CN114696933B - AI receiver based on deep learning technology and use method - Google Patents
AI receiver based on deep learning technology and use method Download PDFInfo
- Publication number
- CN114696933B CN114696933B CN202210337246.2A CN202210337246A CN114696933B CN 114696933 B CN114696933 B CN 114696933B CN 202210337246 A CN202210337246 A CN 202210337246A CN 114696933 B CN114696933 B CN 114696933B
- Authority
- CN
- China
- Prior art keywords
- signal
- channel
- network
- sample
- pilot
- 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
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/391—Modelling the propagation channel
- H04B17/3912—Simulation models, e.g. distribution of spectral power density or received signal strength indicator [RSSI] for a given geographic region
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L25/00—Baseband systems
- H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
- H04L25/03—Shaping networks in transmitter or receiver, e.g. adaptive shaping networks
- H04L25/03006—Arrangements for removing intersymbol interference
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L25/00—Baseband systems
- H04L25/02—Details ; arrangements for supplying electrical power along data transmission lines
- H04L25/03—Shaping networks in transmitter or receiver, e.g. adaptive shaping networks
- H04L25/03006—Arrangements for removing intersymbol interference
- H04L25/03165—Arrangements for removing intersymbol interference using neural networks
Abstract
The invention discloses an AI receiver and a method based on deep learning technology, comprising three neural networks, a channel characteristic extractor, a signal characteristic extractor and a signal detection classifier, wherein the channel characteristic extractor estimates and extracts channel information according to the received signal, the output of the channel characteristic extractor is fed to the signal characteristic extractor to help the signal characteristic extractor to extract characteristic signals after eliminating channel influence, and the characteristic signals are finally fed to the signal detection classifier for recovery. The invention solves the problem that the non-communication exclusive characteristic of the deep learning architecture is not matched with the real-time and accurate requirements of communication signal detection, simultaneously improves the transfer capability and the generalization capability of the AI receiver, and embeds the knowledge in the digital communication field into the design of the neural network, so that the designed communication system receiver has higher generalization capability and robustness.
Description
Technical Field
The invention relates to the technical field of AI receivers, in particular to an AI receiver and a method based on a deep learning technology.
Background
Signal detection is an important component of wireless communication systems. The receiver of the traditional communication system implements signal detection according to a physical mathematical model of a channel, but when the actual communication system is implemented, due to the non-ideality of actual devices, a plurality of non-linear links (such as a non-linear power amplifier and a low-precision ADC) are introduced, so that an equivalent channel model of the communication system becomes very complex, and the traditional method cannot identify and compensate non-linear factors, so that the performance of the actual communication system is greatly reduced.
Disclosure of Invention
In order to solve the above technical problems, an object of the present invention is to provide an AI receiver and a method based on a deep learning technique, which solve the problem that a non-communication exclusive characteristic of a deep learning architecture is not matched with real-time and accurate requirements of communication signal detection, improve the mobility and generalization capabilities of the AI receiver, and embed digital communication Domain Knowledge (Domain Knowledge) into the design of a neural network, so that the designed communication system receiver has higher generalization capability and robustness.
In order to achieve the purpose, the invention adopts the technical scheme that:
an AI receiver based on deep learning technology comprises three Neural Networks (NN), a channel feature extractor, a signal feature extractor and a signal detection classifier, wherein the channel feature extractor estimates and extracts channel information according to signals, the output of the channel feature extractor is fed to the signal feature extractor to help the signal feature extractor to extract feature signals after eliminating channel influence, and the feature signals are finally fed to the signal detection classifier to be recovered.
The data symbol frame format actually transmitted by the communication system of the characteristic signal is as follows: the method comprises the steps that pilot frequency + useful signals are used, original data are mapped to corresponding constellation points through grouping at a transmitting end and are transmitted through a transmitter; the signal is propagated through an actual channel, a sample signal obtained by a receiving end is composed of a pilot signal sample and a useful signal sample, and an AI receiver performs signal detection according to the received samples.
The symbols sent by the transmitting terminal in digital communication are taken from corresponding constellations, and the corresponding transmitting signals are taken from a finite set of Gamma = { s = } 1 ,s 2 ,...,s m And transmitting a signal to a receiving end after the signal passes through a channel, wherein the received signal is y = f (x), the mapping f is related to the actual channel characteristic, and the communication aims to recover the original transmitted signal as much as possible according to the received signal y.
