CN104811276A - DL-CNN (deep leaning-convolutional neutral network) demodulator for super-Nyquist rate communication - Google Patents
DL-CNN (deep leaning-convolutional neutral network) demodulator for super-Nyquist rate communication Download PDFInfo
- Publication number
- CN104811276A CN104811276A CN201510220785.8A CN201510220785A CN104811276A CN 104811276 A CN104811276 A CN 104811276A CN 201510220785 A CN201510220785 A CN 201510220785A CN 104811276 A CN104811276 A CN 104811276A
- Authority
- CN
- China
- Prior art keywords
- cnn
- demodulator
- nyquist rate
- filter
- super
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- 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 DL-CNN (deep learning-convolutional neutral network) demodulator for super-Nyquist rate communication. For digitally modulated signals via a strict band-limited channel and even more than Nyquist rate bandwidth limitation, the demodulator directly extracts signal features including intersymbol interference from received signal samples subjected to filtering by a shock filter or an ordinary bandpass filter at the receiving end by the aid of a CNN, the CNN is trained by the aid of DL, and associated decision is made by adopting a single symbol element or multiple symbols, so that the signal samples can be classified by the DL-CNN under the higher intersymbol interference environment, intersymbol interference demodulation of the super-Nyquist rate modulation signals is realized, and good modulation performance and higher adaption capability are achieved as compared with a conventional amplitude integral judgment modulator.
Description
Technical field
The present invention relates to the communications field, the judgement demodulation problem of intersymbol interference modulation signal when the character rate particularly communicated exceedes Nyquist rate.
Background technology
Along with the development of information technology, wide-band mobile communication and smart mobile phone, radio-frequency spectrum is not only caused to become scarce resource, and the annual power consumption of the communication base station spread all over the country also exceedes over ten billion degree, therefore, advanced information society's active demand has the information transmission system of high spectrum utilization and high-energy utilance.
1. super Nyquist speed
At digital communicating field, baud rate and modulation rate, refer to the speed of valid data signal modulated carrier, can be regarded as the number of institute's transmission symbol (also claiming code element) in the unit interval, unit is Symbol/s or baud (Baud or Bd).And Nyquist (Nyquist) theorem is pointed out, the chip rate of particular channel can not exceed 2 times of low pass channel bandwidth, and thus 2Bd/Hz is just called Nyquist rate.Existing communication system, for improving the rate of information throughput (i.e. bit rate, unit is bit/s or bps) in per unit band, can only increase the number that each communication code element is quantized, namely introduce high order modulation.This is the common practise of communication technical field and classical technology, and associated has again following two common practise:
1) Nyquist rate is the theoretical maximum transmission rate under noiseless state, and actual physics channel has various interference unavoidably, therefore channel capacity will be restricted by Shannon (Shannon) formula;
2) chip rate exceedes Nyquist rate and not can not communicate, but intersymbol is bound to produce intersymbol interference (ISI), and existing optimal receiver model and correlation demodulation theoretical, be all be based upon there is no intersymbol interference prerequisite under.
Therefore, based on the common practise of this area, several conclusions are also obvious below:
1) communication undertaken by the narrower channel of bandwidth of Nyquist rate modulation signal is exactly super Nyquist rate communication (chip rate that namely communicates exceedes Nyquist rate);
2) super Nyquist rate communication directly can improve the availability of frequency spectrum;
3) the narrower bandpass filtering of bandwidth (being called for short " filtering of super band limit ") is applied to Nyquist rate modulation signal, obtain super Nyquist rate signal;
4) for same chip rate, transmission can adopt narrower channel and receiver bandwidth with reception super Nyquist rate signal, this contributes to reducing receiver noise factor, improve receiver sensitivity, and be expected to obtain higher received signal to noise ratio (SNR) under same transmitting power, or farther communication distance;
5) realize the key of super Nyquist rate communication, be interference signal between energy correct demodulation code, and prior art is divided into two steps usually: first eliminate intersymbol interference by the technological means such as channel equalization, liftering, recover normal modulation signal; Conventional method is adopted to complete demodulation again.
In a word, super Nyquist rate modulation signal is the result of losing high fdrequency component because normal speed modulation signal is subject to narrow band logical (carrier (boc) modulated signals) or low pass (baseband signal) filtering restriction in essence, and first existing reception treatment technology adopts channel equalization or liftering to eliminate intersymbol interference, equivalent band resistance (for carrier (boc) modulated signals) or high pass (for baseband signal) filter to be adopted exactly relatively to compensate the high fdrequency component of Received signal strength at frequency domain, this inevitably also can correspondingly elevator belt external noise, cause the deterioration of received signal to noise ratio before demodulation.So, can the high-performance demodulator of direct demodulation super Nyquist rate modulation signal, be the key that simultaneously can promote the communication system availability of frequency spectrum and capacity usage ratio.
2. degree of depth study-convolutional neural networks (DL-CNN)
2006, University of Toronto professor Geoffrey Hinton has delivered one section of article about many hidden layers deep neural network on " science " magazine, open the research tide of degree of depth study (Deep learning, be called for short DL) in academia and industrial quarters.
