当前位置:网站首页>Application of deep learning in physical layer signal processing

Application of deep learning in physical layer signal processing

2020-12-08 11:09:11 Huawei cloud developer community

Abstract : This paper mainly introduces the application of physical layer based on deep learning , A new method based on depth is proposed Q The Internet (DQN) Of MIMO System location information verification scheme , The receiver exploits the depth in the changeable and unknown channel environment Q The network is constantly updated .

01 introduction

With the explosive growth of mobile traffic 、 Communication scenarios with high reliability and low delay bring more complexity and computational challenges to current networks . According to the IBM reports , Move data to 2020 The year will exceed 40 One trillion Gbits, Than 2009 Increase in 44 times , The total number of connected devices will reach 500 Billion . To meet this demand , New communication theories and innovative technologies are needed to meet 5G System requirements . In recent years, the development of deep learning paradigm has caused academic and Industrial Research on wireless communication technology based on deep learning , The results show that deep learning technology can improve the performance of wireless communication system , And it has the potential to be used in the physical layer for interference adjustment 、 Channel estimation and signal detection 、 Signal processing and so on .

02 Deep learning paradigm

The concept of deep learning comes from artificial neural networks (ANN) The study of , from Hinton And other people in 2006 in . Pictured 1 Shown , Deep learning through the establishment of hierarchical structure of ANN, It often contains an input layer 、 Multiple hidden layers and one output layer . Each layer uses different weights to connect with adjacent layers , By extracting and filtering the input information layer by layer , It can realize end-to-end supervised learning and unsupervised learning . Deep neural networks include feedforward neural networks (FNN)、 Cyclic neural network (RNN)、 Convolutional neural networks (CNN)、 Against generative networks (GAN) And deep belief networks . One of them is based on gating RNN, For example, long-term and long-term memory (LSTM) The network has a certain memory function for input , Therefore, it is often used in physical layer signal processing and channel state information estimation . Besides , Deep learning can also participate in building reinforcement learning (RL) System , Form deep reinforcement learning , For example, depth Q The Internet (DQN)[1], It can be used to optimize the physical layer signal processing strategy .

1) Long and short term memory network

As RNN A variation of , Long and short term memory network can effectively solve the problem of gradient explosion or disappearance of simple cyclic neural network .RNN Historical information is stored in a hidden state . In simple RNN in , Every moment of the hidden state is rewritten , Therefore, it can be regarded as a kind of short-term memory . And in the LSTM In the network , Memory units hold key information longer than short-term memory .LSTM Network access mechanism to control the path of information transmission . The value of gate mechanism is in 0 To 1 Between , To control the proportion of information passing through .LSTM The network mainly includes 3 A door , The forgetting gate controls how much information the internal state of the last moment needs to be forgotten ; The input gate controls how much information the candidate state holds at the current time ; The output gate controls how much information the current internal state needs to output to the external state .

2) depth Q The Internet

DQN take CNN And Q Combine learning , use Q The objective function of learning is used to construct the objective function of deep learning , Using memory playback mechanism to solve the problem of data relevance , The system stability is solved by iterative updating . Suppose the state of the environment at the moment is , Agent according to certain strategy To take action , And get a reward . then , The environment moves to the next state with transition probability at the moment . stay DQN in , The agent interacts with the environment through a series of actions , The goal is to maximize cumulative rewards .

meanwhile , It uses the experience playback based on convolution neural network Q Continuous approximation of functions . In the replay of experience , The agent uses ξ-greedy To select actions , And save the learning experience of each moment in the experience pool . In the parameter update loop of the algorithm , Random sampling or batch random sampling of samples in the memory pool , adopt Q Learn to update the model parameters . And pass CNN Based on previous experience , Keep approximating the largest Q value .CNN The loss function is approximate Q Value and reality Q The deviation between values , The weight of neural network is adjusted by gradient descent algorithm , The value of the loss function can be reduced continuously .

03  Physical layer signal processing application based on deep learning

In recent years , There have been some deep learning applications in the physical layer in academia and industry , Deep learning can improve physical layer performance . This section from the physical layer signal processing point of view , From channel status information (CSI) It is estimated that 、 Signal encoding and decoding 、 Four aspects of interference adjustment and signal detection are given to illustrate the existing related work .

1) Based on deep learning CSI It is estimated that

The precise CSI Acquisition is very important to guarantee the link performance of wireless communication system . The wireless network chooses the specific signal control scheme according to the channel estimation state , for example , When CSI When it's low , The physical layer adopts low order modulation scheme to resist the bad communication state, so as to reduce the bit error rate .5G The communication system adopts multiple input and multiple output (MIMO)、 Millimeter wave and non orthogonal multiple access (NOMA) Technology , So that both sides of the communication have more transmission channels , Channel estimation becomes more complex . Conventional CSI The estimation scheme needs to perform matrix operations with high complexity , It is limited by computing resources and time delay .