A working method of an AI receiver based on a deep learning technology comprises the following steps;
in the data transmission process, data of a transmitting end is divided into a pilot signal and a useful signal, in one frame of data, the pilot signal is in front of the useful signal, the useful signal is followed by the useful signal, a sample signal is obtained at a receiving end after channel transmission, the pilot signal is transmitted to the receiving end to obtain a pilot sample signal, the useful signal is transmitted to the receiving end to obtain a useful signal sample, when the data is transmitted, the pilot sample reaches the receiving end first, the pilot signal is known by a receiver and a transmitter, the pilot signal and the pilot sample signal are input to a channel characteristic extractor, the channel characteristic extractor is realized by a fully-connected neural network, the module estimates channel information according to the transmitted pilot signal and the pilot receiving sample, the pilot signal is changed into the pilot sample after actual channel action, the difference between the pilot signal and the pilot sample is caused by a channel, the neural network of the channel characteristic extractor intelligently estimates the channel information according to the pilot and the sample, and finally the output of the channel characteristic extractor represents the channel information.
After the useful signal sample arrives, the signal sample and the output of the channel characteristic extractor are used as input and sent to a channel equalization network, the channel equalization network performs equalization processing on the signal sample according to channel information, information related to a channel in the signal sample is eliminated as much as possible, parameters of the channel equalization network are not trainable and are generated by the channel characteristic extraction network, the data transmission process of a neural network is that network weight coefficients are multiplied by the input, so that when the parameters of the channel equalization network are generated by the output of the channel characteristic extraction network, the weight coefficients of the channel equalization network are channel state information, after the useful signal sample arrives, multiplication-like operation is achieved after the useful signal sample passes through the channel equalization network, the output of the channel equalization network is a signal characteristic vector, and finally a signal classifier performs signal detection according to the signal characteristic vector.
The channel equalization network introduces a link discriminator to a receiver module in a training stage, the link discriminator is at the same position as a signal classifier, the link discriminator is a neural network, the input of the link discriminator is a signal characteristic vector and a sending signal, and the link discriminator is used for estimating the current state as far as possible according to the inputThe channel state experienced by the channel, the link identification network and the signal detection network have respective optimization targets, and the optimization targets of the channel characteristic extraction network and the channel equalization network are that as far as possible, the link discriminator cannot realize accurate judgment according to the signal characteristic vector, but the signal detection network can detect according to the signal characteristic network, which means that the channel related information in the sample signal is rejected by the channel equalization network, so that the link identification network cannot perform accurate estimation based on the signal characteristic vector, and finally the probability of network guess pair tends to:
the signal detection network enables high quality signal decisions, which means that the AI receiver learns how to perform channel estimation and channel equalization based on the sample signal.
The invention has the beneficial effects that:
the communication system receiver based on the deep learning technology combines the field knowledge of the communication system with the deep learning technology, the channel feature extractor, the signal feature extractor and the signal detection classifier work cooperatively, and a link discriminator is introduced in the training stage, so that the generalization capability of a network is improved.
When the nonlinear loop segment is not considered, the AI receiver, the LS channel estimation algorithm of the traditional communication and an accurate error rate curve are basically superposed, so that the AI receiver can realize the performance consistent with that of the traditional communication algorithm, and the great potential of the AI receiver designed by the inventor is proved.
When two nonlinear loops, I/Q imbalance and nonlinear power amplifier, are considered, the performance of the AI receiver is far better than that of the LS channel estimation algorithm in conventional communication, the channel estimation and detection in the conventional method cannot consider the factor of I/Q imbalance, the system still performs channel estimation and equalization according to the method under normal conditions, the result of channel estimation is poor, and rapid performance degradation is caused. But the AI receiver is realized based on the neural network, the hidden layer of the neural network can be integrated with any function at a certain precision, and when the nonlinear link is introduced, the AI receiver can intelligently capture the rule of the nonlinear link according to the change of the data, further formulate a corresponding judgment detection rule and compensate, and has better performance. This is also the greatest advantage of the AI receiver over conventional communications.
The deep learning technology is an advantageous tool, the neural network has the advantages of feature extraction and classification, so that the deep learning technology is more advantageous for communication in complex scenes considering non-ideal factors in engineering practice, and the performance superior to the traditional communication algorithm can be realized by designing a communication system receiver based on the deep learning technology in the complex scenes.
Drawings
FIG. 1 shows the contents of the present invention.
Fig. 2 is a model diagram of a wireless communication system according to the present invention.
Fig. 3 is a schematic diagram of the decision based on the sample signal at the receiving end according to the present invention.
Fig. 4 is a schematic diagram of a receiver structure in conventional communication.
Fig. 5 is a diagram illustrating a format of a transmission data frame.
Fig. 6 is a schematic diagram of an AI receiver architecture.
Fig. 7 is a schematic diagram of a neural network implementation of the AI receiver.
Fig. 8 is a BPSK/QPSK signal constellation.
Fig. 9 is a schematic diagram of training and testing of an AI receiver.
Fig. 10 is a 16-QAM constellation.
FIG. 11 is a graph illustrating the effect of non-linear factors: I/Q imbalance (left) nonlinear power amplifier (right).
Fig. 12 is a simulation diagram of a non-linear element flat rayleigh fading channel.
Fig. 13 is a simulation diagram of a flat rayleigh fading channel under I/Q imbalance.
