WO2021203242A1 - 基于深度学习的mimo多天线信号传输与检测技术 - Google Patents

基于深度学习的mimo多天线信号传输与检测技术 Download PDF

Info

Publication number
WO2021203242A1
WO2021203242A1 PCT/CN2020/083563 CN2020083563W WO2021203242A1 WO 2021203242 A1 WO2021203242 A1 WO 2021203242A1 CN 2020083563 W CN2020083563 W CN 2020083563W WO 2021203242 A1 WO2021203242 A1 WO 2021203242A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
neural network
mimo
deep learning
transmitting
Prior art date
Application number
PCT/CN2020/083563
Other languages
English (en)
French (fr)
Inventor
刘婵梓
曲春晓
陈高
刘新宇
周清峰
Original Assignee
东莞理工学院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 东莞理工学院 filed Critical 东莞理工学院
Priority to CN202080000955.0A priority Critical patent/CN111630787B/zh
Priority to CN202310024809.7A priority patent/CN116054888A/zh
Priority to PCT/CN2020/083563 priority patent/WO2021203242A1/zh
Publication of WO2021203242A1 publication Critical patent/WO2021203242A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • 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
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • 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
    • H04L25/0202—Channel estimation
    • H04L25/024—Channel estimation channel estimation algorithms
    • H04L25/0254—Channel estimation channel estimation algorithms using neural network algorithms
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Definitions