Use deep learning to get CSI The correlation between information space-time and uplink and downlink , It has been proved that it can improve CSI Estimated efficiency , And reduce the amount of data needed for uplink and downlink reference information [2]. Pictured 2 Shown , The paper [3] Put forward history CSI The data is extracted by a two-dimensional convolution neural network , Then a one-dimensional convolution neural network is used to extract the state eigenvector from the frequency eigenvector . Last , One LSTM The Internet is used for CSI State prediction . Because the two-dimensional convolution neural network was originally used to process image data , therefore , The author will CSI Raw data is divided into cells , Each cell corresponds to a picture pixel . Of each band CSI The pixels corresponding to the auxiliary information form a channel . therefore ,N The data of each frequency band will be converted into N Pixel information of channels , And input it into the learning framework .

2) Codec based on deep learning

The application of deep learning in source coding and channel coding , It is also proved that it can improve the coding efficiency and reduce the network's BER. The joint coding scheme based on deep learning framework can encode the source of this paper through the cyclic neural network ( structured ), And then you put structured information into two-way LSTM The Internet , And finally output the final transmission of binary data stream . At the receiving end ,LSTM For decoding processing . The paper [4] The encoder with full connection depth neural network is proposed , Used to improve the confidence propagation algorithm based on HPDC Decoding efficiency .O’Shea Et al. [5] The entire physical layer is modeled as one that contains modulation 、 Self encoder for channel coding and signal classification functions , The convolution neural network is used to train the self encoder . Pictured 3 Shown , In the learning framework of multi dense layer neural network , The input signal is encoded as a single hot code (One-hot encoding), The wireless channel is modeled as a noise layer . Cross entropy loss function and stochastic gradient descent algorithm are used to train the model , At the output, the output signal with the highest probability is taken as the decoding result .

3) Interference adjustment based on deep learning

MIMO The interference adjustment in the system uses the Linear Precoding technique to adjust the transmitted signal , The interference signal at the receiver can be controlled in a reduced dimension subspace , To break through MIMO Throughput limitation caused by system interference . In the existing work, the results have shown that , Using deep learning can improve the throughput of interference adjustment network , And get the optimization results .He Et al. [6] The adoption of DQN To obtain the optimal user selection strategy under interference adjustment . In this mechanism , The central scheduler is used to collect all channel status and cache status for each user , And allocate channel resources to each user . The time-varying process of channel is modeled by a finite state Markov model , The state of the system is defined as the channel state and cache state of each user . The central scheduler is used to train the best strategy for the system , The corresponding system action is defined as whether to allocate channel resources to each user for data transmission , To maximize interference and adjust network throughput .DQN It can also be used to eliminate the interference between secondary users and primary users in cognitive radio networks , Secondary users use frequency hopping and mobility to defend against jammers [7].

4) Signal detection based on deep learning

be based on DL The detection algorithm can significantly improve the performance of communication system , Especially when the traditional processing module needs joint optimization or the channel cannot be characterized by the common analysis model . The paper [8] A five layer fully connected DNN The framework is embedded in OFDM Joint channel estimation and signal detection in the receiver . The received signal and the corresponding transmission data and pilot are used as input ,DNN You can infer channel information , And it can be used to predict the data sent . stay MIMO It's being tested , The iterative method based on Bayesian optimal detector has been proved to have better performance and moderate computational complexity . But in many more complex environments , Unknown channel distribution conditions will limit the effectiveness of the detector . Using deep learning algorithms , The model parameters can be recovered according to certain input data , So as to improve the adaptive ability of the detector . meanwhile , In some cases , Deep learning algorithm can also make use of some semantic information , For example, the location of the receiver and the information of the surrounding vehicle nodes , To do beam prediction , To improve system performance .

04  be based on DQN Signal detection mechanism of

In a location-based service scenario , Vehicles or users need to constantly send beacon messages to report their location , So as to improve location services and network performance . But some vehicles or users will choose to send fake locations to get more resources , Affect the effectiveness of network services .

stay MIMO In the system , Transmission signals often contain rich information ( Arrival angle 、 Receiving power, etc ) Signal detection technology can be used to verify the location of beacon message at the receiving end . We propose that it is based on DQN Signal detection mechanism of , It can be used for MIMO In the system, the sender's location information verification and the detection of information forger . The main idea is , The receiver uses maximum likelihood estimation to test the received signal , When the received signal passes the detection test , It is considered that the transmitted signal comes from the location reported by the sender . otherwise , Think the sender reported false location information . In order to improve the detection performance in the changeable channel state , At the receiving end, based on DQN To predict the benefits of different detection thresholds , And select the optimal detection threshold . The system framework is shown in the figure 4 Shown .

1) System model   Suppose the zero false in the test is set as the sending node reports the real location information , The alternative hypothesis reports false location information for the sending node . At every moment , The receiver receives the signal from the sender and the real position between the sender and the receiver 、 The channel state is related to the angle of arrival . Under the condition of known transmission information and transmission power , The receiver can use maximum likelihood detection to test the received signal .