FIG. 14 is a schematic diagram of data visualization of internal analysis of signals under I/Q imbalance.
Fig. 15 is a schematic diagram of a simulation of a flat rayleigh fading channel with nonlinear power.
Detailed Description
The invention will be further explained with reference to the drawings
As shown in fig. 1: the AI receiver is designed based on deep learning techniques. In order to solve the problem that the non-communication exclusive characteristic of a deep learning architecture is not matched with the real-time and accurate requirements of communication signal detection, and meanwhile improve the migration capability and the generalization capability of an AI receiver, digital communication Domain Knowledge (Domain Knowledge) is embedded into the design of a neural network. Aiming at a classic communication system receiver model, a knowledge-driven signal detector architecture is designed by adopting a modular design method.
Under this framework, a preliminary AI receiver was designed: three Neural Networks (NN) were utilized: a channel feature extractor, a signal feature extractor and a signal classifier form the whole receiver and estimate the transmitted symbols. In addition, based on experimental results, a new component called a link discriminator is integrated into the architecture based on the existing AI receiver by using the knowledge in the learning idea and communication of the countermeasure type to improve the versatility of the detector. As shown in fig. 2:
the communication system comprises a transmitter, a channel and a receiver, wherein a channel characteristic extractor, a signal characteristic extractor and a signal detection classifier form the whole receiver and estimate transmitted symbols;
the channel characteristic extractor is realized by a full-connection neural network, the module is used for estimating channel information according to a transmitted pilot signal and a pilot receiving sample, the pilot signal is changed into the pilot signal sample after the actual channel action, the difference between the two signals is caused by a channel, so the channel information can be estimated according to the pilot signal and the sample thereof, and finally the output of the channel characteristic extractor represents the channel information. After the useful signal sample arrives, the signal sample and the output of the channel feature extractor are used as input to be sent to a channel equalization network, and the network is used for carrying out equalization processing on the signal sample according to channel information and rejecting information related to a channel in the signal sample as far as possible. The parameters of the channel equalization network are not trainable and are generated by the channel characteristic extraction network, the data transmission process of the neural network is input and multiplied by the network weight coefficient, so when the parameters of the channel equalization network are generated by the output of the channel characteristic extraction network, the weight coefficient of the channel equalization network is channel state information, after a useful signal sample arrives, multiplication-like operation is realized after the useful signal sample passes through the channel equalization network, the knowledge of a traditional receiver is used for reference, the output of the channel equalization network is a 'signal characteristic vector', and finally, a signal classifier carries out signal detection according to the signal characteristic vector. Under ideal conditions, the channel related information in the sample signal is eliminated by the channel equalization network and does not exist in the signal feature vector, and the signal classifier can realize accurate detection according to the signal feature vector.
The working steps and the principle of the AI receiver are specifically as follows:
during the data transmission process, the data at the transmitting end can be divided into a pilot signal and a useful signal, wherein in one frame of data, the pilot signal is in front of the useful signal and the useful signal is immediately behind the pilot signal. After channel transmission, a sample signal is obtained at a receiving end, a pilot signal is transmitted to the receiving end to obtain a pilot sample signal, and a useful signal is transmitted to the receiving end to obtain a useful signal sample. When data is transmitted, pilot signal samples arrive at a receiving end first, and since the pilot signal is known by both a receiver and a transmitter, the pilot signal and the pilot signal samples are sent to a 'channel characteristic extractor' as input. The channel characteristic extractor is realized by a fully-connected neural network, the module is used for estimating channel information according to a transmitted pilot signal and a pilot frequency receiving sample, the pilot signal is changed into a pilot signal sample after the actual channel action, the difference between the pilot signal and the pilot signal sample is caused by a channel, the neural network of the channel characteristic extractor part can intelligently estimate the channel information according to the pilot frequency and the sample thereof, and finally the output of the channel characteristic extractor represents the channel information.
After the useful signal sample arrives, the signal sample and the output of the channel characteristic extractor are used as input to be sent to a channel equalization network, and the network is used for carrying out equalization processing on the signal sample according to channel information and eliminating the signal sample and the channel related to the signal sample as much as possibleThe information of (1). The parameters of the channel equalization network are not trainable and are generated by the channel characteristic extraction network, so the design just uses the knowledge of the traditional communication, and the receiver of the traditional communication system estimates the channel H according to the pilot signal firstly and then uses the channel H + And Y realizes channel equalization. The data transmission process of the neural network is that the input is multiplied by the network weight coefficient, so when the parameters of the channel equalization network are generated by the output of the channel characteristic extraction network, the weight coefficient of the channel equalization network is channel state information, after a useful signal sample arrives, multiplication-like operation is realized after the channel equalization network, the knowledge of a traditional receiver is used for reference, the output of the channel equalization network is a 'signal characteristic vector', and finally, the signal classifier carries out signal detection according to the signal characteristic vector. Under ideal conditions, the channel related information in the sample signal is eliminated by the channel equalization network and does not exist in the signal feature vector, and the signal classifier can realize accurate detection according to the signal feature vector.