  • the present invention relates to the field of mobile communication technology, and in particular to MIMO multi-antenna signal transmission and detection technology based on deep learning.
  • MIMO Multiple-Input Multiple-Output
  • SISO Single-Input Single-Output
  • MIMO includes the SIMO (Single-Input Multiple-Output) system and the MISO (Multiple-Input Single-Output) system. It can make full use of space resources and achieve multiple transmissions and multiple receptions through multiple antennas.
  • MIMO systems have been widely used In wireless communication-mobile devices and networks generally use multiple antennas to enhance connectivity, improve network speed and user experience. Massive MIMO is a key factor in the ultra-high data rate of 5G, which can bring greater network capacity, wider signal coverage and better user experience, and bring the potential of 5G to a whole new level.
  • space diversity refers to the use of multiple transmitting antennas to send signals with the same information through different paths, and at the same time to obtain multiple independently fading signals of the same data symbol at the receiver end, thereby obtaining improved reception reliability with diversity.
  • space diversity refers to the use of multiple transmitting antennas to send signals with the same information through different paths, and at the same time to obtain multiple independently fading signals of the same data symbol at the receiver end, thereby obtaining improved reception reliability with diversity.
  • space diversity refers to the use of multiple transmitting antennas to send signals with the same information through different paths, and at the same time to obtain multiple independently fading signals of the same data symbol at the receiver end, thereby obtaining improved reception reliability with diversity.
  • N r the maximum diversity gain can be obtained as N r .
  • the gain of multiple paths is also used to improve the reliability of the system.
  • the maximum diversity gain that can be obtained is N t N r .
  • the unreliability of wireless communication is mainly caused by the time-varying and multipath characteristics of wireless fading channels. How to reduce the impact of multipath fading on base stations and mobile stations without increasing power and sacrificing bandwidth is very important. The only way is to use anti-fading technology, and the effective method to overcome multipath fading is various diversity technologies. Diversity technology is mainly used to combat channel fading. On the contrary, the fading characteristics in the MIMO channel can provide additional information to increase the degree of freedom in communication.
  • the transmission data rate can be provided, which is called spatial multiplexing.
  • SNR Signal to Noise Ratio
  • MIMO relies on its two major advantages: (1) Improve the capacity of the channel. Between the MIMO access point and the MIMO client, multiple spatial streams can be sent and received at the same time. The channel capacity can increase linearly with the increase in the number of antennas. Therefore, the MIMO channel can be used to double the wireless channel capacity. Without increasing the bandwidth and antenna transmission power, the spectrum utilization rate can be doubled; (2) Improve the reliability of the channel. Using the spatial multiplexing gain and spatial diversity gain provided by the MIMO channel, multiple antennas can be used to suppress channel fading. The application of the multi-antenna system enables parallel data streams to be transmitted at the same time, which can significantly overcome channel fading and reduce the bit error rate. It has become a core technology used in 802.11n. 802.11n is IEEE's new wireless LAN technology after 802.11a/b/g, with a speed of up to 600Mbps. At the same time, MIMO technology can improve the performance of existing 802.11a/b/g networks.
  • Artificial intelligence refers to the intelligence displayed by machines made by humans.
  • Artificial intelligence refers to the technology of presenting human intelligence by simulating some human thinking processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.) through ordinary computer programs. It mainly includes the principles of computer realization of intelligence and manufacturing similar to humans. Brain-smart computers enable computers to achieve higher-level applications. We firmly believe that in the next few years, artificial intelligence will continue to "promote directly", create value for traditional industries, and profoundly change our daily lives, such as robotics, language recognition, image recognition, expert systems, etc.
  • Machine learning is a subset of artificial intelligence. At present, machine learning mainly solves classification problems, clustering problems, regression problems, etc., and has been widely used in character recognition, machine translation, speech recognition, search engines, face recognition, unmanned driving and other fields. The most critical of all machine learning algorithms is deep learning.
  • Deep learning comes from the research of artificial neural networks.
  • the concept of deep learning was first put forward by Professor GE. Hinton, a dean in the field of machine learning. His two core views are: (1) Artificial neural networks with multiple hidden layers have excellent feature learning capabilities, which are very useful for learning. The characteristic data of the network has a deeper display, which can better classify or visualize the final network data; (2) Deep neural networks can overcome the difficulty of training their own network parameters through "layer-by-layer initialization", while initializing layer by layer This can be achieved through unsupervised learning.
  • a multilayer perceptron with multiple hidden layers is a deep learning structure. Deep learning forms a more abstract high-level representation attribute category or feature by combining low-level features to discover distributed feature representations of data.
  • Any neural network can have any number of layers, inputs or outputs. The layer between the input neurons and the last layer of output neurons is the hidden layer of the deep neural network.
  • Existing technology such as the authorization announcement date of 20180619 and the authorization announcement date of CN105610477B, proposed a compressed sensing-based multiple-transmit-multiple-receive system enhanced signal multiplexing method.
  • the random measurement matrix in the compressed sensing technology is selected as the signal compression multiplexing matrix, and then make full use of the sparse characteristics of the transmitted signal on the over-complete redundant dictionary, and reconstruct through compressed sensing
  • the algorithm extracts the high-dimensional transmission signal from the low-dimensional received multiplexed signal, thereby greatly improving the signal multiplexing gain under the condition of the given number of transmit and receive antennas of the MIMO system, and better meet the application requirements of the MIMO system for broadband transmission , And has the advantage of ensuring that the receiving end uses the mature optimized reconstruction algorithm in the compressed sensing field to reconstruct the multiple data streams sent by the transmitting end through the compression multiplexing step with a high probability, and has the advantages of small modification
  • the Chinese patent application number 201510473741.6 discloses a complex neural network channel prediction method, which mainly solves the channel fading problem caused by channel time variation in the MIMO system.
  • the technical solution is: 1.
  • the base station measures the channel to obtain a channel coefficient training sequence containing estimation errors; 2. Obtains corresponding training samples and expected output according to the obtained channel coefficient sequence; 3. Inputs training samples for complex wavelet neural network training , Get the final network weight; 4.
  • the base station uses the trained complex wavelet neural network to predict the channel coefficients.
  • the method is simple and easy to implement, has good effects, and is suitable for reducing the influence of the channel time change on the MIMO system channel.
  • the Chinese patent application number 201810177829.7 discloses a method for modeling a wireless channel based on a neural network.
  • the present invention first processes the received signal fed back by the user to obtain estimated channel parameters; then obtains the three-dimensional geographic information of the scatterers according to the two-dimensional image, and clusters them, and finally uses the channel parameters and geographic information as the input of the neural network , Receive the signal as the output, and train the nonlinear time-varying neural network model.
  • This method obtains a more accurate channel model within acceptable complexity, and can meet the channel modeling requirements of scenarios such as massive MIMO technology, frequency band expansion, and high mobility used in future 5G communication systems.
  • the Chinese patent application number 201810267976.3 (Deep neural network massive MIMO system detection method based on BP algorithm, Southeast University) provides a deep neural network massive MIMO system detection method based on BP algorithm, which is iteratively propagating beliefs
  • the algorithm factor graph is expanded and mapped to the neural network structure to construct a deep neural network for large-scale MIMO system detection; the neurons of the deep neural network correspond to the nodes in the iterative algorithm factor graph, and the number of neurons in each layer is equal to the iteration
  • the number of symbol nodes in the algorithm factor graph; the mapping function between hidden layers is the update formula of the confidence information in the iterative algorithm, and the number of hidden layers is equal to the number of iterations of the iterative algorithm.
  • the present invention also provides a MIMO detection method for constructing two deep neural networks based on the two information propagation iterative algorithms of damped confidence propagation and maximum and confidence propagation respectively.
  • the invention achieves a lower bit error rate without increasing the online operation complexity, and has robustness to various channel conditions and antenna ratios.
  • the Chinese patent application number 201610327115.0 (a codebook selection method based on deep learning under massive MIMO, Chongqing University of Posts and Telecommunications) relates to a codebook selection method based on deep learning under massive MIMO.
  • the method includes: collecting pilot information in the test area to construct a pilot training sequence, and then obtaining pilot training samples; performing neural network iterative learning on the pilot training samples to obtain the final network weight value; according to the output of the learned neural network Channel, select the optimal codeword from the complete codebook. Then, the unknown area and the test area are matched with channel information to obtain the wireless channel, and then the codeword corresponding to the wireless channel is obtained.
  • the invention can effectively, accurately and quickly establish a wireless channel model and codebook query, avoids channel estimation in an unknown area and greatly reduces the complexity of channel selection codebook in an unknown area.
  • the Chinese application number 201811626005.X patent discloses a low-complexity MIMO-NOMA system signal detection method based on an improved gradient projection method, involving wireless Communication technology.
  • the idea of convex optimization algorithm is used to transform the system model into a strict quadratic programming problem; then the problem is solved iteratively, and the results of each iteration are preprocessed to achieve the goal The signal is effectively detected.
  • the invention breaks through the problem of slow algorithm convergence in the traditional detection method. Preprocessing the results of each iteration can not only make the detection results quickly converge, but also detect the set of active users.
  • the implementation process is simple and the application range is wide. .
  • the Chinese patent application number 201910014714.0 discloses a method for designing beamforming matrix of MIMO system based on deep learning.
  • the steps are as follows. First, use a known algorithm to obtain the depth Learn the training sample set required by the network; then build a deep learning neural network model, initialize the relevant parameters of the model and use the training sample set for training; then use the pilot to obtain the channel and send it to the neural network to predict the beamforming matrix coefficients, and finally the channel and The beamforming matrix coefficients are combined to form the beamforming matrix.
  • the method uses the beamforming matrix obtained by the deep learning neural network to take into account the performance and the algorithm complexity at the same time, and can reduce the delay under the premise of ensuring the performance, so that the MIMO system can provide real-time services.
  • the Chinese patent application number 201810182937.3 discloses a machine learning-based MIMO link adaptive transmission method, which uses an unsupervised learning self-encoding algorithm for feature extraction With dimensionality reduction, the introduction of the idea of deep learning can reduce the feature dimension and computational complexity on the premise of preserving the main information state information.
  • the present invention uses a logistic regression algorithm to construct the mapping relationship between channel state information and transmission parameters, which is different from the fixed parameterized model in the past and can be trained based on sample data. Under the condition that the quality of the data set is better and covers all states, it can be more improved.
  • the mapping relationship between channel state information and transmission parameters can be established well, and compared with the traditional single equivalent signal-to-noise ratio, the channel state information can be more fully utilized.
  • the present invention also selects CQI based on the channel matrix, and studies the MIMO link adaptation method based on machine learning through the channel matrix and noise variance, which is not restricted by receiver design and has universal applicability.
  • the Chinese patent application number 201710495044.X discloses a joint precoding and antenna selection method for MIMO systems based on deep learning, which includes the following steps: First, use the existing antenna selection method to generate the training data set required for deep learning; then, establish a deep learning model, use the training data to train the deep learning model and save it; then use the saved deep learning model to complete the antenna selection; finally
  • the MIMO subsystem is designed for optimal precoding.
  • the invention utilizes deep learning technology to design MIMO system joint precoding and antenna selection, and can achieve lower computational complexity under the condition of obtaining a good system signal-to-noise ratio.
  • Chinese Patent Application No. 201910242525.9 proposes a deep signal detection method for high-speed rail.
  • the high-speed rail first determines its location information through GPS during driving, judges the area where it is located, and selects the corresponding neural network model, and then trains the real-time received signal input In the neural network, the signal sent by the base station is output in real time.
  • the system performance of the present invention is greatly improved, the signal detection bit error rate is reduced, and the algorithm is more robust.
  • the method used in the present invention does not need to estimate the channel and saves pilot overhead.
  • Chinese application number 201810279530.2 invention patent discloses a visible light communication MIMO anti-interference and noise reduction method based on BP neural network, involving MIMO in the field of visible light communication Antenna technology.
  • the system includes the transmitting end of the system device, the signal processing part of the receiving end of the system device and the signal processing part of the BP neural network which are connected in sequence.
  • This method is as follows: 1) The electrical signal is loaded on the LED array and emitted in the form of light signal; 2) The photodetector at the receiving end of the device converts the light signal into an electrical signal; 3) Multiple electrical signals are removed by a low-pass filter High-frequency interference; 4) After training, the BP neural network performs noise reduction and interference elimination processing on multi-channel signals, and finally becomes a binary serial data stream through parallel-serial conversion.
  • the present invention improves the transmission performance of the existing MIMO technology; combines the neural network with the visible light MIMO technology to give play to the advantages of the neural network in removing interference and noise in wireless communication; and adopts the neural network receiving and processing technology to make the entire VLC system more stable.
  • Chinese Application No. 201710213235.2 Invention Patent discloses a joint equalization method of visible light channel based on orthogonal mapping and Probabilistic Neural Network, including device transmitter and device receiver The signal is transmitted from the transmitting end of the device to the receiving end of the device through the visible light MIMO channel; the visible light MINO channel is a multiple-input multiple-output channel; the joint equalization is a combination of pre-equalization and post-equalization; It is a joint equalization scheme that combines pre-equalization technology and post-equalization technology, that is, a visible light multiple-input multiple-output channel joint equalization method based on orthogonal mapping and probabilistic neural networks, which can effectively suppress the interference between channels of visible light MIMO communication systems. Improve the reliability of data transmission.
  • the embodiment of China Patent No. 201910125325.5 discloses a MIMO decoding method, device and storage medium based on deep learning.
  • the training data set includes multiple training data; then the neural network is trained based on the training data set to obtain the trained neural network model; when the MIMO signal to be decoded is received, the training data set will be decoded.
  • the MIMO signal is input to the neural network model for MIMO decoding, and then the MIMO decoding result output by the neural network model is obtained.
  • a neural network model for joint MIMO detection and channel decoding is designed based on deep learning, MIMO detection and channel decoding are regarded as a joint decoding process, and the approximation of the output result of the neural network model is improved through training , To ensure the overall performance of MIMO decoding, with higher decoding accuracy and faster decoding speed.
  • the Chinese Patent Application No. 201810757547.4 discloses a machine learning-assisted massive MIMO downlink user scheduling method, which includes the following steps: S1: the base station sends data from users The uplink detection signal obtains the characteristic mode energy coupling matrix in the characteristic direction; S2: The base station uses the characteristic mode energy coupling matrix to assist in the calculation of the sum rate under various user and beam combinations through the method of machine learning; S3: adopts the greedy algorithm to achieve the sum rate User scheduling based on the maximum rate criterion to obtain the optimal user beam pairing combination.
  • the present invention obtains statistical channel information through the uplink sounding signal, and uses the sum rate maximization criterion for user scheduling.
  • the base station only has statistical channel information
  • the approximate calculation of the rate and the rate can be accurately achieved, which greatly reduces the complexity of user scheduling under large-scale antennas, and the performance is close to the best. Excellent, with good applicability and robustness.
  • China 201610353881.4 invention patent (a modulation recognition method under MIMO related channels based on machine learning algorithms, Beijing University of Posts and Telecommunications) is a modulation recognition method under MIMO related channels based on machine learning algorithms, which belongs to the field of communications; the specific steps are as follows: : First, use space-time coding for each data stream at the transmitting end of the communication device, and each codeword is transmitted through Nt transmitting antennas; then, calculate the MIMO channel matrix H according to the correlation matrix of the receiving end of the device and the correlation matrix of the transmitting end of the device ;According to the MIMO channel matrix H, the received signal on each receiving antenna is calculated and corrected; finally, each receiving antenna extracts the characteristics of the corrected signal, performs training tests on the extracted eigenvalues, and calculates that the sample belongs to Modulation recognition mode; the advantages are: strong robustness and generalization ability for non-Gaussian channels, and modulation system recognition in more complex environments can be achieved through parameter iteration; by extracting the characteristics of high-order moments and high
  • MIMO technology makes space a resource that can be used to improve performance and can increase the coverage of wireless systems.
  • MIMO technology has become one of the key technologies in the field of wireless communication.
  • MIMO technology has been increasingly applied to various wireless communication systems.
  • the complexity of the implementation of the MIMO technology is greatly increased, which limits the number of antennas used and cannot fully utilize the advantages of the MIMO technology.
  • the neural network realizes the mathematical network through the computer, which is recognized as the "outlet" to solve some of the bottlenecks encountered in current communications.
  • the existing related research on MIMO conditions has not yet determined the number of transmitting antennas and receiving antennas.
  • This application combines the latest research progress of compressed sensing technology and neural network technology, and proposes a signal multiplexing transmission and detection scheme in a MIMO system based on deep learning and compressed sensing.
  • this application is to introduce a compression multiplexing module and a demultiplexing module at the receiving end of the device to reduce the number of antennas required while simultaneously transmitting the same amount of data, thereby increasing the multiplexing gain and capacity of the MIMO system; on the other hand, with Compared with the existing MIMO spatial multiplexing technical solutions that only focus on eliminating the interference of adjacent data, this application not only focuses on eliminating the interference of adjacent data, but also pays more attention to how to achieve On the basis of ensuring the detection performance of the receiving end of the device, more data streams are multiplexed and transmitted to the receiving end of the device to obtain multiplexing gain and transmission capacity that exceed the inherent MIMO system.
  • the MIMO device based on deep learning and compressed sensing proposed in this application is mainly composed of the compression multiplexing module 104 of the transmitting end 1 of the device, and the first neural network signal processing module 202 and the second neural network signal processing module 203 of the receiving end 2 of the device. Constituted.
  • a random number generator 101 or original information bit generation module
  • a bit-level processing module 102 or a modulation module 103 are further provided in the transmitting terminal 1 of the device
  • a channel estimation module 201 is also provided in the receiving terminal 2 of the device.
  • the device transmitting terminal 1 the original data generated by the random number generator 101 passes through the bit-level processing module 102 and the modulation module 103 successively to generate a modulated signal, and the modulated signal is then compressed
  • the multiplexing module 104, the compression multiplexing module 104 performs compression and dimensionality reduction and multiplexing processing on the transmitted signal, and the compressed and multiplexed signal is transmitted through the transmitting antenna
  • the device receiving end 2 the channel estimation module 201 performs channel on the received signal It is estimated that the input of the first neural network signal processing module 202 is obtained based on the received signal and the estimated channel state information, the input of the second neural network signal processing module 203 is obtained based on the output of the first neural network signal processing module 202, and the second The neural network model 203 reconstructs and outputs the original transmission data stream x.
  • the above process specifically includes the following:
  • a MIMO multi-antenna signal transmission and detection device based on deep learning includes: a compression multiplexing module for compressing and dimensionality reduction of the modulated signal; the transmitter of the device is used for a given number of transmitting and receiving antennas The target signal processed by the compression multiplexing module is transmitted through the transmitting antenna; the receiving end of the device is used to process the received signal to realize the reconstruction of the target signal, and the receiving end of the device includes the first The neural network signal processing module and the second neural network signal processing module, wherein the second neural network signal processing module uses the first neural network model constructed by the first neural network signal processing module to extract the low-dimensional target signal The resolved high-dimensional sparse signal ⁇ is input into the second neural network model constructed by it to reconstruct the original signal x.