2) Maximum likelihood detection   The receiver uses the maximum likelihood detection algorithm to verify the received signal , Detection rules are defined as :

Which represents the detection threshold , The value range is . And represent the normal and false reports respectively . And are the posterior distributions of the observed signals under the null hypothesis and the alternative hypothesis respectively . according to [9] Available , The result of hypothesis test ( False positive rate and loss rate ) With the actual location of the sender 、 Reporting location 、 Channel condition is related to detection threshold . For the receiver , The actual location of the sender 、 The reported location and channel status are unknown or partially known environmental variables , In the process of continuous information exchange with the sender , In this paper, the receiver can put forward DQN To continuously optimize the selection of detection threshold , So as to improve the accuracy of signal detection .

3) be based on DQN Detection threshold optimization of

In the mechanism proposed in this paper , The state space of the receiver is divided into two dimensions , The first dimension is the channel state from sender to receiver , The second dimension is the result of channel detection . The channel state space includes a series of channel indexes after quantization , It is assumed that the state transition of the channel follows the Markov process , That is, the state of the channel at the current time is only related to the state of the previous time . As a result, there are four kinds of state space : Real data test results are true 、 The real data test result is false ; False data detection results are true and false data detection results are false . In every move , The direct reward at the receiver is related to the test results , When the test results are correct, you can get a positive benefit , When the detection result is wrong, the negative benefit is obtained . The action of receiver is defined as the threshold of signal detection , A series of quantitative spatial detection thresholds include . At every moment , The receiver's hybrid strategy is to choose the probability of different detection threshold . Based on the introduction of the second chapter of this article DQN principle , After each experience, the receiver , Test threshold of your choice 、 The corresponding status results and benefits are stored in the experience pool , utilize CNN Yes Q Function for training prediction , Continuously optimize the selection of detection threshold .

05  Summary and suggestions for future development

In this paper , We have proved the great potential of deep learning in physical layer communication through existing work and cases . In addition to the several application directions described above , Deep learning has also been applied in end-to-end communication system . however , At present, there is no conclusion whether the performance of end-to-end communication system based on deep learning will eventually exceed the performance of traditional communication system . in addition , Physical layer applications based on deep learning need data driven , In order to improve the training efficiency of deep learning model , It is possible to fuse modules that require long training , And you need to consider the trade-off between good performance and training efficiency . The rise of deep learning applications is largely due to the availability of various datasets , However, there are still few data sets for wireless communication . Data security and privacy issues further limit the access to communication data in the real world . But for communication applications based on deep learning , Need some open telecom data sets to publish and share . Last ,5G Complex and changeable communication environment , Include MIMO、 Millimeter wave communications and NOMA Technology, etc. , It also brings great potential for the application of deep learning .

reference

[1] Mnih, Volodymyr, et al. "Human-levelcontrol through deep reinforcement learning." Nature 518.7540(2015): 529. https://www.nature.com/articles/nature14236.

[2] A. Mousavi and R. G. Baraniuk, “Learning toInvert: Signal Recovery via Deep Convolutional Networks,” Proc. IEEE Int’l.Conf. Acoustics Speech Signal Process. (ICASSP’17), New Orleans, LA, Mar. 2017,pp. 2272–76.

[3] C. Luo, J. Ji, Q. Wang, X. Chen and P. Li,"Channel State Information Prediction for 5G Wireless Communications: ADeep Learning Approach," in IEEE Transactions on Network Science andEngineering, early access.

[4] E. Nachmani, Y. Be’ery, and D. Burshtein,“Learning to decode linear codes using deep learning,” in Proc. Communication,Control, and Computing (Allerton), 2016, pp. 341–346.

[5] T. O’Shea and J. Hoydis, "An Introduction to Deep Learning for thePhysical Layer," in IEEE Transactions on Cognitive Communications andNetworking, vol. 3, no. 4, pp. 563-575, Dec. 2017.

[6] Y. He, C. Liang, F. R. Yu, N. Zhao, and H.Yin, “Optimization of cache-enabled opportunistic interference alignmentwireless networks: A big data deep reinforcement learning approach,” in Proc.IEEE Int. Conf. Commun. (ICC), May 2017, pp. 1–6.

[7] G. Han, L. Xiao, and H. V. Poor,“Two-dimensional anti-jamming communication based on deep reinforcementlearning,” in Proc. IEEE Int. Conf. Acoust. Speech Signal Process. (ICASSP),New Orleans, USA, Mar. 2017, pp. 2087–2091.

[8] H. Ye, G. Y. Li, and B.-H. F. Juang, “Power ofDeep Learning for Channel Estimation and Signal Detection in OFDM Systems,”IEEE Wireless Commun. Lett., vol. 7, no. 1, Feb. 2018, pp. 114–17.

[9] Bai, Lin, Jinho Choi, and Quan Yu. “SignalProcessing at Receivers: Detection Theory.” Low Complexity MIMO Receivers,Springer, Cham, 2014. pp.5-28.

This article is shared from Huawei cloud community 《 Research on the application of deep learning in signal processing of physical layer 》, Original author : Suddenly .

 

Click to follow , The first time to learn about Huawei's new cloud technology ~

版权声明
本文为[Huawei cloud developer community]所创,转载请带上原文链接,感谢
https://chowdera.com/2020/12/20201208110854516o.html