In practice, because the true value of H is ignored, the channel characteristic extraction network cannot well remove the channel-related information from the received signal, and the signal characteristic vector depends on the channel state information. In this case, it is likely that the symbol detection error rate can still be kept at a very low level on the training data set, since during the training phase the network can detect from the received signal without fully equalizing out the channel information. However, the dependence of the signal feature vector on the channel state information means that the signal detection rules also depend on the channel state information. Therefore, when the network is used for a new channel (test phase), the learned signal detector cannot meet the design target, because the channel condition of the new device is different from that of the training device, and the detection performance of the AI receiver is lower than that of the test phase, so the generalization capability is weak.
In the training stage, channel information is not known, a discrete channel state set is established during training, different channel states are labeled, and each channel state has a corresponding channel index value. In order to ensure that the channel equalization network can realize perfect channel equalization operation, in the training stageThe segments introduce a "link discriminator" to the receiver module, at the same location as the signal classifier. The link discriminator is also a neural network, the inputs of which are the signal feature vectors and the transmitted signals, and the function of which is to estimate, as far as possible from the inputs, the channel conditions experienced by the current channel. The link identification network and the signal detection network have respective optimization targets, and the optimization targets of the channel characteristic extraction network and the channel equalization network are that the link identification device cannot realize accurate judgment according to the signal characteristic vector as much as possible, but the signal detection network can detect according to the signal characteristic network. That is, the channel related information in the sample signal is rejected by the channel equalization network, so that the link discrimination network cannot perform accurate estimation based on the signal feature vector, and finally the probability of network guess tends to:
the signal detection network can realize high-quality signal judgment, which means that the AI receiver learns how to perform channel estimation and channel equalization based on the sample signal, and the network has high generalization capability. As shown in fig. 3: the signal types of the digital communication at the transmitting end are limited, if the signal at the transmitting end has M types, the original signal is distorted after passing through the signal corresponding to M constellation points in a constellation diagram, and the signal sample obtained by the receiving end may be far from the original signal. And the estimation recovery for the signal is essentially for the received signal
The sample space is divided, the original signal changes after passing through the channel, and the influence caused by the channel needs to be eliminated from the received sample as much as possible when the receiving end needs to realize signal recovery according to the obtained scattered sample signal, namely, the channel equalization operation in the traditional communication, which can prove that after the channel equalization operation is carried out, the minimum distance criterion under the discrete vector is equivalent to the maximum likelihood judgment criterion, and the signal judgment can be carried out according to the minimum distance criterion.
FIG. 4 illustrates a wide range of conventional communication systemsThe receiver architecture employed. Firstly, the pilot frequency part of the received signal passes through a channel estimation module to obtain the estimation of a channel coefficient matrix H
In that
On the basis of the received signal
Using channel equalizers H + To obtain
Finally based on the processed signal
A minimum distance criterion is applied for the decision.
The above receiver architecture can achieve good detection performance under most channel conditions. This classical receiver design is applied to signal detection when the channel conditions are unknown. The classical design structure of fig. 4 is applied to a neural network-based AI receiver, and the entire AI receiver is composed of a plurality of neural network modules. Before describing a specific AI receiver architecture.
As shown in fig. 5: the data frame format adopted in the data transmission is as follows:
in a frame signal, the pilot ratio is:
the pilot frequency occupation ratio is very low, which means that the efficiency of signal transmission is higher, and the requirement of small sample learning is met. Combining a conventional communication system receiver structure and a data frame format.
As shown in fig. 6: when data is transmitted, a pilot signal arrives first, the pilot signal is a signal known to both the transceiver, and corresponding channel information is extracted by using the pilot signal and a received signal after the pilot signal passes through a channel. The received pilot signal and the corresponding transmitted pilot are used as inputs to a "channel feature extraction network" that functions similarly to a channel estimator. The useful signal gets the corresponding sample signal after passing through the channel, the output of the channel characteristic extraction and the sample signal sequence are fed to the 'channel equalization network', and the channel equalization network can be regarded as a channel equalization module. The output of the channel equalization network is called the "signal feature vector". Ideally, the channel related information in the sample signal is removed by the channel equalization network and does not exist in the signal feature vector, which also indicates that the network realizes a perfect channel equalization function. The AI receiver structure of the above figure introduces a "link discrimination network", and if there is no link discrimination network, the detection decision is directly made on the signal feature vector, but because the true value of H is ignored, the channel feature extraction network cannot well remove the channel related information from the received signal, and the signal feature vector depends on the channel state information. In this case, it may be that the symbol detection error rate can still be kept at a very low level on the training data set, since during the training phase the network is able to detect from the received signal which does not fully equalize the channel information. However, the dependence of the signal feature vector on the channel state information means that the signal detection rules also depend on the channel state information. Therefore, when the network is used for a new channel (test phase), the learned signal detector cannot meet the design target, because the channel condition of the new device is different from that of the training device, and the detection performance of the AI receiver is lower than that of the test phase, so the generalization capability is weak.