  • the receiving end of the device further includes a channel estimation module configured to perform channel estimation based on the low-dimensional target signal received by the receiving end of the device and subjected to compression and dimensionality reduction processing, and to obtain channel parameters
  • the matrix is used as the input of the first neural network model.
  • the first neural network signal processing module uses a deep learning BP algorithm to create a neural network.
  • the sparse representation ⁇ is used as a sample, a first training sample set is constructed, the neural network is trained, and the first neural network model is obtained.
  • the redundant dictionary D is formed by taking all possible combinations of the transmission signal vector x as different column vectors of the redundant dictionary, so as to realize the sparse representation ⁇ of the transmission signal vector.
  • the second neural network signal processing module uses a deep learning BP algorithm to create a neural network.
  • the determined sparse representation ⁇ is used as a sample, a second training sample set is constructed, the neural network is trained, and the second neural network model is obtained.
  • the transmitting end of the device uses a random number generator to generate a set of random 0 and 1 binary bit sequences to form the original data; the original data is modulated by BPSK to generate a modulated signal x.
  • the modulated 1 signal x passes through the compression multiplexing module and is compressed into a ⁇ 1 signal, and then the device transmits the compressed and multiplexed signal z via a transmitting antenna.
  • a MIMO multi-antenna signal transmission and detection system based on deep learning constructs a MIMO end-to-end transmission model and obtains a neural network signal based on the target signal received by the receiving end of the device after compression and dimensionality reduction processing and the estimated channel state information
  • the processed input uses neural network signal processing to reconstruct the original signal.
  • a MIMO multi-antenna signal transmission and detection method based on deep learning includes at least one of the following steps: when the number of transmitting and receiving antennas is given, the target signal after the compression multiplexing process is transmitted through the transmitting antenna; the depth is used in advance.
  • the learned BP training algorithm is used to obtain the first neural network model and the second neural network model; the first neural network model constructed is used to solve the high-dimensional sparse signal ⁇ from the low-dimensional target signal;
  • the signal ⁇ is input into the constructed second neural network model to reconstruct the original signal x.
  • FIG. 1 is a schematic block diagram of the signal processing flow of the MIMO multi-antenna signal transmission and detection system based on deep learning provided by the present invention
  • FIG. 2 is a schematic block diagram of a preferred signal compression multiplexing and detection processing process provided by the present invention
  • Figure 3 is a bit error rate performance curve of the classic detection algorithm ZF and the multi-antenna signal transmission and detection technology of the MIMO system of the present invention under different transceiver antenna configurations;
  • Figure 4 is a bit error rate performance curve of the classic detection algorithm ZF and the multi-antenna signal transmission and detection technology of the MIMO system of the present invention under different transceiver antenna configurations;
  • FIG. 5 is a schematic diagram of the module connection of the signal transmission and detection system of the MIMO multi-antenna system based on artificial intelligence and compressed sensing technology provided by the present invention.
  • Device transmitter 101 Random number generator
  • Bit-level processing module 103 Modulation module
  • Compression multiplexing module 2 Device receiving end
  • Channel estimation module 202 The first neural network signal processing module
  • the technical solution adopted by the present invention is a signal transmission and detection technology and detection method of a MIMO multi-antenna system based on artificial intelligence and compressed sensing technology in a MIMO system.
  • the technical solution is as follows:
  • Signal processing at the transmitting end 1 of the device For a MIMO communication system equipped with N t transmitting antennas and N r receiving antennas.
  • the l signal x after channel coding and signal modulation at the transmitting end of the system is compressed into a ⁇ l signal by the compression multiplexing module 104, and then the compressed and multiplexed signal z is transmitted via the transmitting antenna.
  • A is the compressed dimension reduction matrix of N t rows and l columns
  • a Gaussian random matrix is selected as the compressed sensing multiplexing matrix/compressed dimensionality reduction matrix A.
  • the compression ratio ⁇ is determined by the size of the compressed dimensionality reduction matrix A in the compressed sensing technology.
  • the link at the receiving end 2 of the device is roughly the inverse process of the link at the transmitting end 1 of the device.
  • y is the received signal vector of Nr ⁇ 1, which is the l-channel modulation symbol transmitted by the N t transmitting antennas received by the receiving end 2 of the device
  • z is the transmitted signal vector of Nt ⁇ 1
  • n is the Gaussian white of Nr ⁇ 1 Noise vector
  • H is the channel propagation matrix of Nr ⁇ Nt, and it is a deterministic matrix that remains constant within a coherent time interval.
  • the receiving end 2 of the apparatus may estimate the channel propagation matrix H according to the pilot signal inserted in the transmission data.
  • the above-mentioned channel estimation that is, the device receiving end 2 determines the state (uncertainty) of the wireless transmission channel through data processing on the device transmitting end 1.
  • a commonly used method is based on the non-blind channel estimation of pilot symbols, that is, the device transmitter 1 sends known pilot information, and the device receiver 2 processes the information to obtain the channel state.
  • the input signal of the training set is y
  • the output signal is the sparse representation ⁇ of the original signal x on the over-complete redundant dictionary D.
  • the second neural network model the input signal of the training set is ⁇
  • the output signal is the original sending data stream That is, l modulation symbols transmitted by N t transmission antennas at the transmitting end of the device after compression and multiplexing.
  • Compressed sensing also known as compressed sampling or sparse sampling, is a method of finding sparse solutions for underdetermined linear systems. This method has been in existence for at least forty years. Thanks to the work of David Donoho, Emmanuel Candès, and Tao Zhexuan, this field has developed significantly recently. In recent years, compressed sensing technology has been widely used in fifth-generation mobile communication systems, and has received a lot of attention and research.
  • Compressed sensing originates from acquiring and reconstructing sparse or compressible signals. Candès and Donoho in the literature "Compressed sensing,” IEEE Transactions on Information Theory, vol.52, no.4, pp. 1289-1306, 2006 and “Compressive sampling,” In: Proceedings of International Congress of Mathematicians, Switzerland: European Mathematical Society Publishing House, pp.1433-1452, 2006 formally put forward the concept of compressed sensing, using the sparsity of the original signal, compared with the Nyquist theory, can restore the original high-dimensionality from fewer measured values Signal. The core idea is to combine compression and sampling.
  • the non-adaptive linear projection (measured value) of the signal is collected, and then the original signal is reconstructed from the measured value according to the corresponding reconstruction algorithm.
  • the traditional signal acquisition and processing process mainly includes four parts: sampling, compression, transmission and decompression.
  • the sampling process must satisfy Shannon's sampling theorem, that is, the sampling frequency cannot be lower than 2 times the highest frequency in the analog signal spectrum.
  • the compressed sensing theory is different from the traditional Nyquist sampling theorem. As long as the signal x is compressible or sparse in a certain transform domain D, then an observation matrix A that is not related to the transform domain D can be used to transform the high The one-dimensional sparse signal is projected onto a low-dimensional space, and then the original signal is reconstructed with high probability from these few projections by solving an optimization problem. Under this theoretical framework, the sampling rate is not determined by the bandwidth of the signal, but by the structure and content of the information in the signal. Compressed sensing theory mainly includes three aspects: signal sparse representation, coding sampling and reconstruction algorithm.
  • "Over-complete basis” means that the number of atoms in it greatly exceeds the dimensionality of the original signal. Since the ubiquitous signals in nature are generally not sparse, the sparse representation of the signal means that when the signal is projected into a certain transform domain D, only a few elements are non-zero, and the resulting transform vector is said to be sparse or nearly sparse.
  • x D ⁇
  • the projection matrix must meet the Restricted Isometry Property (RIP) condition, and then the linear projection measurement of the original signal is obtained by the product of the original signal and the measurement matrix.
  • the RIP condition is defined as follows: If there is a constant ⁇ K ⁇ (0,1] for all signals ⁇ with a sparsity of K, the matrix A satisfies the following formula
  • the matrix A satisfies the constraint equidistant property of order K, where the sparsity K refers to the number of non-zero elements of the signal ⁇ .
  • the reconstruction algorithm of compressed sensing can still completely recover the original signal from the measurement number much smaller than the signal dimension.
  • the compression ratio ⁇ determines the number of transmitting and receiving antennas that can be reduced and the device receiving end 2 reconstruction M.Davenport pointed out in the theorem 3.5 of his doctoral thesis "Random observation on random observations: Sparse signal acquisition and processing”: A is a matrix that satisfies the 2K-order RIP constant ⁇ 2K ⁇ (0,1], as long as C is a constant approximately equal to 0.28, and the original signal can be recovered. At the same time, Donoho in the document “Extensions of compressed sensing," Signal Processing, vol. 86, no. 3, pp.
  • the under-determined problem can be solved by the reconstruction algorithm in compressed sensing.
  • the solution of this problem requires exhaustive list of all possible permutations of non-zero values in the sparse vector ⁇ , so it is difficult to solve.
  • researchers have proposed a series of algorithms for obtaining sub-optimal solutions, mainly including: greedy pursuit algorithm, convex relaxation method, Bayesian algorithm, combination algorithm, etc.
  • Each algorithm has its advantages and disadvantages.
  • the convex relaxation method requires the least number of observations to reconstruct the signal, but it often has a heavy computational burden.
  • Greedy tracking algorithm is between these types of algorithms in terms of running time and sampling efficiency, and its anti-noise performance is unstable.
  • the appropriate reconstruction algorithm can be selected according to different environments. Once the sparse representation vector is obtained, the original signal can be recovered.
  • the traditional MIMO signal transmission process the transmitted data stream s is processed by space-time coding, digital-to-analog conversion and analog modules, and then separated into Nt sub-data streams, which are simultaneously transmitted through Nt transmitting antennas at the same frequency.
  • the transmitted signal propagates through the reflection and scattering of the wireless channel, and these parallel sub-signals arrive at the receiving end 2 of the device at different times through different paths, and are received by Nr antennas.
  • the receiving end 2 of the device uses signal processing technology to jointly process the signals received by each antenna, thereby recovering the original data stream.
  • the codeword is modulated and sent, and the receiving end 2 of the device performs signal detection and reconstruction of the original signal.
  • a compression multiplexing module 104 is added to the transmitting end 1 of the device in this application. Firstly, the modulated signal is compressed and dimensionally reduced. The selection of the compressed dimension-reduction matrix does not require channel state information. The measurement matrix in the compressed sensing technology can be selected as the signal compression matrix to complete the dimensionality-reduction and multiplexing processing of the transmitted signal. , Reduce the amount of data.
  • the receiving end 2 of the device of this application is divided into the following two steps to reconstruct the signal: (1) Train the first neural network model (NN1) through the BP (back propagation) algorithm in deep learning, and receive it from low-dimensional Solve the high-dimensional sparse signal ⁇ from the signal. (2) Train the second neural network model (Neural Network Model, NN2) through the BP algorithm of deep learning, and reconstruct the original signal x.
  • the backpropagation (BP) neural network is a concept proposed by scientists led by Rumelhart and McClelland in 1986. It is a multi-layer feedforward neural network trained according to the error backpropagation algorithm.
  • the BP algorithm takes the network error square as the objective function, and uses the gradient descent method to calculate the minimum value of the objective function.
  • the BP network adds several layers (one or more layers) of neurons between the input layer and the output layer. These neurons are called hidden units. They have no direct connection with the outside world, but changes in their state can affect the input. The relationship with the output, each layer can have several nodes.
  • the calculation process of BP neural network is composed of forward calculation process and reverse calculation process.
  • the input pattern is processed layer by layer from the input layer to the hidden unit layer, and then to the output layer.
  • the state of each layer of neurons only affects the state of the next layer of neurons. If the desired output cannot be obtained in the output layer, then switch to back propagation, return the error signal along the original connection path, and modify the weight of each neuron to minimize the error signal.
  • the aforementioned "backpropagation" is an algorithm used to effectively calculate gradients in a neural network, or more generally, a feedforward computational graph. It can be boiled down to applying the chain rule of differentiation from the network output, and then propagating the gradient backwards. The first application of backpropagation can be traced back to Vapnik et al.
  • BP network is mainly used in the following four aspects: 1) function approximation: train a network with input vector and corresponding output vector to approximate a function; 2) pattern recognition: use a pending output vector to associate it with the input vector; 3 ) Classification: classify the appropriate way defined by the input vector; 4) data compression: reduce the dimensionality of the output vector to facilitate transmission or storage.
  • this application uses the BP algorithm to train the neural network to achieve function approximation.
  • the present invention can be based on the existing related MIMO system signal multiplexing technology, and compared with the traditional MIMO scheme, by introducing the compression multiplexing module 104 and the demultiplexing module of the receiving end 2 of the device, the required antenna is reduced. Based on the number of data, the same amount of data is transmitted at the same time, which improves the multiplexing gain and capacity of the MIMO system. Compared with the existing MIMO spatial multiplexing technical solution, this application no longer only focuses on eliminating the interference of adjacent data, but more on how to ensure the detection performance of the receiving end 2 of the device under the condition of a given number of transmitting antennas. On the basis of MIMO, multiple data streams are multiplexed and transmitted to the receiving end 2 of the device to obtain multiplexing gain and transmission capacity that exceed the inherent MIMO system.
  • This embodiment integrates the MIMO multi-antenna signal transmission and detection technology based on deep learning and compressed sensing proposed in this application, and gives a detailed description of the specific implementation steps of the present invention with examples.
  • the information source is generated by using the random number generator 2 to generate a 0,1 bit sequence.
  • Modulation is to modulate bit data, including BPSK, QPSK, 16QAM and 64QAM.
  • This embodiment uses BPSK modulation as an example for description.
  • S1 Signal processing at the transmitter 1 of the device.
  • the random number generator 101 is used to generate a 0,1 bit sequence to form the original data.
  • S12 Generate signal x after BPSK modulation.
  • the sending data of each group can be different.
  • S13 Compression multiplexing processing, that is, the sparse vector transmission data is multiplied by the compression multiplexing matrix A to obtain the data vector z.
  • S2 Signal detection at the receiving end 2 of the device.
  • the signal matrix H is estimated.
  • the first neural network signal processing module 202 solves the high-dimensional sparse signal ⁇ from the low-dimensional target signal through the first neural network model constructed by the first neural network signal processing module 202.
  • the input signal of the first neural network model is y
  • the output signal is the sparse representation ⁇ of the original signal x on the over-complete redundant dictionary D.
  • the second neural network signal processing module reconstructs the original signal x by inputting the sparse signal ⁇ into the second neural network model constructed by it.
  • the input signal of the second neural network model is the sparse representation ⁇ of the original signal x on the over-complete redundant dictionary D, and the output signal is the original transmission data stream
  • the neural network model is composed of three parts: the input layer (layer1), the middle layer (layer2, .., L-1), and the output layer (layerL).
  • the input layer plays the role of signal transmission and is responsible for receiving
  • each unit of the input layer represents a feature
  • the middle layer can be a single middle layer or multiple middle layers, which play the role of internal information processing and are responsible for information transformation
  • the output layer plays the role of outputting information to the outside, and each of the output layers
  • the unit represents a category.
  • a BP neural network is used to simulate a mapping function, which can map the input space data to the output space; the BP neural network will try to fit an original device receiving end 2 signal y and original signal x.
  • the sparse representation on the co-dictionary D represents the function between ⁇ , and the sparse representation ⁇ and the original transmitted data stream The function between; based on the mapping function model generated by the trained BP neural network, the sparse representation ⁇ of the original signal x theoretically calculated on the over-complete redundancy dictionary D can be restored according to the signal y received by the device receiving end 2, and
  • the sparse representation ⁇ calculated according to the above theory reconstructs the original transmission data x.
  • a cost function is used to measure the difference between the output of the BP neural network and the real output, and the BP neural network is trained so that the output of the network input (the signal received by the device receiving end 2) through the neural network can be as close as possible to the theoretical output.
  • the gradient descent method is used in this application to solve the neural network parameters.
  • the optimal neural network weight is solved, the first neural network model or the second neural network model is established. Create a BP neural network, collect a large amount of sample data, and artificially label the correct classification results, and then use these labeled data to train the created neural network.
  • each layer in the neural network is constantly adjusting its weight and bias until it can accurately output the target value.
  • the weight parameter matrix between the layers of the neural network Represents, where the superscript of the weight parameter w represents the number of layers, and the number of nodes in each of the two adjacent layers of the subscript.
  • E.g Represents the weight of the line segment between the first node of Layer1 and the second node of Layer2 in the input layer. These weights determine the role of the model, and the goal of the neural network is to calculate the weights through samples.
  • E.g Represents the input value of the first node of Layer2 and brings it into the Logistic function to get the output
  • the output of the neural network for: Introduce a nonlinear operator: It can be derived:
  • Training the neural network model The process of training the neural network model is mainly divided into two steps, one is to calculate the cost function J( ⁇ ), and the other is to adjust the parameter ⁇ to make the cost function value J( ⁇ ) as small as possible.
  • the forward propagation algorithm is used as follows to calculate the output of the sample under the current neural network model for each sample, find the cost function, and then update the weight parameter according to the output. Define the cost function J( ⁇ ), where m is the number of samples. Since the neural network has K outputs, its cost function also calculates the cost of K outputs accordingly.
  • the calculation formula is:
  • the back propagation algorithm is used to adjust the parameter ⁇ as follows, so that the cost function value J( ⁇ ) is as small as possible.
  • the backpropagation algorithm updates each weight coefficient by finding the partial derivative of the cost function with respect to each weight coefficient. For example, first, calculate the gradient of the last layer: (1) calculate the gradient of the cost function value to the nonlinear operator, (2) calculate the gradient of the output of the neural network to the bias and the weight between adjacent layers.
  • Figure 3 shows the bit error rate performance of different transceiver antenna configurations of the MIMO system after using compressed sensing and neural network signal transmission and detection technologies.
  • a flat fading channel is assumed here.
  • 4 transmitting antennas can only transmit 4 data symbols at the same time.
  • the receiving end 2 of the device also uses 2 or 3 receiving antennas and uses the models 1 and 2 trained by the neural network to obtain the reconstructed transmission signal
  • the bit error rate performance of the solution of this application is shown in the figure (2 ⁇ 2)-4.
  • the first number in the brackets represents the number of transmitting antennas
  • the second number represents the number of receiving antennas
  • the last number represents the original data length.
  • the number of receiving antennas has increased, and the scheme is recorded as (3 ⁇ 3)-4.
  • the solution proposed in this application can ensure the bit error rate under the condition of high SNR, while reducing the number of required transmitting and receiving antennas.
  • the above-mentioned zero-forcing detection algorithm uses the filter matrix W ZF to multiply the received signal y to eliminate the interference between the transmitted signals, thereby estimating each transmitted symbol.
  • Figure 4 shows the bit error rate performance of different transceiver antenna configurations of the MIMO system after using compressed sensing and neural network signal transmission and detection technologies.
  • a flat fading channel is assumed here.
  • 20 transmitting antennas can only transmit 20 data symbols at the same time.
  • the receiving end 2 of the device also uses 10 receiving antennas.
  • the bit error rate performance of this scheme is shown in the figure (10 ⁇ 10)-20.
  • the first number in the brackets indicates the number of transmitting antennas, and the second number indicates the number of receiving antennas.
  • the last number indicates the length of the original data.
  • our proposed scheme can reduce the number of transceiver antennas required while ensuring the bit error rate under the condition of high SNR. It can be seen that the solution proposed in this application can reduce the number of transceiver antennas required while ensuring the bit error rate under the condition of high SNR.
  • the enhanced spatial multiplexing method proposed by the present invention can be combined with neural network technology on the basis of the existing MIMO system, and on the basis of reducing the number of antennas, the same amount of data can be transmitted and the required antennas can be reduced. Increase the multiplexing gain and system capacity.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Neurology (AREA)
  • Power Engineering (AREA)
  • Radio Transmission System (AREA)