1) degree of depth study is a branch of machine learning, and main feature is obtained the expression for the different level of abstraction of initial data, and then improve the accuracy of the tasks such as classification and prediction.A pile is such as had to input I (signal as gathered under a pile varying environment), suppose that we devise the system S of a n layer, by the parameter in adjustment System, its output is made to remain input I, so just automatically can obtain a series of level characteristics inputting I, i.e. S1 ..., Sn.Therefore, be different from the shallow-layer learning algorithms such as traditional Support Vector Machine (SVM), DL is without the need to relying on artificial experience sample drawn feature, but the training data (namely utilizing " large data ") by building machine learning model and the magnanimity with a lot of hidden layer, carry out the feature that automatic learning is more useful, thus the final accuracy promoting classification or prediction.
2) convolutional neural networks (Convolutional Neural Networks, be called for short CNN) be the one of artificial neural net, its weights shared network structure makes it more to be similar to biological neural network, reduces the complexity of network model, decreases weights number.CNN is first real learning algorithm of successfully training multitiered network structure, and it utilizes spatial relationship to reduce the parameter needing study, to improve the training performance of general forward direction BP (backpropagation) algorithm.In CNN, sub-fraction data (local receptor field) input as the lowermost layer of hierarchical structure, information is transferred to different layers more successively, every layer obtains the most notable feature of observation data by digital filter, the notable feature of the observation data to translation, convergent-divergent and invariable rotary can be obtained accordingly, because the local receptor field of data allows neuron or processing unit to may have access to most basic feature.Therefore, the main feature of CNN is: convolution (carrying out local by local receptor field to node between layers to connect), weights are shared and pond (down-sampling).And in traditional BP neural net, each node layer is a linear one dimensional arrangement state, be entirely connected between layer and the network node of layer.
In recent years, based on the degree of depth study convolutional neural networks (DL-CNN) pattern classification with identify that particularly the field such as speech recognition and image recognition obtains immense success.CNN has self study and adaptive ability, and can carry out feature extraction by the input to a collection of mutual correspondence provided in advance, output data, analyze and grasp potential rule between the two, this process is called " training " of network.For new data, detection judgement directly can be carried out according to the rule of training out before.Therefore, in voice or image recognition, speech samples or image pixel directly as the input of network, can avoid feature extraction complicated in tional identification algorithm and data reconstruction processes.
In Fig. 1, above one deck be m layer, below one deck be m-1 layer.Can find out, each node on m layer is only connected with 3 nodes of m-1 layer corresponding region, this subrange is also called receptive field, connected by local, greatly reduce weights and connect number, for the clean input of a Far Left node on m layer, just equal the long-pending cumulative of all and that this node is connected last layer neuron node value and corresponding weights, such computational process is called convolution.Weights are shared, refer to all neuron nodes on a characteristic pattern all with same convolution nuclear phase convolution, extracted a kind of feature, if need to extract various features, so every layer has just had multiple characteristic patterns.For image recognition, in theory by convolution kernels different for imagery exploitation by obtaining multiple image after convolution, then these images are directly utilized to classify, but amount of calculation is too large like this, utilize pond (down-sampling) to operate data volume to be reduced, retain original characteristics of image to a certain extent simultaneously.Here, the region in pond is nonoverlapping, and the receptive field of convolution is overlapping.
For the character identification system LeNet-5 based on CNN that New York Univ USA Yann Le professor Cun proposes, its network configuration, as Fig. 2, comprises input layer and output layer always has 8 layers.In Convolution sums sub-sampling procedures, convolution process comprises: with a trainable filter f
xdeconvolute an image inputted (first stage is original input picture, and the stage is below exactly convolution characteristic pattern), then adds a biased b
x, obtain convolutional layer C
x; Sub-sampling procedures comprises: 4 the pixel summations of every neighborhood become 1 pixel, then by scalar W
x+1weighting, then increase biased b
x+1, then by a Sigmoid activation primitive, produce the Feature Mapping figure S that is about reduced 4 times
x+1.In CNN, each feature extraction layer (C layer) is used for asking the computation layer (S layer) of local average and second extraction followed by one, and this distinctive twice feature extraction structure makes network have higher distortion tolerance when identifying to input amendment.
Summary of the invention
Goal of the invention: in order to overcome the deficiencies in the prior art, the invention provides a kind of can the high-performance demodulator of direct demodulation super Nyquist rate modulation signal, for overcoming the existing communication system availability of frequency spectrum and the low technical problem of capacity usage ratio.
We notice:
1) demodulation of M system code element signal of communication is the classification of maximum M code element usually, thus simply too much than situations such as image recognitions in theory;
2) super Nyquist rate communication or bandwidth efficient channel are exactly convolution for the mechanism of signal transmission, and thus CNN should also be suitable for this situation;
3) restriction of channel width and the impact of intersymbol interference, make the correlation being provided with local between the front and back code element sample of super Nyquist rate modulation signal, the local receptor field of this local correlations and CNN should have certain corresponding relation;
4) by selecting different signal to noise ratios and bandwidth (or super Nyquist speed), modulated pattern can form enough " large data " and carry out degree of depth study for CNN.
Therefore, based on degree of depth learning method, convolutional neural networks is trained, DL-CNN is allowed to learn and after remembering the code element characteristic sum intersymbol interference pattern of super Nyquist rate modulation signal inherence, pattern classification and non-linear judgement are carried out to the super Nyquist rate modulation signal sampling value sequence of its input, is namely expected to realize the correct demodulation for super Nyquist rate modulation signal.