The link identification network has the function of estimating the index value h of the channel state experienced by the current signal as much as possible by taking the real transmission data as a label according to the obtained signal characteristic vector index . And the signal detection network performs decision detection according to the obtained signal characteristic vector to minimize a loss function cross-entropy. Output probability vector of softmax function can be used in network training
And cross-entropy of the one-hot encoded vector of the message sequence as a loss function, i.e.:
L loss ＝-ln(b i ) (1)
in the formula:
b i -outputting the probability vector
The ith element of (2);
the signal detection network is therefore a training process with supervised learning. The link identification network and the signal detection network have respective optimization targets, and ideally, channel related information in the sample signal is removed by the channel equalization network, so that the link identification network cannot carry out accurate estimation based on the signal characteristic vector, and finally the probability of network guess tends to be as follows:
however, the signal detection network can realize high-quality signal judgment, the final goal is achieved, the AI receiver learns how to carry out channel estimation and channel equalization based on the sample signal, and the network has high generalization capability.
The neural network structure of the AI receiver is shown in fig. 7 below, in which the corresponding transmissions are represented by directional line segments of different colors. In a specific design process, the equalization network in fig. 7 is also implemented by a neural network, but the weights of the network are not trainable but generated by a channel feature extraction network. Because in the classical communication theory, the channel equalization is performed by estimating the pseudo-inverse of the channel H and the received signal vector
And multiplication. If the channel equalization is performed by a trainable neural network, the channel characteristics and the received signal are fed into the channel equalization, and how they are processed is determined by the structure and parameters of the neural network. Because the operation of the neural network is very large with the operation of the classical multiplier-equalizerIn contrast, neural networks for channel equalization cannot achieve channel equalization in a standard manner. The output of the channel feature extraction network (i.e., the channel features) is used as a parameter of the feature extraction network. And multiplying the channel characteristics by the received signal to obtain an output according to the operation rule of the neural network. Thus, the feature extraction network operates according to a standard channel equalization method, and can apply Domin Knowledge to AI receivers.
And (3) simulation result analysis:
1 simulation Condition and parameter configuration
For the AI receiver, the test is carried out under a flat Rayleigh fading channel, and the performance of the system is judged through an error rate curve obtained by simulation. The conditions of the simulation are divided into two cases:
the first method is to consider the influence of a nonlinear link in actual implementation, adopt a BPSK/QPSK modulation mode, and compare the performance of the AI receiver with the detection performance under the traditional LS channel estimation algorithm.
The second one is to consider the nonlinear factors (I/Q imbalance and nonlinear power amplifier) in the actual communication system, and adopt the 16-QAM modulation method to compare the AI receiver performance with the detection performance under the traditional LS channel estimation algorithm in such a more complex environment.
In the process of completing the task of the project, all the work is completed on a computer in a simulation mode, and software used for the simulation mainly comprises Pycharm and Matlab. Wherein, pycharm mainly completes the training and testing work of the AI receiver; matlab mainly completes simulation work of a traditional communication algorithm in a corresponding scene, and meanwhile, drawing of all result graphs in an experiment is completed through Matlab.
Flat Rayleigh fading channel without considering nonlinear link
The channel environment adopted by system simulation in the scene is a flat Rayleigh fading channel, and the flat fading channel can be equivalent to a single-tap filter. In a flat rayleigh fading channel, the module value of the filter tap follows rayleigh distribution, both the real part and the imaginary part follow gaussian distribution, and the discrete vector expression of the received signal can be expressed as:
y＝h·x+n (2)
in the formula:
x-the transmitter's transmitted signal (in discrete vector form);
y-the received signal (in discrete vector form) after the transmitted signal has passed through the flat rayleigh fading channel;
n-flat rayleigh fading channel gaussian noise, n obeying a complex gaussian distribution: n to CN (0,N) 0 ) The real and imaginary parts are normally distributed:
h-flat rayleigh fading channel filter tap, h obeys complex gaussian distribution: h CN (0,1), the real part and the imaginary part obey normal distribution respectively:
the modulation mode adopted by the signal at the transmitting end is BPSK/QPSK, and the corresponding constellation diagram is shown in fig. 8 as follows:
analyzing an accurate bit error rate expression:
the QPSK signal can be regarded as two independent BPSK signals, so from the perspective of the constellation diagram, the performance of the two modulation schemes is completely consistent. According to the relevant theory of digital communication, the precise bit error rate expression of the BPSK/QPSK signal under the AWGN channel is as follows:
in the formula:
snr — bit signal to noise ratio:
the error rate performance under the flat Rayleigh fading channel can be analyzed on the basis of the AWGN channel, and the conditional error rate under one-time channel realization is as follows:
training and testing of AI receivers:
because the actual channel model is not known, the label data trained by the AI receiver is obtained by sampling the actual physical channel, i.e. the data X can be transmitted in a specific time period i Through the channel at this time, a sample signal Y is obtained i When the channel information is unknown, a specific index mark can be set for the channel at the moment
Indicates the set of tag data (X) i ,Y i ) Generated under the ith channel, the aim of the link discrimination network is to estimate the index value as accurately as possible
100 discrete channel states are sampled to generate corresponding label data for training of the AI receiver. The statistical distribution of the tested channel environment is unchanged, but the channel information during testing and the channel information during training are completely independent, and each frame of data is supposed to experience the same channel realization, and the channel realization among different frame of data is mutually independent. I.e. a training, permanent use mode. Measuring the error rate performance of the AI receiver under the test channel, the specific steps are as shown in fig. 9:
flat Rayleigh fading channel under consideration of nonlinear element
In this scenario, the modulation scheme adopted by the signal is 16-QAM, and the corresponding constellation is as shown in fig. 10,
the performance of a communication system is degraded by non-linear elements in the actual communication system, which are caused by non-idealities of physical devicesProperties such as quantization error, low precision ADC, etc., consider two non-linear factors: I/Q imbalance and nonlinear power amplification. In the analysis of noise, the actual noise z is different from the gaussian assumption of the received noise that is generally adopted k May follow a non-gaussian distribution (e.g., in an industrial environment, transmitted symbols may experience impulsive interference). In practical applications, it is often not possible to model transmitter non-idealities and receiver noise accurately. Therefore, the conventional model-based signal detection method cannot be applied. Non-linear factors can cause distortion of the signal constellation, as shown in fig. 11:
non-linearity due to transmitter non-ideality:
1. I/Q imbalance
2. Non-linear power amplifier
Wherein
And
with v =1, β a ＝0.25,
β φ ＝0.25。
Impulse noise model:
1. input-output relationship under flat rayleigh fading channel:
y k ＝h k x k +n k
wherein h is k CN (0,1) and n k Represents the equivalent noise of the receiving end due to the influence of impulse noise, n k Is not in a complex gaussian distribution subject to the standard.
n k ＝w k +i k
Wherein the content of the first and second substances,
is the receiving end thermal noise; i all right angle k The impulse noise is expressed and modeled as a Bernoulli-Gaussian random process, which can be expressed as: product of the real Bernoulli Process and the Complex Gaussian Process i k ＝b k g k
Wherein b is k ～B(1,0.1)，g k ～CN(0,5σ 2 ). Equivalent noise n of receiver under the model k Has a variance of σ 2 。
2 simulation results analysis
Flat Rayleigh fading channel without considering nonlinear link
When the nonlinear link in actual communication is not considered, the error rate performance under the BSK/QPSK modulation mode is compared, and the specific simulation result is shown in fig. 12 as follows:
from the simulation result of the upper graph, it can be seen that, when the nonlinear link is not considered, the AI receiver, the LS channel estimation algorithm of the traditional communication and the accurate bit error rate curve are basically overlapped, which indicates that the AI receiver can realize the performance consistent with the traditional communication under the corresponding data frame format, and can approach the lower mathematical bound of theory, thus proving that the AI receiver designed by the inventor has great potential, which also lays a foundation for the following study of the nonlinear link and the simulation, and provides possibility.
Flat Rayleigh fading channel under consideration of nonlinear link
Considering the nonlinear link in actual communication, the bit error rate performance under the 16-QAM modulation mode is compared.
The specific simulation results when considering the I/Q imbalance are shown in FIG. 13 below:
as can be seen from the simulation results of the above diagram, when the I/Q imbalance nonlinear loop is considered, the performance of the AI receiver is far better than that of the LS channel estimation algorithm in conventional communication, the channel estimation and detection in the conventional method cannot consider the I/Q imbalance factor, the system still performs channel estimation and equalization according to the method under normal conditions, the channel estimation result is very poor, and rapid performance degradation is caused.
But the AI receiver is realized based on the neural network, the hidden layer of the neural network can integrate any function at a certain precision, and when the I/Q unbalanced nonlinear link is introduced, the AI receiver can intelligently capture the rule of the nonlinear link according to the change of the data, further formulate a corresponding judgment detection rule for compensation, and have better performance. This is also the greatest advantage of AI receivers over conventional communications.
For the deep analysis of an AI receiver and a traditional LS channel estimation and equalization algorithm under the condition of I/Q imbalance, the propagation of internal signals can be observed, the signals after final equalization and the corresponding judgment results are observed, and the difference between the two methods can be visually seen. Internal analysis is shown in fig. 14: from the figure, it can be found that the point with wrong judgment can be found visually by comparing the actual signals after the equalization of the two systems with the judgment result, the equalization is not thorough in the traditional method, and the misjudgment occurs to many points.