Abstract

本发明涉及基于深度学习的MIMO多天线信号传输与检测技术，提出了基于深度学习的MIMO多天线信号传输与检测装置，该装置包括：压缩复用模块，用于对调制后的信号进行压缩降维处理；装置发射端，用于在收发天线数给定的情况下将经过压缩复用模块处理后的目标信号通过发射天线发射；装置接收端，用于对接收到的信号进行处理以实现其对目标信号的重构，装置接收端中包括第一神经网络信号处理模块和第二神经网络信号处理模块，其中，第二神经网络信号处理模块通过将由第一神经网络信号处理模块利用其所构建的第一神经网络模型来从低维的目标信号中所解出的高维的稀疏信号θ输入其所构建的第二神经网络模型来重构得到原始信号x。

Description

基于深度学习的MIMO多天线信号传输与检测技术 技术领域
本发明涉及移动通讯技术领域，尤其涉及基于深度学习的MIMO多天线信号传输与检测技术。
背景技术
MIMO(Multiple-Input Multiple-Output)技术指在装置发射端和装置接收端分别使用多个发射天线和接收天线，使信号通过装置发射端与装置接收端的多个天线传送和接收，从而改善通信质量。根据收发两端天线数量，相对于普通的SISO(Single-Input Single-Output)系统，MIMO包括SIMO(Single-Input Multiple-Output)系统和MISO(Multiple-Input Single-Output)系统。它能充分利用空间资源，通过多个天线实现多发多收，在不增加频谱资源和天线发射功率的情况下，可以成倍的提高系统信道容量，显示出明显的优势，MIMO系统已经广泛应用于无线通信中——移动设备和网络普遍都会使用多根天线来增强连接性、提升网络速度和用户体验。大规模MIMO是5G超高数据速率的关键因素，可以带来更大网络容量，更广信号覆盖和更好的用户体验，将5G的潜力发挥到一个全新的水平。
根据空时映射方法的不同，MIMO技术大致可以分为两类：空间分集和空间复用。空间分集是指利用多根发送天线将具有相同信息的信号通过不同的路径发送出去，同时在接收机端获得同一个数据符号的多个独立衰落的信号，从而获得分集提高的接收可靠性。举例来说，在慢瑞利衰落信道中，使用一根发射天线N r根接收天线，发送信号通过N r个不同的路径。如果各个天线之间的衰落是独立的，可以获得最大的分集增益为N r。对于发射分集技术来说，同样是利用多条路径的增益来提高系统的可靠性。在一个具有N t根发射天线N r根接收天线的系统中，如果天线对之间的路径增益是独立均匀分布的瑞利衰落，可以获得的最大分集增益为N tN r。无线通信的不可靠性主要是由无线衰落信道时变和多径特性引起的，如何在不增加功率和不牺牲带宽情况下，同时减少多径衰落对基站和移动台的影响就显得很重要。唯一方法是采用抗衰落技术，克服多径衰落的有效方法是各种分集技术。分集技术主要用来对抗信道衰落。相反，MIMO信道中的衰落特性可以提供额外的信息来增加通信中的自由度。从本质上来讲，如果每对发送接收天线之间的衰落是独立的，那么可以产生多个并行的子信道。如果在这些并行的子信道上传输不同的信息流，可以提供传输数据速率，这被称为空间复用。但是，在高SNR(Signal to Noise Ratio，信噪比)的情况下，传输速率是自由度受限的。
MIMO凭借其两大优势：(1)提高信道的容量。MIMO接入点到MIMO客户端之间，可以同时发送和接收多个空间流，信道容量可以随着天线数量的增大而线性增大，因此可以利用MIMO信道成倍地提高无线信道容量，在不增加带宽和天线发送功率的情况下，频谱利用率可以成倍地提高；(2)提高信道的可靠性。利用MIMO信道提供的空间复用增益及空间分集增益，可以利用多天线来抑制信道衰落。多天线系统的应用，使得并行数据流可以同时传送，可以显著克服信道的衰落，降低误码率，已经成为一项运用于802.11n的核心技术。802.11n是IEEE继802.11a/b/g后全新的无线局域网技 术，速度可达600Mbps。同时，MIMO技术可改进已有802.11a/b/g网络的性能。
随着使用天线数目的增加，MIMO技术实现的复杂度大幅度增高，从而限制了天线的使用数目，不能充分发挥MIMO技术的优势。目前，如何在保证一定的系统性能的基础上降低MIMO技术的算法复杂度和实现复杂度，成为业界面对的巨大挑战。
人工智能指由人制造出来的机器所表现出来的智能。通常人工智能是指通过普通计算机程序模拟人的某些思维过程和智能行为(如学习、推理、思考、规划等)，来呈现人类智能的技术，主要包括计算机实现智能的原理、制造类似于人脑智能的计算机，使计算机能实现更高层次的应用。我们坚信，未来数年，人工智能依然会“长驱直入”，为传统产业创造价值，并深刻改变我们的日常生活，如机器人领域、语言识别领域、图像识别领域、专家系统等。机器学习是人工智能的一个子集。机器学习目前主要解决分类问题、聚类问题、回归问题等，已广泛应用于字符识别、机器翻译、语音识别、搜索引擎、人脸识别、无人驾驶等领域。所有机器学习算法中最关键的是深度学习。
深度学习的概念源于人工神经网络的研究。深度学习的概念最早由机器学习领域的泰斗——多伦多大学GE.Hinton教授提出，其两个核心观点是：(1)含有多隐层的人工神经网络有着很优秀的特征学习能力，对学习到的特征数据有着更深刻的展示，可以更好的分类或可视化最终得到的网络数据；(2)深度神经网络可以通过“逐层初始化”来克服训练其本身网络参数上的难度，而逐层初始化可以通过无监督学习来实现。含多隐层的多层感知器就是一种深度学习结构。深度学习通过组合低层特征形成更加抽象的高层表示属性类别或特征，以发现数据的分布式特征表示。任何神经网络都可以有任何数量的层，输入或输出。输入神经元和最后一层输出神经元之间的层是深层神经网络的隐藏层。
围绕应用深度学习解决MIMO的问题，国内外也有不少发明专利成果。
现有技术如授权公告日为20180619的授权公告日为CN105610477B的专利文献提出的一种基于压缩感知的多发多收系统增强信号复用方法，在现有的MIMO技术的基础上，在现有的相关MIMO系统信号复用技术的基础上，选择压缩感知技术中的随机测量矩阵作为信号压缩复用矩阵，然后充分利用发送信号在过完备的冗余字典上呈现的稀疏特性，通过压缩感知重构算法，从低维的接收复用信号中解出高维的发送信号，从而大幅度提高给定MIMO系统收发天线数条件下的信号复用增益，更好地满足MIMO系统对宽带传输的应用要求，且具有确保接收端运用压缩感知领域成熟的优化重构算法高概率地重构出发送端经过压缩复用步骤发送的多路数据流、对现有MIMO系统修改小的优点。
中国申请号201510473741.6专利(自复数神经网络信道预测方法，西安电子科技大学)公开了一种复数神经网络信道预测方法，主要解决MIMO系统中由于信道时变而导致的信道衰落问题。其技术方案是：1.基站对信道进行测量得到含有估计误差的信道系数训练序列；2.根据得到的信道系数序列得到相应的训练样本和期望输出；3.输入训练样本进行复数小波神经网络训练，得到最终的网络权值；4.基站利用训练后的复数小波神经网络进行信道系数预测。本方法简便易行，效果良好，适用于减少由信道时变对MIMO系统信道的影响。
中国申请号201810177829.7专利(一种基于神经网络的无线信道建模方法，东南大学)公开了一种基于神经网络的无线信道建模方法。本发明首先对用户反馈的接收信号进行处理，得到估计的信道参数；然后根据二维图像得到散射体的三维地理信息，并对它们进行聚类，最后将信道参数、地理信息作为神经网络的输入，接收信号作为输出，训练得到非线性时变的神经网络模型。该方法在可接受的复杂度内得到更加准确的信道 模型，能够满足未来5G通信系统中采用的大规模MIMO技术、频带扩展、高移动性等场景的信道建模需求。
中国申请号201810267976.3专利(基于BP算法构建的深度神经网络大规模MIMO系统检测方法，东南大学)提供了一种基于BP算法构建的深度神经网络大规模MIMO系统检测方法，该方法通过将置信传播迭代算法因子图展开并映射到神经网络结构上来构建用于大规模MIMO系统检测的深度神经网络；深度神经网络的神经元对应于迭代算法因子图中的结点，各层神经元个数均等于迭代算法因子图中符号结点的个数；隐藏层之间的映射函数为迭代算法中置信信息的更新公式，隐藏层的数量等于迭代算法的迭代次数。具体地，本发明还提供了基于阻尼置信传播和最大和置信传播这两种信息传播迭代算法分别构建两种深度神经网络的MIMO检测方法。本发明在不增加在线运算复杂度的情况下，达到更低的误码率，并且对各种信道状况和天线配比都具有稳健性。
中国申请号201910063733.2专利(基于深度学习的优化MIMO检测方法，上海大学)通过构建MIMO端到端传输模型，根据MIMO装置接收端收到的信号y(t)和估计的不完美信道状态信息得到模型的复时域向量作为深度神经网络(DNN)的输入，利用DNN得到装置发射端比特流的估计值与现有技术根据做硬判决得到发送比特流的估计值相比，本发明能够在不完美的信道信息下，提高准确度和检测速率，保证在低复杂度算法下实现低误比特率的检测性能，同时在含有固有信道误差的情况下具有很好的鲁棒性。
中国申请号201610327115.0专利(一种大规模MIMO下基于深度学习的码本选择方法，重庆邮电大学)涉及一种大规模MIMO下基于深度学习的码本选择方法。该方法包括：采集测试区的导频信息构建导频训练序列，进而得到导频训练样本；对导频训练样本进行神经网络迭代学习，得到最终的网络权重值；根据学习后的神经网络输出的信道，从完备码本中选出最优码字。之后将未知区与测试区进行信道信息匹配，得到其无线信道，进而得到与无线信道对应的码字。本发明能有效、准确、快速地建立无线信道模型与码本查询，避免了未知区的信道估计且大大降低了未知区信道选择码本的复杂度。
中国申请号201811626005.X专利(基于改进梯度投影法的低复杂度MIMO-NOMA系统信号检测方法，重庆邮电大学)公开一种基于改进梯度投影法的低复杂度MIMO NOMA系统信号检测方法，涉及无线通信技术。根据系统活跃用户的稀疏特性，利用凸优化算法思想，将系统模型转化为严格的二次规划问题；然后对该问题进行迭代求解，并对每次迭代结果进行预处理操作，达到对活跃用户及其信号有效的检测。本发明突破了传统检测方法中算法收敛速度慢的问题，对每次迭代结果进行预处理操作，不仅可使检测结果快速收敛，而且还能检测出活跃用户集合，其实现过程简单，应用范围广泛。
中国申请号201910014714.0专利(一种基于深度学习的MIMO系统波束成形矩阵设计方法，南京邮电大学)公开了一种基于深度学习的MIMO系统波束成形矩阵设计方法，步骤如下，首先利用已知算法获得深度学习网络所需要的训练样本集；接着构建深度学习神经网络模型，初始化模型相关参数并利用训练样本集进行训练；然后利用导频获取信道送入神经网络预测出波束成形矩阵系数，最后将信道与波束成形矩阵系数组合，构成波束成形矩阵。该方法利用深度学习神经网络获得的波束成形矩阵能够同时兼顾性能与算法复杂度，可以在保证性能的前提下降低时延，使得MIMO系统能提供实时服务。
中国申请号201810182937.3专利(一种基于机器学习的MIMO链路自适应传输方法，东南大学)公开了一种基于机器学习的MIMO链路自适应传输方法，使用非监督学习自编码算法进行特征的提取与降维，引入深度学习的思想，能够在保留主要信息状 态信息的前提下降低特征维度与计算复杂度。本发明利用逻辑回归算法构建信道状态信息与传输参数间的映射关系，不同于以往固定的参数化模型，能够基于样本数据进行训练，在数据集质量较优，覆盖所有状态的情况下，能够更好的建立信道状态信息与传输参数间的映射关系，并相较于传统的单一等效信噪比，能够更充分的利用信道状态信息。此外，本发明还基于信道矩阵进行CQI选择，通过信道矩阵和噪声方差研究基于机器学习的MIMO链路自适应方法不受接收机设计的约束，具有普适性。
中国申请号201710495044.X专利(一种基于深度学习的MIMO系统联合预编码和天线选择方法，浙江理工大学)公开了一种基于深度学习的MIMO系统联合预编码和天线选择方法，包括以下步骤：首先通过已有的天线选择方法产生深度学习所需要的训练数据集；接着，建立深度学习模型，利用训练数据训练深度学习模型并保存；然后利用保存的深度学习模型完成天线选择；最后对所选择的MIMO子系统进行最优预编码设计。本发明利用深度学习技术设计MIMO系统联合预编码和天线选择，能够在获得良好系统信噪比的情况下实现较低的计算复杂度。
中国申请号201910242525.9号发明专利(一种面向高铁的深度信号检测方法，深圳大学)提出了一种面向高铁的深度信号检测方法，首先，收集数据，依据高铁沿线不同环境类型，在高铁沿线各个场景内收集若干发送信号和接收信号；其次，划分场景，通过数据分析将每个场景进一步划分为多个区域以满足神经网络的兼容性；再次，建立深度高铁信号检测神经网络模型；接着，离线训练高铁信号检测神经网络；最后，进行在线实时检测信号，高铁在行驶过程中先通过GPS确定其位置信息，判断其所在的区域，并选择相应的神经网络模型，接着将实时接收到信号输入训练好的神经网络中，实时输出基站端发送信号。本发明的系统性能有了很大的提高，信号检测误比特率降低，算法更具鲁棒性。本发明所使用的方法无需对信道进行估计，节省了导频开销。
中国申请号201810279530.2号发明专利(基于BP神经网络的可见光通信MIMO抗扰降噪方法，中南民族大学)公开了一种基于BP神经网络的可见光通信MIMO抗扰降噪方法，涉及可见光通信领域的MIMO天线技术。本系统包括依次连通的系统装置发射端、系统装置接收端信号处理部分和BP神经网络信号处理部分。本方法是：1)电信号加载到LED阵列上，以光信号的形式发射出去；2)装置接收端光电探测器将光信号转为电信号；3)多路电信号经低通滤波器去除高频干扰；4)BP神经网络经过训练后对多路信号进行降噪以及消除干扰的处理，最后通过并串转换成为二进制串行数据流。本发明改进了现有MIMO技术的传输性能；将神经网络与可见光MIMO技术结合，发挥神经网络在无线通信中去扰降噪方面的优势；采用神经网络接收处理技术使整个VLC系统更加稳定。
中国申请号201710213235.2发明专利(基于正交映射与概率神经网络的可见光信道联合均衡方法，中山大学)公开一种基于正交映射与概率神经网络的可见光信道联合均衡方法，包括装置发射端和装置接收端，信号通过可见光MIMO信道从装置发射端传送到装置接收端；所述的可见光MINO信道为多输入多输出信道；所述的联合均衡为前置均衡与后置均衡相结合；本发明采用的是前置均衡技术与后置均衡技术相结合的联合均衡方案，即基于正交映射与概率神经网络的可见光多输入多输出信道联合均衡方法，能有效抑制可见光MIMO通信系统信道之间的干扰，提高数据传输可靠性。
中国201910125325.5号发明专利(一种基于深度学习的MIMO解码方法、装置及存储介质，深圳市宝链人工智能科技有限公司)实施例公开了一种基于深度学习的MIMO解码方法、装置及存储介质，通过构建MIMO解码的训练数据集合，训练数据集合中包括多个训练数据；然后基于训练数据集合对神经网络进行训练，得到训练完成的神 经网络模型；在接收到待解码MIMO信号时，将待解码MIMO信号输入至神经网络模型进行MIMO解码，然后得到神经网络模型所输出的MIMO解码结果。通过本发明的实施，基于深度学习来设计用于联合MIMO检测和信道解码的神经网络模型，将MIMO检测和信道解码视为联合解码过程，并且通过训练来改善神经网络模型所输出结果的近似性，保证了MIMO解码的整体性能，具有更高的解码准确性以及更快的解码速度。
中国申请号201810757547.4专利(一种机器学习辅助的大规模MIMO下行用户调度方法，东南大学)公开了一种机器学习辅助的大规模MIMO下行用户调度方法，包括以下步骤：S1：基站通过用户发送的上行探测信号获取特征方向上的特征模式能量耦合矩阵；S2：基站利用特征模式能量耦合矩阵，通过机器学习的方法辅助进行各种用户和波束组合下的和速率计算；S3：采用贪婪算法实现和速率最大准则的用户调度，获取最优用户波束配对组合。本发明通过上行探测信号获取统计信道信息，采用和速率最大化准则进行用户调度。在基站仅有统计信道信息的情况下，通过有针对性的特征提取以及神经网络的设计，精确地实现和速率的近似计算，极大地降低大规模天线下用户调度的复杂度，并且性能接近最优，具有较好的适用性和鲁棒性。
中国201610353881.4号发明专利(一种基于机器学习算法的MIMO相关信道下的调制识别方法，北京邮电大学)是一种基于机器学习算法的MIMO相关信道下的调制识别方法，属于通信领域；具体步骤为：首先，将通信装置发射端的每个数据流分别采用空时编码，每个码字分别通过Nt根发射天线发射出去；然后，根据装置接收端的相关矩阵和装置发射端的相关矩阵计算MIMO信道矩阵H；根据MIMO信道矩阵H，计算每个接收天线上的接收信号并进行修正；最后，每根接收天线分别对修正后的信号进行特征提取，针对提取的特征值进行训练测试，计算出该样本属于的调制识别模式；优点在于：对非高斯信道的鲁棒性和泛化能力较强，通过参数迭代可实现更加复杂环境下的调制体制识别；通过提取高阶矩和高阶累积量的特征，在较高信噪比下，信号特征差异明显，便于机器学习算法的分类。
MIMO技术的应用，使空间成为一种可以用于提高性能的资源，并能够增加无线系统的覆盖范围。MIMO技术已经成为无线通信领域的关键技术之一，通过近几年的持续发展，MIMO技术已经越来越多地应用于各种无线通信系统。随着使用天线数目的增加，MIMO技术实现的复杂度大幅度增高，从而限制了天线的使用数目，不能充分发挥MIMO技术的优势。同时，从1956年正式提出人工智能学科算起，50多年来，取得长足的发展，成为一门广泛的交叉和前沿科学。作为人工智能的一个重要分支，从上面的描述可以看出，神经网络通过计算机来实现数学上的网络，是公认解决目前通信遇到的部分瓶颈问题的“出路”。
综合上述目前国内外围绕神经网络和通信问题关键技术的研究，目前已有不少的研究成果，这些成果已经从MIMO的信道估计、检测信号等角度提出了可行的解决方案。但与此同时，如何在保证一定的系统性能的基础上提高系统容量、降低MIMO技术的算法复杂度和实现复杂度，成为业界面对的巨大挑战。现有围绕MIMO条件下的相关研究还没有从给定发送天线和接收天线数的条件下，结合神经网络技术提高系统复用增益这个角度来研究信号传输与检测的方案。
此外，一方面由于对本领域技术人员的理解存在差异；另一方面由于发明人做出本发明时研究了大量文献和专利，但篇幅所限并未详细罗列所有的细节与内容，然而这绝非本发明不具备这些现有技术的特征，相反本发明已经具备现有技术的所有特征，而且申请人保留在背景技术中增加相关现有技术之权利。
发明内容
针对如何在保证一定的系统性能的基础上提高系统容量、降低MIMO技术的算法复杂度和实现复杂度的问题，现有围绕MIMO条件下的相关研究还没有从给定发送天线和接收天线数的条件下，结合神经网络技术提高系统复用增益这个角度来研究信号传输与检测的方案。本申请结合压缩感知技术的最新研究进展和神经网络技术，提出了一种基于深度学习和压缩感知的MIMO系统中信号的复用传输与检测方案，一方面，与传统的MIMO方案相比，本申请通过引入压缩复用模块、以及装置接收端的解复用模块，在减少所需天线数的基础上，同时传输相同的数据量，提高了MIMO系统的复用增益和容量；另一方面，与现有的仅仅关注于消除相邻数据的干扰的MIMO空间复用技术方案相比，本申请不仅关注于消除相邻数据的干扰，而更关注于如何在给定发送天线数的条件下，在保证装置接收端检测性能的基础上，将更多的数据流复用传输到装置接收端，获得超过MIMO系统固有的复用增益和传输容量。
本申请所提出的基于深度学习和压缩感知的MIMO装置主要由装置发射端1的压缩复用模块104、以及装置接收端2的第一神经网络信号处理模块202和第二神经网络信号处理模块203所构成。此外优选地，装置发射端1中还设置有随机数发生器101(或称原始信息比特生成模块)、比特级处理模块102、调制模块103，装置接收端2中还设置有信道估计模块201。
优选地，如图5所示，装置发射端1：由随机数发生器101产生的原始数据先后经过比特级处理模块102和调制模块103，生成调制后的信号，经调制后的信号继而经过压缩复用模块104，压缩复用模块104对发送信号进行压缩降维和复用处理，经压缩复用后的信号经发射天线发射；装置接收端2：由信道估计模块201对接收到的信号进行信道估计，基于接收到的信号和估计的信道状态信息得到第一神经网络信号处理模块202的输入，基于第一神经网络信号处理模块202的输出得到第二神经网络信号处理模块203的输入，第二神经网络模型203重构输出原始的发送数据流x。
上述过程具体包括以下内容：
基于深度学习的MIMO多天线信号传输与检测装置，该装置包括：压缩复用模块，用于对调制后的信号进行压缩降维处理；装置发射端，用于在收发天线数给定的情况下将经过所述压缩复用模块处理后的目标信号通过发射天线发射；装置接收端，用于对接收到的信号进行处理以实现其对目标信号的重构，所述装置接收端中包括第一神经网络信号处理模块和第二神经网络信号处理模块，其中，第二神经网络信号处理模块通过将由第一神经网络信号处理模块利用其所构建的第一神经网络模型来从低维的目标信号中所解出的高维的稀疏信号θ输入其所构建的第二神经网络模型来重构得到原始信号x。
根据一种优选实施方式，所述装置接收端还包括信道估计模块，其被配置为根据所述装置接收端收到的经过压缩降维处理的低维目标信号进行信道估计并将得到的信道参数矩阵作为第一神经网络模型的输入。
根据一种优选实施方式，所述第一神经网络信号处理模块通过深度学习的BP算法创建神经网络，以所述装置接收端的接收信号向量y和基于所述装置发射端的发送信号向量x所确定的稀疏表示θ作为样本，构造第一组训练样本集，训练所述神经网络，得到所述第一神经网络模型。
根据一种优选实施方式，通过将所述发送信号向量x的所有可能组合分别作为冗余字典的不同列向量的方式组成冗余字典D，实现所述发送信号向量的稀疏表示θ。
根据一种优选实施方式，所述第二神经网络信号处理模块通过深度学习的BP算法创建神经网络，以所述装置发射端的发送信号向量x和所述基于所述装置发射端的发送信号 向量x所确定的稀疏表示θ作为样本，构造第二组训练样本集，训练所述神经网络，得到所述第二神经网络模型。
根据一种优选实施方式，所述装置发射端采用随机数发生器产生一组随机的0、1二进制比特序列，构成原始数据；原始数据经过BPSK调制，产生调制后的信号x。
根据一种优选实施方式，经调制后的l路信号x经过所述压缩复用模块，压缩成ρ1路信号，然后所述装置将压缩复用后的信号z经发射天线发射。
根据一种优选实施方式，压缩复用后的ρl路信号z是通过计算式z＝Ax来得到的，其中，A是N t行l列的压缩降维矩阵，ρ＝N t/l∈(0,1]代表压缩比例。
基于深度学习的MIMO多天线信号传输与检测系统，所述系统通过构建MIMO端到端传输模型，根据装置接收端收到的经过压缩降维处理的目标信号和估计的信道状态信息得到神经网络信号处理的输入，利用神经网络信号处理重构出原始信号。
基于深度学习的MIMO多天线信号传输与检测方法，所述方法包括至少一个以下步骤：收发天线数给定的情况下将经过所述压缩复用处理后的目标信号通过发射天线发射；提前利用深度学习的BP训练算法，得到第一神经网络模型和第二神经网络模型；通过所构建的第一神经网络模型来从低维的目标信号中解出高维的稀疏信号θ；通过将所述稀疏信号θ输入所构建的第二神经网络模型来重构得到原始信号x。
附图说明
图1是本发明提供的基于深度学习的MIMO多天线信号传输与检测系统的信号处理流程的示意框图；
图2是本发明提供的优选的信号压缩复用与检测处理过程示意框图；
图3是不同收发天线配置下经典检测算法ZF和本发明MIMO系统多天线信号传输与检测技术的误码率性能曲线；
图4是不同收发天线配置下经典检测算法ZF和本发明MIMO系统多天线信号传输与检测技术的误码率性能曲线；和
图5是本发明提供的基于人工智能和压缩感知技术的MIMO多天线系统的信号传输与检测系统的模块连接示意图。
附图标记列表
1：装置发射端                 101：随机数发生器
102：比特级处理模块           103：调制模块
104：压缩复用模块             2：装置接收端
201：信道估计模块             202：第一神经网络信号处理模块
203：第二神经网络信号处理模块
具体实施方式
下面结合附图对本发明进行详细说明。
如图2所示，本发明采用的技术方案是一种MIMO系统中基于人工智能和压缩感知技术的MIMO多天线系统的信号传输与检测技术和检测方法，其技术方案如下所述：
1.装置发射端1信号处理：对于一个配有N t根发送天线和N r根接收天线的MIMO通信系统。系统发射端经过信道编码、信号调制后的l路信号x经过压缩复用模块104，压缩成ρl路信号，然后将压缩复用后的信号z经发射天线发射。其中压缩复用模块104对 输入信号的压缩处理可以表示为：z＝Ax。其中，x＝[x 1,x 2,…,x l] T代表经过编码调制后的l路调制符号，A是N t行l列的压缩降维矩阵，ρ＝N t/l∈(0,1]代表压缩比例。优选地，本实施例中选择高斯随机矩阵作为压缩感知复用矩阵/压缩降维矩阵A。压缩比例ρ由压缩感知技术中的压缩降维矩阵A的尺寸决定。
2.装置接收端2的信号检测：装置接收端2链路大致是装置发射端1链路的逆过程，装置接收端2接收到的信号为：y＝Hz+n＝HAx+n。其中，y是Nr×1的接收信号向量，即为装置接收端2收到N t根发射天线发射的l路调制符号；z是Nt×1的发送信号向量；n是Nr×1的高斯白噪声向量；H是Nr×Nt的信道传播矩阵，并且是一个确定性的、在一个相干时间间隔内都保持不变的矩阵。装置接收端2可以是根据发送数据中所插入的导频信号估计出信道传播矩阵H。
上述信道估计，即装置接收端2通过对装置发射端1的数据处理，来确定无线传输信道的状态(不确定性)。常用的方法是基于导频符号的非盲信道估计，即装置发射端1发送已知的导频信息，装置接收端2对该信息进行处理，得出信道状态。
3.根据训练得到的第一神经网络模型和第二神经网络模型重构出原始的发送数据流
Figure PCTCN2020083563-appb-000001
其中，第一神经网络模型，训练集的输入信号为y，输出信号是原始信号x在过完备冗余字典D上的稀疏表示θ。第二神经网络模型，训练集的输入信号为θ，输出信号为原始的发送数据流
Figure PCTCN2020083563-appb-000002
即装置发送端N t根发送天线发送的经压缩复用后的l路调制符号。θ为经过编码后的l路调制符号x＝[x 1,x 2,…,x l] T在过完备冗余字典D上的稀疏表示。
如下对上述过程的具体实现步骤进行说明：
针对“压缩复用矩阵的构造”：压缩感知(Compressed sensing)，也被称为压缩采样或稀疏采样，是一种寻找欠定线性系统的稀疏解的方法。这一方法至少已经存在了四十年，由于David Donoho、Emmanuel Candès和陶哲轩的工作，最近这个领域有了长足的发展。近几年，压缩感知技术被大量应用在第五代移动通信系统中，获得了大量的关注以及研究。
压缩感知起源于获取和重构稀疏或可压缩的信号。Candès和Donoho在文献”Compressed sensing,”IEEE Transactions on Information Theory,vol.52,no.4,pp.1289-1306,2006和”Compressive sampling,”In:Proceedings of International Congress of Mathematicians,Switzerland:European Mathematical Society Publishing House,pp.1433-1452,2006中正式提出了压缩感知的概念，利用原始信号的稀疏性，相较于奈奎斯特理论，得以从较少的测量值还原出原来整个的高维信号。其核心思想是将压缩与采样合并进行，首先采集信号的非自适应线性投影(测量值)，然后根据相应重构算法由测量值重构原始信号。传统的信号获取和处理过程主要包括采样、压缩、传输和解压缩四个部分。其采样过程必须满足香农采样定理，即采样频率不能低于模拟信号频谱中最高频率的2倍。
压缩感知理论与传统奈奎斯特采样定理不同，只要信号x是可压缩的或在某个变换域D是稀疏的，那么就可以用一个与变换域D不相关的观测矩阵A将变换所得高维的稀疏信号投影到一个低维空间上，然后通过求解一个优化问题从这些少量的投影中以高概率重构出原信号。在该理论框架下，采样速率不决定于信号的带宽，而决定于信息在信号中的结构和内容。压缩感知理论主要包括信号的稀疏表示、编码采样和重构算法三个方面。
信号的稀疏表示指的是将原始信号表示为在适当选取的一组过完备基(字典D＝[d 1,d 2…d p]，或称变换域)上的稀疏线性组合，其中d 1,d 2…d p为字典中的原子。“过完备基” 的意思是其中的原子数大大的超过原始信号的维数。由于自然界中普遍存在的信号一般都不是稀疏的，信号的稀疏表示就是将信号投影到某个变换域D时，只有少数元素是非零的，则称所得到的变换向量是稀疏或者近似稀疏的，即x＝Dθ，θ是原始信号x的一种简洁表达，这是压缩感知的先验条件，即信号必须在某种变换下可以稀疏表示。从理论上来说，总是可以找到一个变换域D，实现信号的稀疏表示。如果原始信号x本身就是稀疏的，则x＝θ。从冗余字典中找到具有最佳线性组合的多项原子来表示一个信号，称作信号的稀疏逼近或高度非线性逼近。
接下来在压缩感知理论中，需要设计压缩采样系统的观测矩阵A，如何采样得到少量的观测值，并保证从中能重构出原始的信号。显然，如果观测过程破坏了原始信号中的信息，重构质量是不可能得到保证的。为了确保信号的线性投影能够保持信号的原始结构，投影矩阵必须满足约束等距性(Restricted Isometry Property，RIP)条件，然后通过原始信号与测量矩阵的乘积获得原始信号的线性投影测量。RIP条件定义如下：如果存在常数δ K∈(0,1]对所有稀疏度为K的信号θ，矩阵A满足下式
Figure PCTCN2020083563-appb-000003
则称矩阵A满足阶数为K的约束等距性质，其中稀疏度K是指信号θ的非零元素的个数。A是N t行l列的压缩降维矩阵。压缩感知技术的优势在于即使N t＞l，(l是指信号的长度)，依然可以从N r(N r＝N t)次测量值中恢复出长度为l的原始数据。令ρ＝N t/l∈(0,1]代表压缩比例。根据压缩感知原理，只要测量矩阵A满足RIP条件，即使A是行数远远小于列数的矩阵，将信号θ投影到了一个维度减少的空间上，依然可以通过压缩感知的重构算法从远小于信号维度的测量数中完整地恢复出原始信号。压缩比例ρ决定了能够减少的发送和接收天线数和装置接收端2重构的性能。M.Davenport在其博士论文“Random observation on random observations:Sparse signal acquisition and processing”的定理3.5中指出：A是满足2K阶RIP常数δ 2K∈(0,1]的矩阵，只要
Figure PCTCN2020083563-appb-000004
C是约等于0.28的常数，则可恢复出原始信号。同时，Donoho在文献“Extensions of compressed sensing,”Signal Processing,vol.86,no.3,pp.533-548,2006中给出了观测矩阵所必需具备的三个条件，并指出大部分一致分布的随机矩阵都具备这三个条件,均可作为观测矩阵,如:部分Fourier集、部分Hadamard集、一致分布的随机投影(uniform Random Projection)集等。文献“ecoding by linear programming,”IEEE Transactions on Information Theory,vol.51,no.12,pp.4201-4215,2005和”Stable signal recovery from incomplete and inaccurate measurements,”Communications on Pure and Applied Mathematics,vol.59,no.8,pp.1207-1223,2006证明当测量矩阵A是高斯随机矩阵时，A能以较大概率满足RIP条件。所以本申请中选择高斯随机矩阵作为压缩感知复用矩阵A。
针对“装置接收端2的信号重构”：欠定问题可以通过压缩感知中的重构算法解决。该问题的求解需要穷举稀疏向量θ中非零值的所有排列可能,因而难于求解。鉴于此，研究人员提出了一系列求得次最优解的算法,主要包括：贪婪追踪算法、凸松弛法、贝叶斯 算法、组合算法等。每种算法都有其优缺点。凸松弛法重构信号所需的观测次数最少，但往往计算负担很重。贪婪追踪算法在运行时间和采样效率上都位于这几类算法之间，并且抗噪性能不稳定。可以根据不同的环境选择合适的重构算法，一旦得到稀疏表示向量，就可以恢复出原始的信号。
传统的MIMO信号传输流程：发射数据流s经过空时编码、数模转换和模拟模块处理，被分离为Nt路子数据流，以相同的频率分别经过Nt根发射天线同时发射出去。发射的信号经过无线信道的反射、散射等传播，这些并行子信号经过不同的路径在不同的时刻到达装置接收端2，由Nr根天线接收。装置接收端2采用信号处理技术，对各个天线接收到的信号进行联合处理，从而恢复出原始数据流。码字经过调制后发送，装置接收端2进行信号检测重构原始信号。
与上述传统MIMO信号传输流程相比，我们所提出的是基于深度学习和压缩感知技术相结合的信号传输与检测方案。如图1所示，本申请在装置发射端1增加了压缩复用模块104。首先将调制后的信号进行压缩降维处理，压缩降维矩阵的选择不需要信道状态信息，选择压缩感知技术中的测量矩阵即可作为信号压缩矩阵，完成对发送信号的压缩降维和复用处理，降低数据量。本申请装置接收端2分为以下两步来重构信号：(1)通过深度学习中的BP(back propagation)算法训练出第一神经网络模型(Neural Network Model，NN1)，从低维的接收信号中解出高维的稀疏信号θ。(2)通过深度学习的BP算法训练出第二神经网络模型(Neural Network Model，NN2)，重构出原始信号x。
其中，反向传播(back propagation，BP)神经网络是1986年由Rumelhart和McCl elland为首的科学家提出的概念，是一种按照误差逆向传播算法训练的多层前馈神经网络。BP算法以网络误差平方为目标函数、采用梯度下降法来计算目标函数的最小值。BP网络是在输入层与输出层之间增加若干层(一层或多层)神经元，这些神经元称为隐单元，它们与外界没有直接的联系，但其状态的改变，则能影响输入与输出之间的关系，每一层可以有若干个节点。BP神经网络的计算过程由正向计算过程和反向计算过程组成。正向传播过程，输入模式从输入层经隐单元层逐层处理，并转向输出层，每一层神经元的状态只影响下一层神经元的状态。如果在输出层不能得到期望的输出，则转入反向传播，将误差信号沿原来的连接通路返回，通过修改各神经元的权值，使得误差信号最小。上述“反向传播”是一种在神经网络中用来有效地计算梯度的算法，或更一般而言，是一种前馈计算图(feedforward computational graph)。其可以归结成从网络输出开始应用分化的链式法则，然后向后传播梯度。反向传播的第一个应用可以追溯到1960年代的Vapnik等人，但论文Learning representations by back-propagating errors常常被作为引用源。目前，在人工神经网络的实际应用中，绝大部分的神经网络模型都采用BP网络及其变化形式。它也是前向网络的核心部分，体现了人工神经网络的精华。
BP网络主要用于以下四个方面：1)函数逼近：用输入向量和相应的输出向量训练一个网络逼近一个函数；2)模式识别：用一个待定的输出向量将它与输入向量联系起来；3)分类：把输入向量所定义的合适方式进行分类；4)数据压缩：减少输出向量维数以便于传输或存储。这里，本申请采用BP算法训练神经网络，实现函数的逼近。
本发明可以在现有的相关MIMO系统信号复用技术的基础上，和传统的MIMO方案相比，通过引入压缩复用模块104、以及装置接收端2的解复用模块，在减少所需天线数的基础上，同时传输相同的数据量，提高MIMO系统的复用增益和容量。与现有的MIMO空间复用技术方案相比，本申请不再仅仅关注于消除相邻数据的干扰，而更关注于如何在给定发送天线数的条件下，在保证装置接收端2检测性能的基础上，将更多的数据流复用传输到装置接收端2，获得超过MIMO系统固有的复用增益和传输容量。
实施例
本实施例融合本申请提出的基于深度学习和压缩感知的MIMO多天线信号传输与检测技术，对本发明的具体实施步骤进行举例详细说明。
首先，信息源的产生是采用随机数发生器2产生0,1比特序列。
调制是对比特数据进行调制，包括BPSK、QPSK、16QAM和64QAM等。
本实施例采用BPSK调制作为例子进行说明。
根据如图2所示出的信号在装置发射端1和装置接收端2的处理流程，具体步骤如下：
S1：装置发射端1的信号处理。
S11：采用随机数发生器101产生0,1比特序列，构成原始数据。
S12：经过BPSK调制，产生信号x。每一组的发送数据都可以不相同。
S13：压缩复用处理，即稀疏向量发送数据乘以压缩复用矩阵A，得到数据向量z。A表示压缩感知里面的测量矩阵，这里选择为N t×l的高斯矩阵。其中，ρ＝N t/l代表压缩比，表示天线减少数量的比例。最后，将数据经信道发射出去。
S2：装置接收端2的信号检测。
S21：接收到的信号为y＝Hz+n＝HAx+n，其中，n表示噪声。根据信道估计模块201，估计信号矩阵H。
S22：提前利用深度学习的BP训练算法，得到第一神经网络模型和第二神经网络模型。
具体地，当发送信号为x时，装置接收端2信号为y，同时，从理论上来说，我们总可以找到一个合适的基，实现信号的稀疏表示。Temlyakov在文献”Nonlinear Methods of Approximation,IMI Research Reports,Dept.of Mathematics,University of South Carolina,2001中指出字典D的选择应尽可能好地符合被逼近信号的结构，其构成可以没有任何限制。这里将x的所有可能组合，分别作为冗余字典D的不同列向量，组成冗余字典D，实现x的稀疏表示：除了相应索引位置为1，其他位置均为0的稀疏向量θ。利用y和θ，构成了第一组训练样本；θ和x，构成了第二组训练样本。经由深度学习的BP训练算法，第一组训练样本得到第一神经网络模型，第二组训练样本得到第二神经网络模型，当输入为信号y时，即得到重构的发送数据
Figure PCTCN2020083563-appb-000005
S23：第一神经网络信号处理模块202通过其所构建的第一神经网络模型来从低维的目标信号中解出高维的稀疏信号θ。第一神经网络模型的输入信号为y，输出信号是原始信号x在过完备冗余字典D上的稀疏表示θ。
S24：第二神经网络信号处理模块通过将所述稀疏信号θ输入其所构建的第二神经网络模型来重构得到原始信号x。第二神经网络模型的输入信号是原始信号x在过完备冗余字典D上的稀疏表示θ，输出信号为原始的发送数据流
Figure PCTCN2020083563-appb-000006
如下对步骤S22～S24进一步说明：神经网络模型由输入层(layer1)、中间层(layer2,..,L-1)、输出层(layerL)三部分组成，输入层起信号传输作用，负责接收外部输入信息，输入层每个单元代表一个特征；中间层可以是单中间层或多中间层，其起内部信息处理作用，负责信息变换；输出层起向外部输出信息作用，输出层的每个单元代表一个类别。本申请中利用BP神经网络模拟一个映射函数，此函数可以把输入空间数据映射到输出空间；BP神经网络会尽量的去拟合一个原始的装置接收端2信号y和原始信号x在过完备冗余字典D上的稀疏表示θ之间的函数，以及稀疏表示θ与原始的发送数据流
Figure PCTCN2020083563-appb-000007
之间的函数；基于训练的BP神经网络产生的映射函数模型，可以根据装置接收端2接收 到的信号y还原出理论推算的原始信号x在过完备冗余字典D上的稀疏表示θ，以及根据上述理论推算出的稀疏表示θ重构出原始的发送数据x。
本申请中使用代价函数来衡量BP神经网络输出与真实输出间的差异，训练BP神经网络使得网络的输入(装置接收端2接收信号)经过神经网络后的输出能够尽可能的接近理论输出。为了使代价函数最小，本申请中使用梯度下降法来求解神经网络参数，当求解到最优的神经网络权重，建立第一神经网络模型或第二神经网络模型。创建BP神经网络，收集大量的样本数据，并且人为的标记正确的分类结果，然后用这些标记好的数据来训练所创建的神经网络。在这个过程中，根据当前的输出值以及被标记的正确的目标值之间的差异，神经网络中的每一层都在不断的调整自身的权重和偏置，直到能够准确输出目标值。
针对训练神经网络时需用到的两个参数——权重和偏置进一步说明：本申请中，神经网络各层间的权重参数矩阵，用
Figure PCTCN2020083563-appb-000008
表示，其中，权重参数w的上标表示层数，下标的相邻两层各自的第几个节点。例如，
Figure PCTCN2020083563-appb-000009
表示输入层Layer1的第1个节点与Layer2的第2个节点的线段的权重。这些权重决定了模型的作用，神经网络的目标就是通过样本来计算权重。每一个中间层和输出层的节点都是一个Logistic函数g(z)＝a。例如
Figure PCTCN2020083563-appb-000010
表示Layer2的第1个节点的输入值，带入Logistic函数得到输出
Figure PCTCN2020083563-appb-000011
神经网络各层间的偏置参数矩阵为：B＝[b 1b 2…b n] T，已知神经网络的输入为：Y＝[y 1y 2…y n] T,神经网络的输出为：
Figure PCTCN2020083563-appb-000012
引入一个非线性算子：
Figure PCTCN2020083563-appb-000013
Figure PCTCN2020083563-appb-000014
则可以推导出：
Figure PCTCN2020083563-appb-000015
初始化权重参数：将权重参数w随机初始化为[-ε,ε]之间，ε是预设的一个足够小的值。
训练神经网络模型：训练神经网络模型的过程主要分为两步，一是计算代价函数J(θ)，二是调整参数θ，使得代价函数值J(θ)尽量小。如下采用正向传播算法，对每一个样本计算当前神经网络模型下该样本的输出，求出代价函数，再根据输出来更新权重参数。定义代价函数J(θ)，m是样本的个数，由于神经网络有K个输出，其代价函数也相应地计算了K个输出的代价，其计算公式为：
Figure PCTCN2020083563-appb-000016
如下采用反向传播算法调整参数θ，使得代价函数值J(θ)尽量小。反向传播算法通过求代价函数关于各个权重系数的偏导数，以此来更新各个权重系数。例如，首先，计算最后一层的梯度：(1)计算代价函数值对非线性算子的梯度，(2)计算神经网络的输出对偏置及相邻层间权重的梯度。并按照梯度的负方向更新梯度；其次，计算倒数第二层的梯度：(1)计算上层回传的误差对非线性算子的梯度，(2)计算H n-1(H为每层经过激活函数后的输出)对偏置及相邻层间权重的梯度，并按照梯度的负方向更新梯度；最后，经过一层一层回传之后，最后计算第一层的梯度：(1)计算第二层回传的误差对非线性算子的梯度；(2)计算H 1对偏置及相邻层间权重的梯度。并按照梯度的负方向更新梯度。如此，在第一次反向传播过程循环结束后，继续进行上面的正向传播得到输出，后向传播更新参数这两步，直到均方差最小，即完成神经网络模型的训练过程。
图3给出了MIMO系统不同收发天线配置在采用了压缩感知和神经网络的信号传输与检测技术后的误码率性能。这里假定平坦衰落信道。在传统的方案下，4根发射天线只能同时发送4个数据符号。应用本申请的方案，先采用BPSK调制信号，得到原始信号x 4×1，随机高斯矩阵A 4ρ×4作为压缩降维矩阵，从而得到z。如ρ＝0.5，则只需要2根发射天线，就可以实现原始数据的发送。如ρ＝0.75，则只需要3根发射天线，就可以实现原始数据的发送。装置接收端2，同样通过2或者3根接收天线，采用神经网路训练出的模型1和2，得到重构的发送信号
Figure PCTCN2020083563-appb-000017
本申请方案的误码率性能如图中(2×2)-4所示，括号内第一个数字表示发送天线数，第二个数字表示接收天线数，最后的数字表示原始数据长度。接收天线数有所增加，该方案记为(3×3)-4。和MIMO系统传统信号检测算法—迫零检测(ZF，Zero Forcing)比起来，本申请所提出的方案可以在高SNR的条件下，兼顾保证误码率的同时，减少需要的收发天线数。
上述迫零检测算法是利用滤波矩阵W ZF左乘接收信号y来消除各发送信号之间的干扰，从而估计出每个发送符号，滤波矩阵为：W ZF＝H -1＝(H HH) -1H H，因此估计信号向量为：
Figure PCTCN2020083563-appb-000018
得到估计信号向量之后将其映射到星座图中具有最近欧氏距离的星座点上，该星座点即作为最优解从而恢复出最终的符号向量X ZF。
图4给出了MIMO系统不同收发天线配置在采用了压缩感知和神经网络的信号传输与检测技术后的误码率性能。这里假定平坦衰落信道。在传统的方案下，20根发射天线只能同时发送20个数据符号。应用本申请的方案，先采用BPSK调制信号，然后将x 20×1分为5组，则每组的向量x i(i＝1,2,3,4,5)的长度等于4。随机高斯矩阵A 4ρ×4作为压缩降维矩阵，从而得到z i。将z i级联起来，得到待发送向量
Figure PCTCN2020083563-appb-000019
如ρ＝0.5，则只需要10根发射天线，就可以实现原始数据的发送。如ρ＝0.75，则需要15根发射天线，就可以实现原始数据的发送。装置接收端2，同样通过10根接收天线，该方案的误码率性能如图中(10×10)-20所示，括号内第一个数字发送天线数，第二个数字表示接收天线数，最后的数字表示原始数据长度。同样的数据，分组数也相同，ρ＝0.75，压缩降维矩阵为A 3×4时，接收天线数有所增加，该方案记为(15×15)-20。和经典检测算法ZF(Zero Forcing)比起来，我们提出的方案可以在高SNR的条件下，兼顾保证误码率的同时，减少需要的收 发天线数。可见，本申请所提出的方案可以在高SNR的条件下，兼顾保证误码率的同时，减少需要的收发天线数。
如上所述，采用本发明所提出的增强空间复用方法，可以在已有的MIMO系统的基础上，结合神经网络技术，在减少天线数的基础上，传输相同的数据量，减少需要的天线数，提高复用增益和系统容量。
需要注意的是，上述具体实施例是示例性的，本领域技术人员可以在本发明公开内容的启发下想出各种解决方案，而这些解决方案也都属于本发明的公开范围并落入本发明的保护范围之内。本领域技术人员应该明白，本发明说明书及其附图均为说明性而并非构成对权利要求的限制。本发明的保护范围由权利要求及其等同物限定。

Claims (10)

  1. 基于深度学习的MIMO多天线信号传输与检测装置，该装置包括：
    压缩复用模块(104)，用于对调制后的信号进行压缩降维处理；
    装置发射端(1)，用于在收发天线数给定的情况下将经过所述压缩复用模块(104)处理后的目标信号通过发射天线发射；
    装置接收端(2)，用于对接收到的信号进行处理以实现其对目标信号的重构，
    其特征是，
    所述装置接收端(2)中包括第一神经网络信号处理模块(202)和第二神经网络信号处理模块(203)，其中，
    第二神经网络信号处理模块(203)通过将由第一神经网络信号处理模块(202)利用其所构建的第一神经网络模型来从低维的目标信号中所解出的高维的稀疏信号θ输入其所构建的第二神经网络模型来重构得到原始信号x。
  2. 根据权利要求1所述的装置，其特征是，所述装置接收端(2)还包括信道估计模块(201)，其被配置为根据所述装置接收端(2)收到的经过压缩降维处理的低维目标信号进行信道估计并将得到的信道参数矩阵作为第一神经网络模型的输入。
  3. 根据权利要求2所述的装置，其特征是，所述第一神经网络信号处理模块(202)通过深度学习的BP算法创建神经网络，以所述装置接收端(2)的接收信号向量y和基于所述装置发射端(1)的发送信号向量x所确定的稀疏表示θ作为样本，构造第一组训练样本集，训练所述神经网络，得到所述第一神经网络模型。
  4. 根据权利要求3所述的装置，其特征是，通过将所述发送信号向量x的所有可能组合分别作为冗余字典的不同列向量的方式组成冗余字典D，实现所述发送信号向量的稀疏表示θ。
  5. 根据权利要求4所述的装置，其特征是，所述第二神经网络信号处理模块(203)通过深度学习的BP算法创建神经网络，以所述装置发射端(1)的发送信号向量x和所述基于所述装置发射端(1)的发送信号向量x所确定的稀疏表示θ作为样本，构造第二组训练样本集，训练所述神经网络，得到所述第二神经网络模型。
  6. 根据权利要求5所述的装置，其特征是，所述装置发射端(1)采用随机数发生器(101)产生一组随机的0、1二进制比特序列，构成原始数据；原始数据经过BPSK调制，产生调制后的信号x。
  7. 根据权利要求6所述的装置，其特征是，经调制后的l路信号x经过所述压缩复用模块(104)，压缩成ρl路信号，然后所述装置将压缩复用后的信号z经发射天线发射。
  8. 根据权利要求7所述的装置，其特征是，压缩复用后的ρ路信号z是通过计算式 z＝Ax来得到的，其中，A是N t行l列的压缩降维矩阵，ρ＝N t/l∈(0,1]代表压缩比例。
  9. 基于深度学习的MIMO多天线信号传输与检测系统，
    其特征是，
    所述系统通过构建MIMO端到端传输模型，根据装置接收端(2)收到的经过压缩降维处理的目标信号和估计的信道状态信息得到神经网络信号处理的输入，利用神经网络信号处理重构出原始信号。
  10. 基于深度学习的MIMO多天线信号传输与检测方法，其特征是，所述方法包括至少一个以下步骤：
    收发天线数给定的情况下将经过所述压缩复用处理后的目标信号通过发射天线发射；
    提前利用深度学习的BP训练算法，得到第一神经网络模型和第二神经网络模型；
    通过所构建的第一神经网络模型来从低维的目标信号中解出高维的稀疏信号θ；
    通过将所述稀疏信号θ输入所构建的第二神经网络模型来重构得到原始信号x。
PCT/CN2020/083563 2020-04-07 2020-04-07 基于深度学习的mimo多天线信号传输与检测技术 WO2021203242A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
CN202080000955.0A CN111630787B (zh) 2020-04-07 2020-04-07 基于深度学习的mimo多天线信号传输与检测技术
CN202310024809.7A CN116054888A (zh) 2020-04-07 2020-04-07 一种天线信号的原始信号重构方法及装置
PCT/CN2020/083563 WO2021203242A1 (zh) 2020-04-07 2020-04-07 基于深度学习的mimo多天线信号传输与检测技术

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/083563 WO2021203242A1 (zh) 2020-04-07 2020-04-07 基于深度学习的mimo多天线信号传输与检测技术

Publications (1)

Publication Number Publication Date
WO2021203242A1 true WO2021203242A1 (zh) 2021-10-14

Family

ID=72273280

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/083563 WO2021203242A1 (zh) 2020-04-07 2020-04-07 基于深度学习的mimo多天线信号传输与检测技术

Country Status (2)

Country Link
CN (2) CN111630787B (zh)
WO (1) WO2021203242A1 (zh)

Cited By (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114039633A (zh) * 2021-11-04 2022-02-11 网络通信与安全紫金山实验室 用于线性轨迹的智能波束预测方法、装置、设备及介质
CN114124168A (zh) * 2021-11-01 2022-03-01 南京邮电大学 一种基于深度学习的mimo-noma系统信号检测方法及系统
CN114337746A (zh) * 2022-01-07 2022-04-12 重庆邮电大学 一种基于深度学习的大规模mimo信号检测方法
CN114500189A (zh) * 2022-01-24 2022-05-13 华南理工大学 一种可见光通信直接预均衡方法、系统、装置及介质
CN114614920A (zh) * 2022-03-10 2022-06-10 南京邮电大学 基于学习因子图的数据与模型联合驱动的信号检测方法
CN114745234A (zh) * 2022-04-02 2022-07-12 华南理工大学 一种深度学习mimo系统信号检测方法及系统
US20220239510A1 (en) * 2021-01-25 2022-07-28 Marvell Asia Pte Ltd Ethernet physical layer transceiver with non-linear neural network equalizers
CN114826461A (zh) * 2022-04-02 2022-07-29 大连大学 一种基于改进自适应无参数递归神经网络的无线信道状态预测方法
CN114844541A (zh) * 2022-04-22 2022-08-02 南通大学 一种基于深度学习的混合预编码和反馈方法
CN114884775A (zh) * 2022-03-31 2022-08-09 南京邮电大学 一种基于深度学习的大规模mimo系统信道估计方法
CN115065446A (zh) * 2022-04-29 2022-09-16 台州宏创电力集团有限公司科技分公司 一种基于流形学习的信道信息反馈方法
CN115060985A (zh) * 2022-06-21 2022-09-16 北京电磁方圆科技有限公司 多通道天线的测试方法及装置
CN115598966A (zh) * 2022-10-26 2023-01-13 电子科技大学（Cn） 一种六自由度并联摇摆台的视觉伺服神经网络控制方法
CN115694571A (zh) * 2022-10-31 2023-02-03 西安科技大学 一种大规模mimo系统中基于深度学习的信号检测方法和装置
CN116032332A (zh) * 2022-12-30 2023-04-28 哈尔滨工程大学 一种基于icgan网络的大规模mimo系统检测模型构建方法
CN116506261A (zh) * 2023-06-27 2023-07-28 南昌大学 一种可见光通信感知方法及系统
CN116506072A (zh) * 2023-06-19 2023-07-28 华中师范大学 基于多任务联邦学习的mimo-noma系统的信号检测方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11625610B2 (en) * 2019-03-12 2023-04-11 Samsung Electronics Co., Ltd Multiple-input multiple-output (MIMO) detector selection using neural network
CN112054832B (zh) * 2020-09-08 2021-10-15 西安交通大学 一种深度学习多输入多输出检测方法
CN116938644A (zh) * 2022-03-31 2023-10-24 华为技术有限公司 一种信号处理方法和装置
CN115361258B (zh) * 2022-07-14 2023-06-16 鹏城实验室 一种大规模mimo稀疏信道估计方法及相关设备
CN115695105B (zh) * 2023-01-03 2023-03-17 南昌大学 基于深度迭代智能超表面辅助通信的信道估计方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103188058A (zh) * 2011-12-27 2013-07-03 财团法人工业技术研究院 通道信息反馈方法及其无线通信装置
US20140293904A1 (en) * 2013-03-28 2014-10-02 Futurewei Technologies, Inc. Systems and Methods for Sparse Beamforming Design
CN106301384A (zh) * 2016-08-26 2017-01-04 辽宁工程技术大学 一种基于分块压缩感知的信号重构方法
CN107368887A (zh) * 2017-07-25 2017-11-21 江西理工大学 一种深度记忆卷积神经网络的结构及其构建方法

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105610477B (zh) * 2016-01-28 2018-06-19 西南交通大学 基于压缩感知的多发多收系统增强信号复用方法
US9990687B1 (en) * 2017-01-19 2018-06-05 Deep Learning Analytics, LLC Systems and methods for fast and repeatable embedding of high-dimensional data objects using deep learning with power efficient GPU and FPGA-based processing platforms
US20190087713A1 (en) * 2017-09-21 2019-03-21 Qualcomm Incorporated Compression of sparse deep convolutional network weights
CN109921882B (zh) * 2019-02-20 2021-08-27 深圳市宝链人工智能科技有限公司 一种基于深度学习的mimo解码方法、装置及存储介质
CN110187321B (zh) * 2019-05-30 2022-07-22 电子科技大学 基于深度学习的复杂环境下雷达辐射源特征参数提取方法
CN110446173B (zh) * 2019-07-31 2021-08-27 中国地质大学（武汉） 一种高效节能的星载无线传感网数据压缩方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103188058A (zh) * 2011-12-27 2013-07-03 财团法人工业技术研究院 通道信息反馈方法及其无线通信装置
US20140293904A1 (en) * 2013-03-28 2014-10-02 Futurewei Technologies, Inc. Systems and Methods for Sparse Beamforming Design
CN106301384A (zh) * 2016-08-26 2017-01-04 辽宁工程技术大学 一种基于分块压缩感知的信号重构方法
CN107368887A (zh) * 2017-07-25 2017-11-21 江西理工大学 一种深度记忆卷积神经网络的结构及其构建方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
CHEN MUHAN, GUO JIAJIA, LI XIAOJINSHI: "An overview of the CSI feedback based on deep learning for massive MIMO systems", CHINESE JOURNAL ON INTERNET OF THINGS, vol. 4, no. 1, 31 March 2020 (2020-03-31), pages 33 - 44, XP055856457, DOI: 10.11959/j.issn.2096−3750.2020.00157 *
LIU CHANZI; ZHOU QINGFENG; WANG XINDI; CHEN KAIPING: "MIMO Signal Multiplexing and Detection Based on Compressive Sensing and Deep Learning", IEEE ACCESS, IEEE, USA, vol. 7, 1 January 1900 (1900-01-01), USA , pages 127362 - 127372, XP011746358, DOI: 10.1109/ACCESS.2019.2937490 *

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220239510A1 (en) * 2021-01-25 2022-07-28 Marvell Asia Pte Ltd Ethernet physical layer transceiver with non-linear neural network equalizers
CN114124168A (zh) * 2021-11-01 2022-03-01 南京邮电大学 一种基于深度学习的mimo-noma系统信号检测方法及系统
CN114124168B (zh) * 2021-11-01 2023-05-26 南京邮电大学 一种基于深度学习的mimo-noma系统信号检测方法及系统
CN114039633A (zh) * 2021-11-04 2022-02-11 网络通信与安全紫金山实验室 用于线性轨迹的智能波束预测方法、装置、设备及介质
CN114337746A (zh) * 2022-01-07 2022-04-12 重庆邮电大学 一种基于深度学习的大规模mimo信号检测方法
CN114500189A (zh) * 2022-01-24 2022-05-13 华南理工大学 一种可见光通信直接预均衡方法、系统、装置及介质
CN114614920A (zh) * 2022-03-10 2022-06-10 南京邮电大学 基于学习因子图的数据与模型联合驱动的信号检测方法
CN114614920B (zh) * 2022-03-10 2023-06-30 南京邮电大学 基于学习因子图的数据与模型联合驱动的信号检测方法
CN114884775A (zh) * 2022-03-31 2022-08-09 南京邮电大学 一种基于深度学习的大规模mimo系统信道估计方法
CN114826461A (zh) * 2022-04-02 2022-07-29 大连大学 一种基于改进自适应无参数递归神经网络的无线信道状态预测方法
CN114745234A (zh) * 2022-04-02 2022-07-12 华南理工大学 一种深度学习mimo系统信号检测方法及系统
CN114826461B (zh) * 2022-04-02 2023-09-26 大连大学 一种基于改进自适应无参数递归神经网络的无线信道状态预测方法
CN114844541A (zh) * 2022-04-22 2022-08-02 南通大学 一种基于深度学习的混合预编码和反馈方法
CN115065446A (zh) * 2022-04-29 2022-09-16 台州宏创电力集团有限公司科技分公司 一种基于流形学习的信道信息反馈方法
CN115065446B (zh) * 2022-04-29 2024-03-29 台州宏创电力集团有限公司科技分公司 一种基于流形学习的信道信息反馈方法
CN115060985A (zh) * 2022-06-21 2022-09-16 北京电磁方圆科技有限公司 多通道天线的测试方法及装置
CN115598966A (zh) * 2022-10-26 2023-01-13 电子科技大学（Cn） 一种六自由度并联摇摆台的视觉伺服神经网络控制方法
CN115694571A (zh) * 2022-10-31 2023-02-03 西安科技大学 一种大规模mimo系统中基于深度学习的信号检测方法和装置
CN116032332A (zh) * 2022-12-30 2023-04-28 哈尔滨工程大学 一种基于icgan网络的大规模mimo系统检测模型构建方法
CN116032332B (zh) * 2022-12-30 2024-04-12 哈尔滨工程大学 一种适用于信道状态信息多变的大规模mimo系统检测模型构建方法
CN116506072A (zh) * 2023-06-19 2023-07-28 华中师范大学 基于多任务联邦学习的mimo-noma系统的信号检测方法
CN116506072B (zh) * 2023-06-19 2023-09-12 华中师范大学 基于多任务联邦学习的mimo-noma系统的信号检测方法
CN116506261A (zh) * 2023-06-27 2023-07-28 南昌大学 一种可见光通信感知方法及系统
CN116506261B (zh) * 2023-06-27 2023-09-08 南昌大学 一种可见光通信感知方法及系统

Also Published As

Publication number Publication date
CN111630787B (zh) 2022-12-20
CN111630787A (zh) 2020-09-04
CN116054888A (zh) 2023-05-02

Similar Documents

Publication Publication Date Title
WO2021203242A1 (zh) 基于深度学习的mimo多天线信号传输与检测技术
WO2021203243A1 (zh) 基于人工智能的mimo多天线信号传输与检测技术
CN109672464B (zh) 基于fcfnn的大规模mimo信道状态信息反馈方法
CN108390706B (zh) 一种基于深度学习的大规模mimo信道状态信息反馈方法
Liao et al. CSI feedback based on deep learning for massive MIMO systems
Sohrabi et al. Deep learning for distributed channel feedback and multiuser precoding in FDD massive MIMO
CN112737985B (zh) 基于深度学习的大规模mimo信道联合估计和反馈方法
CN110912598B (zh) 基于长短时注意力机制的大规模mimo系统csi反馈方法
CN109257309B (zh) 一种高性能的大规模mimo下行链路传输信道估计方法
Chen et al. Deep learning-based implicit CSI feedback in massive MIMO
Ozpoyraz et al. Deep learning-aided 6G wireless networks: A comprehensive survey of revolutionary PHY architectures
Liu et al. A Markovian model-driven deep learning framework for massive MIMO CSI feedback
CN113972939B (zh) 基于双时间尺度和深度学习的天线系统预编码方法和装置
CN116248156A (zh) 基于深度学习的大规模mimo信道状态信息反馈及重建方法
CN114584448A (zh) 基于深度神经网络的sm-ofdm信号分组检测方法
Zhang et al. Pilot-assisted MIMO-V-OFDM systems: Compressed sensing and deep learning approaches
CN113660020A (zh) 一种无线通信信道信息传输方法、系统和解码器
CN114844541B (zh) 一种基于深度学习的混合预编码和反馈方法
CN114759997B (zh) 一种基于数据模型双驱动的mimo系统信号检测方法
Liu et al. CSI feedback based on complex neural network for massive MIMO systems
CN113078931B (zh) 基于qga-omp算法的毫米波大规模mimo混合预编码方法及系统
CN114598574A (zh) 一种基于深度学习的毫米波信道估计方法
Liao et al. Structured neural network with low complexity for MIMO detection
CN116436551A (zh) 信道信息传输方法及装置
Ahmed et al. Energy efficient ADC bit allocation for massive MIMO: A deep-learning approach

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20930471

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 07/03/2023)

122 Ep: pct application non-entry in european phase

Ref document number: 20930471

Country of ref document: EP

Kind code of ref document: A1