Technical scheme: based on above-mentioned thinking, the technical solution used in the present invention is:
A DN-CNN demodulator for super Nyquist rate communication, comprises receiving filter and DL-CNN classification decision device, said two devices cascade; Described DL-CNN classification decision device is made up of the convolutional neural networks grader of multilayer, described convolutional neural networks grader carries out degree of depth study having in different channel width, signal to noise ratio and the sample that communicates under intersymbol interference environment, extracts and memory completes training with the wave character of the modulation signal filter response of intersymbol interference and internal association.
Structurally be made up of modulator and super band limiting filter with the super Nyquist rate modulation device that above-mentioned DN-CNN demodulator is supporting, wherein modulator can be the conventional modulator of any system (as phase shift keying, frequency shift keying, quadrature amplitude modulation etc.), system (binary system, multi-system) and speed, and super band limiting filter also can be any type of band pass filter, just can modulated signal speed super Nyquist speed be controlled by its filtering bandwidth.Therefore, super Nyquist rate modulation device and generalised modulator there is no different in system configuration.
The present invention does not specialize in CNN itself, but suitably revise for the CNN network configuration being used successfully to image recognition, make it the sample sequence being applicable to train super Nyquist rate signal, propose the judgement detection method of the modulation signal anti-ISI based on DL-CNN.Here CNN is trained for grader, the whole code element received is identified.Carry out degree of depth study from from different channel width, signal to noise ratio and " large data " sample that communicates intersymbol interference environment, extract and memory with the wave character of the modulation signal filter response of intersymbol interference and internal association.CNN exports the numeral of its classification of mark for each input symbols, as represented " 0 " code element with 0, represent that " 1 " code element is (until represent " M code element " with M with 1, if for M ary modulation signal), go to control to export corresponding local standard symbol waveform by this numeral again, can judgement be completed.The CNN network that the present invention adopts is of five storeys altogether except input, output layer, the numerical value of training signal to noise ratio and training iterations different according to different ambient As.Training successfully, just can adjudicate with the signal sample sequence of the network succeeded in school to new input.
Further, in the present invention, according to input signal, select the type accepting filter, comprise shock filter and conventional band pass filter two kinds.
Be less than to " asymmetric " modulation signal of code-element period the modulates information period, selection receiving filter is shock filter, this is a kind of band pass filter of particular design, utilize precipitous phase-frequency characteristic amplify the different wave shape of asymmetric modulation signal and promote output signal-to-noise ratio, the modulation signature of signal can be given prominence to.
For the situation that described shock filter lost efficacy, accept filter and be chosen as conventional band pass filter, be applicable to the universal demodulation of any modulation signal, can filter out-band external noise better, promote the demodulation performance under strong intersymbol interference environment.
Such as, a kind of extended binary phase shift keying (Extended Binary Phase Shift Keying is called for short EBPSK) modulation signal of simplification is defined as follows:
s
0(t)＝sinω
ct， 0≤t＜T
Wherein, s
0(t) and s
1t () represents the modulation waveform of code element " 0 " and " 1 " respectively, ω
cfor carrier angular frequencies; Code-element period T=2 π N/ ω
ccontinue for N>=1 carrier cycle, the modulating time length τ=2 π K/ ω of " 1 " code element
ccontinue for the N number of carrier cycle of K <, K and N is integer to ensure that complete cycle modulates.When the modulation duration of τ=T and code element " 0 " and " 1 " is equal, (1) formula deteriorates to classical binary phase shift keying (BPSK) modulation, visible symmetric modulation BPSK is the special case of asymmetric modulation EBPSK, and thus the application's book is discussed for (1) formula and do not lost 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, the time domain waveform of the EBPSK signal then after transmitting terminal 2kHz bandpass filtering as shown in Figure 4, visible for information sequence " 11100 01 010 ", intersymbol interference has made EBPSK modulation signal waveform produce compared with serious distortion and spread to adjacent code element.The power spectrum of filtering front signal is as shown in Fig. 5 (a), and its main lobe width is about 12kHz; The power spectrum of filtered signal is then as shown in Fig. 5 (b), and now the sending filter of 2kHz has cut most of main lobe of modulation signal power spectrum.
Further, in the present invention, described DL-CNN classification decision device adopts single element decision method or multiple-symbol cascading judgement method, and described multiple-symbol cascading judgement method is carry out cascading judgement to the whole sampled values in multiple code-element period.
Multiple-symbol cascading judgement method is still illustrated for (1) formula.For the EBPSK modulation signal after accepting filter, we notice: when channel width constriction, stronger intersymbol interference cause its adjacent code element especially corresponding to continuous print " 1 " impact envelope expansion, overlapping even connected.Statistical independence condition between the code element of classical communication required (or supposed) is no longer satisfied.Accordingly, the present invention considers to adopt multiple-symbol joint-detection mode to design CNN judging module, the i.e. EBPSK impact filtering output waveform sampling be input as corresponding to n code element of CNN classification, it exports the code character (i.e. continuous print n position) also just corresponding to this n code element composition:
1. as n=1, be traditional single element independent detection, it is input as the waveform sampling corresponding to 1 code element, exports then non-" 1 " i.e. " 0 ";
2. as n=2, it is input as the waveform sampling corresponding to adjacent 2 code elements, and exporting is then one of 4 kind of 2 code element code character (2) below: " 11 ", " 10 ", " 01 ", " 00 ";
3. as n=3, it is input as the waveform sampling in succession corresponding to 3 code elements, and exporting is then one of 8 kind of 3 code element code character (3) below: " 000 ", " 001 ", " 010 ", " 011 ", " 100 ", " 101 ", " 110 ", " 111 ".All the other can the like.
Beneficial effect:
The DN-CNN demodulator of a kind of super Nyquist rate communication provided by the invention has the beneficial effect of following several respects:
1, capacity usage ratio is improved.
Favourable from two aspects:
1) DL-CNN judgement detects the global feature and the internal information that fully learn and make use of intersymbol interference modulation signal waveform, thus than being the simple threshold judgement utilizing amplitude information, significantly improve the demodulation performance of receiver, and intersymbol interference is more serious, larger compared to the advantage of traditional amplitude integration judgement demodulator;
2) for same chip rate, transmission can adopt narrower channel and receiver bandwidth with reception super Nyquist rate signal, this contributes to reducing receiver noise factor, improves receiver sensitivity, and is expected to obtain higher received signal to noise ratio under same transmitting power.
And the lifting of receiver demodulation performance and sensitivity is all equivalent to the increase of transmitting power, communication distance can be extended; If keep former communication distance index constant, then transmitting power just can reduce.This is for the energy resource consumption and the electromagnetic pollution that reduce communication system, and extending battery life, realizes " green communications ", have practical significance.
2, the availability of frequency spectrum is improved.
Modulator end one-sided constriction signal spectrum bandwidth no doubt directly can 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 when modulation signal bandwidth constriction, demodulation performance still can be made to remain on acceptable level.And that DL-CNN demodulator utilizes such multitiered network to have is non-linear, adaptability and degree of depth learning ability, by under stronger intersymbol interference to received signal sample classify, thus the anti-ISI demodulation achieved for modulation signal of making a start " speed-raising " modulation signal, make to promote the availability of frequency spectrum and there is feasibility and practical significance.
3, adaptive capacity is enhanced.
1) degree of depth learning ability of DL-CNN demodulator under " large data " support, make it can remember more signal characteristic and the characteristic of channel, thus improve the generalization ability of general nonlinearity demodulator, when real work, there is stronger robustness, adapt to occasion wider;
2) classical Threshold detection or the judgement of amplitude integration is adopted, generally carry out threshold judgement again get some sampling pointwises summation near impact filtering output peak value after, and CNN judgement is the batch processing to all samplings in a code-element period, or even the disposable cascading judgement to all samplings in n code-element period, thus its for the required precision of sample-synchronous far below the former, also lower to bit synchronous requirement;
3) existing communication receiver is compensate or eliminate the intersymbol interference caused because of band limit (even if not yet reaching the band limit of Nyquist rate or super Nyquist speed), first the technical finesses such as channel estimating, channel equalization, liftering will be carried out before demodulation, and DL-CNN demodulator of the present invention saves or " merging " this step by study in advance, it also avoid the deterioration of the signal to noise ratio before the demodulation that causes because out-of-band noise promotes;
4) DL-CNN demodulator can on-line study, thus there are the potentiality of " ten thousand become " adapting to modulation system and the characteristic of channel with " constant " of demodulator network topology or hardware configuration, be conducive to the normalization of demodulator and receiver, versatility, customizable and software radio realize.
Accompanying drawing explanation
Fig. 1 is local receptor field connection diagram between layers in convolutional neural networks.
Fig. 2 is the CNN network structure of a character identification system LeNet-5 based on CNN, comprises input layer output layer and always has 8 layers.Comprising 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 carrier frequency is 30kHz, and when transmitting terminal bandwidth is restricted to 2kHz, N=28, K=5, the time domain waveform comparison diagram of anti-phase EBPSK signal after transmitting terminal band pass filter during 10 sampling rate, sequence of symhols is " 1110001 01 0 ".
Fig. 5 is carrier frequency is 30kHz, modulation parameter N=28, K=5, simplify the power spectrum of EBPSK modulation signal during 10 sampling rate, the information symbol sequence modulated is " 1110001010 ", and wherein Fig. 5 (a) is original modulated signal power spectrum; Fig. 5 (b) is Fig. 5 (a) signal by bandwidth is power spectrum after the transmitting terminal band pass filter of 2kHz.
Signal when Fig. 6 is noiseless in Fig. 5 (b) is by the time domain waveform after receiving filter, and wherein Fig. 6 (a) adopts shock filter as receiving filter, and Fig. 6 (b) adopts FIR band pass filter as receiving filter.
Fig. 7 is transmitting terminal limit band 4kHz, receiving terminal adopts shock filter process, and DL-CNN adopts bit error rate performance when single element, dicode unit and 3 bit decision respectively, and the bit error rate performance of classical amplitude integration judgement.
Fig. 8 is transmitting terminal limit band 2kHz, receiving terminal adopts the process of common FIR band pass filter, and DL-CNN adopts bit error rate performance when single element, dicode unit and 3 bit decision respectively, and the bit error rate performance of classical amplitude integration judgement.
Fig. 9 is transmitting terminal limit band 2kHz, receiving terminal adopts shock filter and common FIR band pass filter to process respectively, DL-CNN adopts bit error rate performance when single element, dicode unit and 3 bit decision respectively, and the bit error rate performance of classical amplitude integration judgement.
Figure 10 is transmitting terminal limit band 1kHz, receiving terminal adopts the process of common FIR band pass filter, and DL-CNN adopts bit error rate performance when single element, dicode unit and 3 bit decision respectively, and the bit error rate performance of classical amplitude integration judgement.
Embodiment
Below in conjunction with accompanying drawing, the present invention is further described.
1. system describes
Specific embodiments of the invention are for the DL-CNN anti-ISI demodulation simplifying limit band EBPSK modulation signal under additive white Gaussian noise (AWGN) channel, its system block diagram is see Fig. 3, the left side of channel is super Nyquist rate modulation device, and the right of channel is the DL-CNN demodulator of anti-ISI.
Here super Nyquist rate modulation device is structurally made up of modulator 1 and super band limiting filter 2, wherein modulator 1 can be the conventional modulator of any system (as phase shift keying, frequency shift keying, quadrature amplitude modulation etc.), system (binary system, multi-system) and speed, and super band limiting filter 2 also can be any type of band pass filter, just can modulated signal speed super Nyquist speed be controlled by its filtering bandwidth.Therefore, super Nyquist rate modulation device and generalised modulator there is no different in system configuration.
DL-CNN demodulator comprises receiving filter and DL-CNN classification decision device 6, said two devices cascade; Described DL-CNN classification decision device 6 is made up of the convolutional neural networks grader of multilayer.
Transmitting terminal sequence of symhols is first through modulator 1, and the simplification EBPSK realizing defining according to 1 formula modulates; Then through super band limiting filter 2, the band communication channel of Nyquist rate can be exceeded with analog bandwidth restriction; Eventually pass channel 3 and be superimposed with white Gaussian noise.
First receiving terminal provides two kinds of possibilities to promote demodulation performances: one is amplify difference between the code element of " 0 " " 1 " by shock filter 4, and filtering part out-of-band noise; Two is the FIR band pass filter 5 filter out-band external noises by routine.Then the EBPSK Received signal strength after is after filtering sent into the DL-CNN trained to classify in decision device 6, demodulate corresponding information symbol.Because EBPSK is binary modulated, thus directly obtain information code current at this.
Launch the concrete enforcement describing demodulator below.
2. receiving filter design
Receiving filter both can Digital Implementation, also can simulated implementation, and just for the ease of simulation comparison, the present embodiment all adopts digital filter to realize.
1) digital shock filter (Digital Impacting Filters)
Described digital shock filter 4 in Fig. 3 is class infinite impulse response (IIR) band pass filters (also can adopt FIR filter), be made up of a pair conjugation zero point and at least two pairs of conjugate poles, zero point, resonance frequency was lower than the carrier frequency of input signal, pole frequency is all higher than the carrier frequency of input signal, and zero frequency can not between pole frequency, the close degree of zero frequency and pole frequency is not inferior to 10 of input signal carrier frequency
-3the order of magnitude, an extremely narrow trap-selecting frequency characteristic (or utilizing any precipitous phase-frequency characteristic) is presented with the centre frequency place ensureing in filter passband, make the filtering output waveform of asymmetric binary modulating signal produce parasitic amplitude modulation at modulates information place to impact, signal to noise ratio is promoted.Theory and design about digital shock filter refers to " for strengthening the impact filtering method of asymmetric binary modulating signal " (patent of invention number: ZL200910029875.3), the application's book just adopts the amplitude integration judgement demodulator of same digital shock filter as receiving filter and the performance of DL-CNN demodulator to contrast, so the numeral that the present embodiment directly continues to use simple zero-3 limit in described patent impacts filter scheme.The transfer function of this digital shock filter is as follows:
Filter coefficient in formula is:
b
0＝1，b
1＝-1.630，b
2＝1；
a
1＝-4.5781931992746454，a
2＝9.6546659241157258，a
3＝-11.692079480819313，
a
4＝8.5756341567768217，a
5＝-3.6121554794765309，a
6＝0.70084076007371199。
2) digital band-pass filter
For the described band pass filter 5 in Fig. 3, the present embodiment adopts digital FIR filter to realize.Only need given required performance of filter to require (its bandwidth of the present embodiment major requirement is consistent with sending filter), can obtain corresponding filter parameter in Matlab software emulation platform, this is the known of the art.
3.DL-CNN network training and design
In order to utilize the sequence of symhols of CNN to new input to carry out Classification and Identification, first to train its interconnective weights coefficient of intrinsic nerve unit, CNN namely will be allowed to learn and remember the object or pattern that will classify.Convolutional neural networks be trained for Training model, the present embodiment directly carries out Matlab training and emulation to DL-CNN.
1) rule
Because the input signal during real work of DL-CNN demodulator is containing interchannel noise, thus also need artificially to add certain noise or disturbance when training, one is the real work situation in order to meet demodulator in the future, and two is that it directly can affect the final generalization ability of CNN demodulator.In the present invention, the concrete size of training noise is determined according to the varying strength of intersymbol interference, but there is no theory at present and can follow, and the present embodiment can only obtain some Experience norms and specific practices according to a large amount of emulation:
1. channel circumstance is more severe, and intersymbol interference is more serious, and training noise should be less.This point is understood that, because for any communication system, be all that Received signal strength and reception environment are more severe usually, receptivity is poorer (even cannot work) also, thus now also just adds larger noise without the need to extra again when CNN trains;
2. the method for many experiments can be taked, to obtain the relatively best network of generalization ability.
2) simulation parameter
All kinds of simulation parameters in the present embodiment are as follows:
Modulate the simplification EBPSK still adopting (1) formula to define to modulate, carrier frequency 30kHz, N=28, K=5, sample rate is 10 times of signal carrier frequency, sending filter bandwidth gets 4kHz, 2kHz and 1kHz respectively, the digital shock filter of formula that receiving filter adopts respectively (2) or common FIR band pass filter.
3) determine that CNN trains noise
Select through many experiments, when transmitting terminal bandwidth is 4kHz, 2kHz, 1kHz, training noise is respectively:
1. receiving terminal adopts shock filter: [-5:0] dB, [-5:0] dB.
2. receiving terminal adopts common FIR band pass filter: [-5:0] dB, [-5:0] dB, [-3:3] dB.
4) selected CNN structure and iterations carry out CNN network training
1. the CNN that the present embodiment adopts is 6 Rotating fields, and comprise 2 convolutional layers, 2 down-sampling layers, 1 input layer and 1 output layer, convolution kernel size is 5.
2. increase the learning ability that iterations can improve CNN, but by the at substantial time, be balance quality and training cost, the present embodiment have employed 10 iteration and trains CNN.
4. performance simulation
The Performance comparision that the DL-CNN demodulator designed as mentioned above and existing amplitude adjudicate demodulator is shown in Fig. 7 ~ Figure 10.Can find out when there is intersymbol interference:
1) in all DL-CNN demodulators, the demodulation performance of most the superior has all had larger lifting (gain is at least at more than 2dB) than amplitude judgement demodulator, and intersymbol interference is larger, and performance boost amount is also larger;
2) Fig. 7 shows, when transmitting terminal bandwidth is restricted to 4kHz, impact filtering+DL-CNN dicode unit cascading judgement best performance, than the performance boost about 2.5dB of classical integral judgement;
3) Fig. 8 and Fig. 9 shows, when transmitting terminal bandwidth is restricted to 2kHz:
1. adopt common FIR band pass filter+DL-CNN demodulator: when low signal-to-noise ratio, single element judgement best performance, dicode unit cascading judgement takes second place, and 3 code element cascading judgements are the poorest; And when high s/n ratio, dicode unit cascading judgement best performance, single element judgement is taken second place, and 3 code element cascading judgements are the poorest;
2. adopt shock filter+DL-CNN demodulator: when low signal-to-noise ratio, single element judgement best performance, dicode unit cascading judgement takes second place, and 3 code element cascading judgements are the poorest; And when high s/n ratio, 3 code element cascading judgements are better than dicode unit cascading judgement, single element judgement is the poorest.
In DL-CNN demodulator, the performance of most the superior is all better than traditional quadrature judgement demodulator and is about 6.5-7dB;
4) Figure 10 shows, when transmitting terminal bandwidth is restricted to 1kHz, in normal bandpass filters+DL-CNN demodulator, and 3 code element cascading judgement best performances, 8-9dB at least better than the performance of traditional quadrature judgement;
5) he number of once adjudicating is determined
Simulation result shows under different transmission bandwidth conditions and receiving terminal processing mode, and best once adjudicates the ununified conclusion of he number.
5. the complexity of multiple-symbol CNN cascading judgement
Obviously, the input layer needed for n code element cascading judgement and the neuron number of output layer are all n times that single element detects, and structure complexity can increase many, and computation complexity is also correspondingly doubled and redoubled.Thus be limited to the complexity of network configuration, the present invention only relate to n=1,2,3 these 3 kinds of cascading judgement modes.Main Conclusions is as follows:
1) training time of DL-CNN demodulator can the corresponding increase along with the growth of n, but Emulating display amplification is few.And, CNN once train just without the need to changing again, so for training time of DL-CNN demodulator without strict demand.
2) Emulating display, along with the growth of n, most of CNN extrapolates the increase of time not obvious.In fact, due to based on CNN classification judgement be whole sample of signal property " batch processing " again of adopting in a full n code-element period, as long as therefore the extrapolation time of CNN be less than n code-element period, real-time process can be ensured.And now for adopting n the code-element period time needed for the whole sample of signal in a full n code-element period, be then the inherent delay that DL-CNN demodulator employing n-code element cascading judgement is introduced.
3) structure complexity of CNN depends on the selection of the network number of plies of CNN used, neuron number and convolution kernel dimensional parameters.
4) when demodulation performance is suitable, certainly answer the CNN that first-selected structure and calculation complexity is low to adjudicate mode, this overall target depending primarily on the general requirement of anti-ISI communication system, particularly receiver contrasts and balances.
The above is only the preferred embodiment of the present invention; be noted that for those skilled in the art; under the premise without departing from the principles of the invention, can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.
Claims (4)
1. a DN-CNN demodulator for super Nyquist rate communication, is characterized in that: comprise receiving filter and DL-CNN classification decision device (6), said two devices cascade; Described DL-CNN classification decision device (6) is made up of the convolutional neural networks grader of multilayer; Described convolutional neural networks grader carries out degree of depth study having in different channel width, signal to noise ratio and the sample that communicates under intersymbol interference environment, extracts and memory completes training with the wave character of the modulation signal filter response of intersymbol interference and internal association.
2. the DN-CNN demodulator of super Nyquist rate communication according to claim 1, is characterized in that: described receiving filter is shock filter (4).
3. the DN-CNN demodulator of super Nyquist rate communication according to claim 1, is characterized in that: described receiving filter is band pass filter (5).
4. the DN-CNN demodulator of super Nyquist rate communication according to claim 1, it is characterized in that: described DL-CNN classification decision device (6) adopts single element decision method or multiple-symbol cascading judgement method, and described multiple-symbol cascading judgement method is carry out cascading judgement to the whole sampled values in multiple code-element period.
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 true CN104811276A (en) | 2015-07-29 |
| CN104811276B 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) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106027438A (en) * | 2016-07-04 | 2016-10-12 | 东南大学 | Anti-aliasing modulation demodulator for analog-digital hybrid amplitude modulation broadcasting system |
| CN106328150A (en) * | 2016-08-18 | 2017-01-11 | 北京易迈医疗科技有限公司 | Bowel sound detection method, device and system under noisy environment |
| CN106886023A (en) * | 2017-02-27 | 2017-06-23 | 中国人民解放军理工大学 | A kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks |
| CN107317778A (en) * | 2017-07-28 | 2017-11-03 | 西安电子科技大学 | BPSK modulating signal phase transition detection methods based on 1D CNN |
| CN108476195A (en) * | 2016-06-03 | 2018-08-31 | 华为技术有限公司 | A kind of communication means, relevant apparatus based on FTN |
| CN108540202A (en) * | 2018-03-15 | 2018-09-14 | 西安电子科技大学 | A kind of satellite communication signals Modulation Mode Recognition method, satellite communication system |
| CN109150409A (en) * | 2018-09-30 | 2019-01-04 | 西安电子科技大学 | The Adaptable System and method of super Nyquist based on supercomposed coding modulation |
| CN109644268A (en) * | 2016-06-24 | 2019-04-16 | 韩国科学技术院 | Coding method and device and coding/decoding method and device including the loop filter based on CNN |
| CN109743271A (en) * | 2019-01-07 | 2019-05-10 | 西安电子科技大学 | Super Nyquist system symbol estimation method based on iterative interference cancellation |
| CN110300301A (en) * | 2018-03-22 | 2019-10-01 | 华为技术有限公司 | Image coding/decoding method and device |
| CN112633340A (en) * | 2020-12-14 | 2021-04-09 | 浙江大华技术股份有限公司 | Target detection model training method, target detection model training device, target detection model detection device and storage medium |
| CN112953565A (en) * | 2021-01-19 | 2021-06-11 | 华南理工大学 | Return-to-zero convolutional code decoding method and system based on convolutional neural network |
| CN113098601A (en) * | 2021-04-13 | 2021-07-09 | 兰州理工大学 | Deep learning pre-equalization-super-Nyquist rate atmospheric optical transmission method |
| CN113098602A (en) * | 2021-04-13 | 2021-07-09 | 兰州理工大学 | Deep learning pre-equalization and detection-based super-Nyquist atmospheric optical communication method |
| CN113114422A (en) * | 2021-04-13 | 2021-07-13 | 兰州理工大学 | Deep learning detection-super-Nyquist rate atmospheric optical transmission method |
| CN113347128A (en) * | 2021-05-27 | 2021-09-03 | 湖北工业大学 | QPSK modulation super-Nyquist transmission method and system based on neural network equalization |
| CN114169504A (en) * | 2021-12-10 | 2022-03-11 | 河北工业大学 | Graph convolution neural network pooling method based on adaptive filtering |
| US11366999B2 (en) * | 2017-03-09 | 2022-06-21 | Oki Electric Industry Co., Ltd. | Information processing device and information processing 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 |
|---|
| 刘娜 等: "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", 《中国科技论文》 * |
Cited By (32)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108476195A (en) * | 2016-06-03 | 2018-08-31 | 华为技术有限公司 | A kind of communication means, relevant apparatus based on FTN |
| CN108476195B (en) * | 2016-06-03 | 2021-02-23 | 华为技术有限公司 | FTN-based communication method, related device and system |
| CN109644268B (en) * | 2016-06-24 | 2022-03-08 | 韩国科学技术院 | Encoding method and apparatus including CNN-based loop filter, and decoding method and apparatus |
| CN109644268A (en) * | 2016-06-24 | 2019-04-16 | 韩国科学技术院 | Coding method and device and coding/decoding method and device including the loop filter based on CNN |
| US11095887B2 (en) | 2016-06-24 | 2021-08-17 | Korea Advanced Institute Of Science And Technology | Encoding and decoding methods and devices including CNN-based in-loop filter |
| CN106027438A (en) * | 2016-07-04 | 2016-10-12 | 东南大学 | Anti-aliasing modulation demodulator for analog-digital hybrid amplitude modulation broadcasting system |
| 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 |
| CN106328150A (en) * | 2016-08-18 | 2017-01-11 | 北京易迈医疗科技有限公司 | Bowel sound detection method, device 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 |
| CN106886023A (en) * | 2017-02-27 | 2017-06-23 | 中国人民解放军理工大学 | A kind of Radar Echo Extrapolation method based on dynamic convolutional neural networks |
| US11366999B2 (en) * | 2017-03-09 | 2022-06-21 | Oki Electric Industry Co., Ltd. | Information processing device and information processing method |
| 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 |
| CN107317778A (en) * | 2017-07-28 | 2017-11-03 | 西安电子科技大学 | BPSK modulating signal phase transition detection methods based on 1D CNN |
| CN108540202A (en) * | 2018-03-15 | 2018-09-14 | 西安电子科技大学 | A kind of satellite communication signals Modulation Mode Recognition method, satellite communication system |
| CN110300301A (en) * | 2018-03-22 | 2019-10-01 | 华为技术有限公司 | Image coding/decoding method and device |
| CN110300301B (en) * | 2018-03-22 | 2023-01-13 | 华为技术有限公司 | Image coding and decoding method and device |
| CN109150409A (en) * | 2018-09-30 | 2019-01-04 | 西安电子科技大学 | The Adaptable System and method of super Nyquist based on supercomposed coding modulation |
| 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 |
| CN109743271A (en) * | 2019-01-07 | 2019-05-10 | 西安电子科技大学 | Super Nyquist system symbol estimation method based on iterative interference cancellation |
| CN112633340A (en) * | 2020-12-14 | 2021-04-09 | 浙江大华技术股份有限公司 | Target detection model training method, target detection model training device, target detection model detection device and storage medium |
| 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 |
| CN112953565A (en) * | 2021-01-19 | 2021-06-11 | 华南理工大学 | Return-to-zero convolutional code decoding method and system based on convolutional neural network |
| CN113098602A (en) * | 2021-04-13 | 2021-07-09 | 兰州理工大学 | Deep learning pre-equalization and detection-based super-Nyquist atmospheric optical communication 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 |
| CN113114422A (en) * | 2021-04-13 | 2021-07-13 | 兰州理工大学 | Deep learning detection-super-Nyquist rate atmospheric optical transmission method |
| CN113098601A (en) * | 2021-04-13 | 2021-07-09 | 兰州理工大学 | Deep learning pre-equalization-super-Nyquist rate atmospheric optical transmission method |
| CN113347128A (en) * | 2021-05-27 | 2021-09-03 | 湖北工业大学 | QPSK modulation super-Nyquist transmission method and system based on neural network equalization |
| CN114169504A (en) * | 2021-12-10 | 2022-03-11 | 河北工业大学 | Graph convolution neural network pooling method based on adaptive filtering |
| CN114169504B (en) * | 2021-12-10 | 2024-04-16 | 河北工业大学 | Self-adaptive filtering-based graph convolution neural network pooling method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104811276B (en) | 2018-04-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104811276A (en) | DL-CNN (deep leaning-convolutional neutral network) demodulator for super-Nyquist rate communication | |
| CN110460402B (en) | End-to-end communication system establishing method based on deep learning | |
| CN107147600B (en) | Digital modulation signal demodulator based on neural network and demodulation method thereof | |
| CN109672639B (en) | Signal demodulation method based on machine learning | |
| CN113630130B (en) | End-to-end digital communication demodulation method | |
| Lin et al. | Learning of time-frequency attention mechanism for automatic modulation recognition | |
| CN108718288B (en) | Method for recognizing digital signal modulation mode based on convolutional neural network | |
| CN113055107B (en) | Interference strategy generation method for radio station with unknown communication mode | |
| CN106027438A (en) | Anti-aliasing modulation demodulator for analog-digital hybrid amplitude modulation broadcasting system | |
| CN104836773A (en) | Bipolar binary shift pulse keying modulation and demodulation method | |
| Park et al. | Deep learning-based automatic modulation classification with blind OFDM parameter estimation | |
| CN114422311B (en) | Signal modulation recognition method and system combining deep neural network and expert priori features | |
| CN116896492A (en) | Modulation and coding joint identification method and system based on multichannel attention network | |
| CN112804010A (en) | Polarity-staggered binary offset pulse keying modulation and demodulation method | |
| Gorday et al. | Flexible FSK learning demodulator | |
| CN113347128B (en) | QPSK modulation super-Nyquist transmission method and system based on neural network equalization | |
| CN109639301A (en) | A kind of FTN equalization methods based on reliability estimating | |
| CN102882823B (en) | Geometric feature discrimination demodulator based on EBPSK (Extended Binary Phase Shift Keying) signal impulse filter response | |
| Zhao et al. | An End-to-End Demodulation System Based on Convolutional Neural Networks | |
| He et al. | A learning based cognitive radio receiver | |
| Shen et al. | Learning the wireless channel: A deep neural network approach | |
| CN114301499B (en) | Intelligent frequency hopping signal receiving method based on CNN-GRU | |
| CN114301749B (en) | Signal anti-interference method based on generation of countermeasure network | |
| Mumtaz et al. | Autonomous modulation classification using single inception module based convolutional neural network | |
| Bouras et al. | Neural-net-based receiver structures for single-and multiamplitude bandlimited signals in CCI and ACI channels |
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 |