When considering a nonlinear power amplifier, the specific simulation result is shown in fig. 15 as follows: from the simulation results of the above figures, when the nonlinear loop of the nonlinear power amplifier is considered, the performance of the AI receiver is far superior to that of the LS channel estimation algorithm in traditional communication.
Claims (5)
1. A working method of an AI receiver based on a deep learning technology is characterized by comprising the following steps;
in the data transmission process, data of a transmitting end is divided into a pilot signal and a useful signal, in one frame of data, the pilot signal is in front of the useful signal, the useful signal is followed by the useful signal, a sample signal is obtained at a receiving end after channel transmission, the pilot signal is transmitted to the receiving end to obtain a pilot sample signal, the useful signal is transmitted to the receiving end to obtain a useful signal sample, when the data is transmitted, the pilot sample reaches the receiving end first, the pilot signal is known by a receiver and a transmitter, the pilot signal and the pilot sample are input to a channel characteristic extractor, the channel characteristic extractor is realized by a fully-connected neural network, the channel characteristic extractor estimates channel information according to the transmitted pilot signal and the pilot receiving sample, the pilot signal is changed into the pilot sample after actual channel action, the difference between the pilot signal and the pilot sample is caused by a channel, the neural network of the channel characteristic extractor intelligently estimates the channel information according to the pilot and the sample, and finally the output of the channel characteristic extractor represents the channel information;
an AI receiver based on deep learning technology comprises three Neural Networks (NN), a channel feature extractor, a signal feature extractor and a signal detection classifier, wherein the channel feature extractor estimates and extracts channel information according to signals, the output of the channel feature extractor is fed to the signal feature extractor to help the signal feature extractor to extract feature signals after eliminating channel influence, and the feature signals are finally sent to the signal detection classifier to be recovered.
2. The method as claimed in claim 1, wherein the data symbol frame format actually transmitted by the communication system of the signature signal is: the method comprises the steps that pilot frequency + useful signals are used, original data are mapped to corresponding constellation points through grouping at a transmitting end and are transmitted through a transmitter; the signal is propagated through an actual channel, a sample signal obtained by a receiving end is composed of a pilot signal sample and a useful signal sample, and an AI receiver performs signal detection according to the received samples.
3. The method of claim 1, wherein the symbols transmitted by the transmitter in digital communication are from a corresponding constellation diagram, and the corresponding transmission signal is from a finite set Γ = { s }, in which 1 ,s 2 ,...,s m And transmitting a signal to a receiving end after the signal passes through a channel, wherein the received signal is y = f (x), the mapping f is related to the actual channel characteristic, and the communication aims to recover the original transmitted signal as much as possible according to the received signal y.
4. The method as claimed in claim 1, wherein after the desired signal sample arrives, the signal sample and the output of the channel feature extractor are used as inputs to be fed into the channel equalization network, the channel equalization network performs equalization processing on the signal sample according to the channel information, and removes information related to a channel in the signal sample as much as possible, the parameters of the channel equalization network are not trainable and are generated by the channel feature extraction network, the data transmission process of the neural network is that the input is multiplied by a network weight coefficient, so that when the parameters of the channel equalization network are generated by the output of the channel feature extraction network, the weight coefficient of the channel equalization network is channel state information, when the desired signal sample arrives, an operation similar to multiplication is performed after passing through the channel equalization network, the output of the channel equalization network is a signal feature vector, and finally the signal classifier performs signal detection according to the signal feature vector.
5. The method as claimed in claim 4, wherein the channel equalization network introduces a link to the receiver module during the training phaseThe discriminator and the signal classifier are in the same position, the link discriminator is a neural network, the input of the link discriminator is a signal characteristic vector and a sending signal, the link discriminator is used for estimating the channel state experienced by the current channel as much as possible according to the input, the link discriminator and the signal detection network have respective optimization targets, the optimization targets of the channel characteristic extraction network and the channel equalization network are that the link discriminator cannot realize accurate judgment according to the signal characteristic vector as much as possible, but the signal detection network can detect according to the signal characteristic network, namely, the channel related information in the sample signal is eliminated by the channel equalization network, so the link discriminator cannot carry out accurate estimation based on the signal characteristic vector, and the probability of final network guess tends to:
the signal detection network enables high quality signal decisions, which means that the AI receiver learns how to perform channel estimation and channel equalization based on the sample signal.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210337246.2A CN114696933B (en) | 2022-04-01 | 2022-04-01 | AI receiver based on deep learning technology and use method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210337246.2A CN114696933B (en) | 2022-04-01 | 2022-04-01 | AI receiver based on deep learning technology and use method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114696933A CN114696933A (en) | 2022-07-01 |
| CN114696933B true CN114696933B (en) | 2023-02-07 |
Family
ID=82140440
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210337246.2A Active CN114696933B (en) | 2022-04-01 | 2022-04-01 | AI receiver based on deep learning technology and use method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114696933B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107070603A (en) * | 2017-04-28 | 2017-08-18 | 电子科技大学 | Space-time block code system signal method of sending and receiving |
| CN109246038A (en) * | 2018-09-10 | 2019-01-18 | 东南大学 | A kind of GFDM Receiving machine and method of data model double drive |
| CN112637093A (en) * | 2020-12-09 | 2021-04-09 | 齐鲁工业大学 | Signal detection method based on model-driven deep learning |
| CN112637094A (en) * | 2020-12-17 | 2021-04-09 | 南京爱而赢科技有限公司 | Multi-user MIMO receiving method based on model-driven deep learning |
| CN113014524A (en) * | 2021-03-03 | 2021-06-22 | 电子科技大学 | Digital signal modulation identification method based on deep learning |
| CN113411122A (en) * | 2021-05-08 | 2021-09-17 | 西安理工大学 | Solar blind ultraviolet light communication self-adaptive signal detection method based on deep learning |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111683024B (en) * | 2020-06-01 | 2021-06-25 | 西北工业大学 | Time-varying OFDM system channel estimation method based on deep learning |
-
2022
- 2022-04-01 CN CN202210337246.2A patent/CN114696933B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107070603A (en) * | 2017-04-28 | 2017-08-18 | 电子科技大学 | Space-time block code system signal method of sending and receiving |
| CN109246038A (en) * | 2018-09-10 | 2019-01-18 | 东南大学 | A kind of GFDM Receiving machine and method of data model double drive |
| CN112637093A (en) * | 2020-12-09 | 2021-04-09 | 齐鲁工业大学 | Signal detection method based on model-driven deep learning |
| CN112637094A (en) * | 2020-12-17 | 2021-04-09 | 南京爱而赢科技有限公司 | Multi-user MIMO receiving method based on model-driven deep learning |
| CN113014524A (en) * | 2021-03-03 | 2021-06-22 | 电子科技大学 | Digital signal modulation identification method based on deep learning |
| CN113411122A (en) * | 2021-05-08 | 2021-09-17 | 西安理工大学 | Solar blind ultraviolet light communication self-adaptive signal detection method based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| 优化的神经网络分类器在自动调制识别中的应用;程莉;《工程研究-跨学科视野中的工程》;20130925(第03期);全文 * |
| 基于深度信念网络的无线信道二分查找方法;毛勇华 等;《计算机工程》;20180731;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114696933A (en) | 2022-07-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| RU2328081C2 (en) | Hybrid correction method and unit with decision feedback | |
| CN109905337B (en) | Channel equalization method based on NARX neural network and block feedback | |
| CN110429965B (en) | Method for detecting uplink signal of large-scale multi-input multi-output system | |
| JP3340190B2 (en) | Decision feedback equalization method and apparatus | |
| CN107566307A (en) | Blind equalizing apparatus and method, data modulation system and method | |
| He et al. | Deep learning-assisted TeraHertz QPSK detection relying on single-bit quantization | |
| Wong et al. | Estimation of transmitter I/Q imbalance using convolutional neural networks | |
| Zhang et al. | Machine learning for joint channel equalization and signal detection | |
| Ye et al. | Bilinear convolutional auto-encoder based pilot-free end-to-end communication systems | |
| CN112422208B (en) | Signal detection method based on antagonistic learning under unknown channel model | |
| CN114696933B (en) | AI receiver based on deep learning technology and use method | |
| Jang et al. | Combined deep learning and SOR detection technique for high reliability in massive MIMO systems | |
| Ranhotra et al. | Performance comparison of blind and non-blind channel equalizers using artificial neural networks | |
| Brink et al. | Deep networks for equalization in communications | |
| Corral et al. | Time-varying channel neural equalisation using Gauss-Newton algorithm | |
| Wang et al. | Signal Detection for IoT Networks with Unknown Channel Models: A Knowledge-Driven Approach | |
| WO2021262052A1 (en) | A context aware data receiver for communication signals based on machine learning | |
| Hanoon et al. | Performance Evaluation of SISO-OFDM Channel Equalization Utilizing Deep Learning | |
| Gorday et al. | LMS to deep learning: How DSP analysis adds depth to learning | |
| Sahu et al. | Neural network training using FFA and its variants for channel equalization | |
| Feng et al. | The design of satellite equalizer based on CNN | |
| Muthumari et al. | Analysis: a neural network equalizer for channel equalization by particle swarm optimization for various channel models | |
| Enescu et al. | Recursive estimation of noise statistics in Kalman filter based MIMO equalization | |
| Muthulakshmi et al. | Signal demodulation without channel equalizer using machine learning techniques | |
| Pydi et al. | Design of Polynomial Function based Blind Equalizer using Particle Swarm Optimization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |