；
3) Fig. 8 and Fig. 9 show, when transmitting terminal bandwidth is limited to 2kHz：
1. use common FIR bandpass filters+DL-CNN demodulators：In low signal-to-noise ratio, single element judgement best performance,
Dicode member cascading judgement takes second place, and 3 symbol cascading judgements are worst；And in high s/n ratio, dicode member cascading judgement best performance is single
Bit decision takes second place, and 3 symbol cascading judgements are worst；
2. use shock filter+DL-CNN demodulators：In low signal-to-noise ratio, single element judgement best performance, dicode member
Cascading judgement takes second place, and 3 symbol cascading judgements are worst；And in high s/n ratio, 3 symbol cascading judgements are combined better than dicode member to be sentenced
Certainly, single element judgement is worst.
The performance of most the superior is superior to traditional quadrature judgement demodulator about 6.5-7dB in DL-CNN demodulators；
4) Figure 10 shows, when transmitting terminal bandwidth is limited to 1kHz, in normal bandpass filters+DL-CNN demodulators, and 3 yards
First cascading judgement best performance, the performance than traditional quadrature judgement are at least good 8-9dB；
5) he number once adjudicated is determined
Simulation result shows under different transmission bandwidth conditions and receiving terminal processing mode, optimal once judgement code
First number does not have unified conclusion.
5. the complexity of multiple-symbol CNN cascading judgements
It will be apparent that the neuron number of the input layer and output layer needed for n symbol cascading judgements is all the n of single element detection
Times, structure complexity can increase many, and computation complexity is also correspondingly doubled and redoubled.Thus it is limited to the complexity of network structure,
The present invention pertains only to this 3 kinds of cascading judgement modes of n=1,2,3.Main Conclusions is as follows：
1) training time of DL-CNN demodulators can accordingly increase with n growth, but emulate and show that amplification is few.And
And CNN is once trained just without changing again, then for DL-CNN demodulators training time without strict demand.
2) emulation display, with n growth, the increase of most of CNN extrapolation times and unobvious.In fact, due to base
In CNN classification judgement be adopt whole sample of signal property " batch processing " again in full n code-element period, as long as therefore CNN
The extrapolation time is less than n code-element period, you can ensures processing in real time.And it is now to adopt whole signals in full n code-element period
The n code-element period time needed for sample is then DL-CNN demodulators using the introduced inherent delay of n- symbol cascading judgements.
3) CNN structure complexity depends on the CNN used network number of plies, neuron number and convolution kernel dimensional parameters
Selection.
4) in the case where demodulation performance is suitable, the CNN judgement modes that preferred structure is low with computation complexity are answered certainly, this
The overall target of the general requirement of anti-ISI communication system, particularly receiver is depended primarily on to contrast and balance.
Described above is only the preferred embodiment of the present invention, it should be pointed out that：For the ordinary skill people of the art
For member, under the premise without departing from the principles of the invention, some improvements and modifications can also be made, these improvements and modifications also should
It is considered as protection scope of the present invention.
Claims (4)
- A kind of 1. DL-CNN demodulators of super Nyquist rate communication, it is characterised in that：Including receiving filter and DL-CNN Classification decision device（6）, said two devices cascade；The DL-CNN classification decision device（6）By the convolutional neural networks grader of multilayer Composition；Communication of the convolutional neural networks grader under with different channel width, signal to noise ratio and intersymbol interference environment Deep learning is carried out in sample, the wave character of the modulated signal filter response of extraction and memory with intersymbol interference and inherence are closed Join so as to complete to train；After training successfully, the signal newly inputted is adopted with the convolutional neural networks grader succeeded in school Sample sequence makes decisions, and decision steps are：DL-CNN classification decision devices（6）A numeral for indicating its classification is exported for each input symbols, then it is digital with this Control is gone to export corresponding local standard symbol waveform, i.e. DL-CNN classification decision devices（6）Input corresponding to n symbol Receiving filter output waveform samples, DL-CNN classification decision devices（6）Output correspond to the code character of this n symbol composition.
- 2. DL-CNN the demodulators of super Nyquist rate communication according to claim 1, it is characterised in that：It is described to connect Receipts wave filter is shock filter（4）.
- 3. DL-CNN the demodulators of super Nyquist rate communication according to claim 1, it is characterised in that：It is described to connect Receipts wave filter is bandpass filter（5）.
- 4. DL-CNN the demodulators of super Nyquist rate communication according to claim 1, it is characterised in that：The DL- CNN classification decision devices（6）Using single element decision method or multiple-symbol cascading judgement method, the multiple-symbol cascading judgement side Method is to carry out cascading judgement to whole sampled values in multiple code-element periods.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510220785.8A CN104811276B (en) | 2015-05-04 | 2015-05-04 | A kind of DL CNN demodulators of super Nyquist rate communication |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510220785.8A CN104811276B (en) | 2015-05-04 | 2015-05-04 | A kind of DL CNN demodulators of super Nyquist rate communication |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104811276A CN104811276A (en) | 2015-07-29 |
| CN104811276B true CN104811276B (en) | 2018-04-03 |
Family
ID=53695801
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510220785.8A Active CN104811276B (en) | 2015-05-04 | 2015-05-04 | A kind of DL CNN demodulators of super Nyquist rate communication |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104811276B (en) |
Families Citing this family (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017206188A1 (en) * | 2016-06-03 | 2017-12-07 | 华为技术有限公司 | Ftn-based communication method and relevant apparatus |
| KR101974261B1 (en) | 2016-06-24 | 2019-04-30 | 한국과학기술원 | Encoding method and apparatus comprising convolutional neural network(cnn) based in-loop filter, and decoding method and apparatus comprising convolutional neural network(cnn) based in-loop filter |
| CN106027438B (en) * | 2016-07-04 | 2020-04-07 | 东南大学 | Anti-aliasing modem for analog-digital hybrid amplitude modulation broadcasting system |
| CN106328150B (en) * | 2016-08-18 | 2019-08-02 | 北京易迈医疗科技有限公司 | Borborygmus sound detection method, apparatus and system under noisy environment |
| CN106886023B (en) * | 2017-02-27 | 2019-04-02 | 中国人民解放军理工大学 | A kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks |
| JP6801529B2 (en) * | 2017-03-09 | 2020-12-16 | 沖電気工業株式会社 | Information processing equipment, information processing methods, and programs |
| CN107317778B (en) * | 2017-07-28 | 2019-11-26 | 西安电子科技大学 | BPSK modulating signal phase transition detection method based on 1D-CNN |
| CN108540202B (en) * | 2018-03-15 | 2021-01-26 | 西安电子科技大学 | Satellite communication signal modulation mode identification method and satellite communication system |
| CN110300301B (en) * | 2018-03-22 | 2023-01-13 | 华为技术有限公司 | Image coding and decoding method and device |
| CN109150409B (en) * | 2018-09-30 | 2021-06-25 | 西安电子科技大学 | super-Nyquist adaptive system and method based on superposition coded modulation |
| CN109743271B (en) * | 2019-01-07 | 2020-04-14 | 西安电子科技大学 | Symbol estimation method of super-Nyquist system based on iterative interference cancellation |
| CN112633340B (en) * | 2020-12-14 | 2024-04-02 | 浙江大华技术股份有限公司 | Target detection model training and detection method, device and storage medium |
| CN112953565B (en) * | 2021-01-19 | 2022-06-14 | 华南理工大学 | Return-to-zero convolutional code decoding method and system based on convolutional neural network |
| CN113114422A (en) * | 2021-04-13 | 2021-07-13 | 兰州理工大学 | Deep learning detection-super-Nyquist rate atmospheric optical transmission method |
| CN113098601B (en) * | 2021-04-13 | 2022-05-24 | 兰州理工大学 | Deep learning pre-equalization-super-Nyquist rate atmospheric optical transmission method |
| CN113098602B (en) * | 2021-04-13 | 2022-05-24 | 兰州理工大学 | Deep learning pre-equalization and detection-based super-Nyquist atmospheric optical communication method |
| CN113347128B (en) * | 2021-05-27 | 2023-03-24 | 湖北工业大学 | QPSK modulation super-Nyquist transmission method and system based on neural network equalization |
| CN114169504B (en) * | 2021-12-10 | 2024-04-16 | 河北工业大学 | Self-adaptive filtering-based graph convolution neural network pooling method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103955718A (en) * | 2014-05-15 | 2014-07-30 | 厦门美图之家科技有限公司 | Image subject recognition method |
| CN104244113A (en) * | 2014-10-08 | 2014-12-24 | 中国科学院自动化研究所 | Method for generating video abstract on basis of deep learning technology |
| WO2014205231A1 (en) * | 2013-06-19 | 2014-12-24 | The Regents Of The University Of Michigan | Deep learning framework for generic object detection |
| CN104537647A (en) * | 2014-12-12 | 2015-04-22 | 中安消技术有限公司 | Target detection method and device |
-
2015
- 2015-05-04 CN CN201510220785.8A patent/CN104811276B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2014205231A1 (en) * | 2013-06-19 | 2014-12-24 | The Regents Of The University Of Michigan | Deep learning framework for generic object detection |
| CN103955718A (en) * | 2014-05-15 | 2014-07-30 | 厦门美图之家科技有限公司 | Image subject recognition method |
| CN104244113A (en) * | 2014-10-08 | 2014-12-24 | 中国科学院自动化研究所 | Method for generating video abstract on basis of deep learning technology |
| CN104537647A (en) * | 2014-12-12 | 2015-04-22 | 中安消技术有限公司 | Target detection method and device |
Non-Patent Citations (1)
| Title |
|---|
| 基于多层卷积神经网络学习的单帧图像超分辨率重建方法;刘娜 等;《中国科技论文》;20150131;第10卷(第2期);第201-205页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104811276A (en) | 2015-07-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104811276B (en) | A kind of DL CNN demodulators of super Nyquist rate communication | |
| CN110460402B (en) | End-to-end communication system establishing method based on deep learning | |
| CN109672639B (en) | Signal demodulation method based on machine learning | |
| CN109067688A (en) | A kind of OFDM method of reseptance of data model double drive | |
| CN107147600B (en) | Digital modulation signal demodulator based on neural network and demodulation method thereof | |
| Lin et al. | Learning of time-frequency attention mechanism for automatic modulation recognition | |
| CN107038421A (en) | Modulation Types recognition methods based on sparse storehouse own coding | |
| CN110163282A (en) | Modulation Mode Recognition method based on deep learning | |
| CN113630130B (en) | End-to-end digital communication demodulation method | |
| CN112733811B (en) | Method for identifying underwater sound signal modulation modes based on improved dense neural network | |
| CN114422311B (en) | Signal modulation recognition method and system combining deep neural network and expert priori features | |
| CN106027438A (en) | Anti-aliasing modulation demodulator for analog-digital hybrid amplitude modulation broadcasting system | |
| CN109194595A (en) | A kind of adaptive OFDM method of reseptance of channel circumstance neural network based | |
| CN109905337A (en) | A kind of channel equalization method based on NARX neural network and blocky feedback | |
| CN113055107B (en) | Interference strategy generation method for radio station with unknown communication mode | |
| CN110659684A (en) | Convolutional neural network-based STBC signal identification method | |
| CN114615118B (en) | Modulation identification method based on multi-terminal convolution neural network | |
| CN112291005A (en) | Bi-LSTM neural network-based receiving end signal detection method | |
| CN114595729A (en) | Communication signal modulation identification method based on residual error neural network and meta-learning fusion | |
| CN111404856B (en) | High-order modulation signal demodulation method based on deep learning network | |
| CN116896492A (en) | Modulation and coding joint identification method and system based on multichannel attention network | |
| Kalade et al. | Using sequence to sequence learning for digital bpsk and qpsk demodulation | |
| Gorday et al. | Flexible FSK learning demodulator | |
| CN113347128B (en) | QPSK modulation super-Nyquist transmission method and system based on neural network equalization | |
| CN115913849A (en) | Electromagnetic signal identification method based on one-dimensional complex value residual error network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| EXSB | Decision made by sipo to initiate substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |