A method and apparatus for reconstructing a raw signal of an antenna signal

Abstract

The application relates to an antenna signal original signal reconstruction method and device, and the method comprises the following steps: obtaining a first neural network model and a second neural network model by using a BP training algorithm of deep learning; solving a high-dimensional sparse signal theta from a low-dimensional target signal by using the constructed first neural network model; and reconstructing an original signal by inputting the sparse signal theta into the constructed second neural network model. The application can simultaneously transmit the same data amount on the basis of reducing the required antenna number, and improve the multiplexing gain and capacity of a MIMO system.

Description

[0001] The original basis for this divisional application is the patent application with application number (202080000955.0), application date of April 7, 2020, entitled "MIMO multi-antenna signal transmission and detection technology based on deep learning". Technical Field

[0002] This invention relates to the field of mobile communication technology, and in particular to a method and apparatus for reconstructing the original signal of an antenna signal. Background Technology

[0003] MIMO (Multiple-Input Multiple-Output) technology refers to the use of multiple transmit and receive antennas at both the transmitting and receiving ends of a device, enabling signals to be transmitted and received through these antennas, thereby improving communication quality. Based on the number of antennas at both ends, MIMO includes SIMO (Single-Input Multiple-Output) systems and MISO (Multiple-Input Single-Output) systems, compared to ordinary SISO (Single-Input Single-Output) systems. It fully utilizes spatial resources, achieving multiple transmissions and receptions through multiple antennas. Without increasing spectrum resources or antenna transmit power, it can significantly increase system channel capacity, demonstrating a clear advantage. MIMO systems are widely used in wireless communication—mobile devices and networks commonly use multiple antennas to enhance connectivity, improve network speed, and enhance user experience. Massive MIMO is a key factor in 5G's ultra-high data rates, bringing greater network capacity, wider signal coverage, and a better user experience, pushing the potential of 5G to a whole new level.

[0004] Based on different space-time mapping methods, MIMO technology can be broadly divided into two categories: spatial diversity and spatial multiplexing. Spatial diversity refers to using multiple transmit antennas to send signals with the same information through different paths, while simultaneously obtaining multiple independently fading signals of the same data symbol at the receiver, thereby achieving improved reception reliability through diversity. For example, in a slow Rayleigh fading channel, using a single transmit antenna N... r One receiving antenna, the transmitted signal passes through N r There are several different paths. If the fading between each antenna is independent, the maximum diversity gain can be obtained as N. r For transmit diversity technology, the gain from multiple paths is also used to improve system reliability. In a system with N... t Root transmitting antenna N rIn a system with a root receiving antenna, if the path gain between antenna pairs is an independent and uniformly distributed Rayleigh fading, 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 fading channels. Therefore, it is crucial to reduce the impact of multipath fading on base stations and mobile stations without increasing power or sacrificing bandwidth. The only solution is to employ anti-fading techniques, and effective methods to overcome multipath fading include various diversity techniques. Diversity techniques are primarily used to combat channel fading. Conversely, the fading characteristics in MIMO channels can provide additional information to increase the degrees of freedom in communication. Essentially, if the fading between each pair of transmit and receive antennas is independent, multiple parallel sub-channels can be generated. Transmitting different information streams on these parallel sub-channels can provide a data transmission rate, a phenomenon known as spatial multiplexing. However, at high SNR (Signal-to-Noise Ratio), the transmission rate is limited in its degrees of freedom.

[0005] MIMO has two major advantages: (1) Increased channel capacity. MIMO access points and MIMO clients can simultaneously transmit and receive multiple spatial streams. The channel capacity increases linearly with the number of antennas, thus multiplying the wireless channel capacity without increasing bandwidth or antenna transmission power; (2) Improved channel reliability. Utilizing the spatial multiplexing and spatial diversity gains provided by MIMO channels, multiple antennas can be used to suppress channel fading. The application of multi-antenna systems allows parallel data streams to be transmitted simultaneously, significantly overcoming channel fading and reducing bit error rate. It has become a core technology used in 802.11n. 802.11n is a new wireless LAN technology from IEEE following 802.11a / b / g, with speeds up to 600Mbps. Meanwhile, MIMO technology can improve the performance of existing 802.11a / b / g networks.

[0006] As the number of antennas used increases, the complexity of implementing MIMO technology increases significantly, thus limiting the number of antennas that can be used and preventing the full realization of the advantages of MIMO technology. Currently, how to reduce the algorithmic and implementation complexity of MIMO technology while ensuring a certain level of system performance has become a major challenge for the industry.

[0007] Artificial intelligence (AI) refers to the intelligence exhibited by machines created by humans. Generally, AI refers to the technology of simulating certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning) through ordinary computer programs to present human intelligence. It mainly includes the principles of computer intelligence, the creation of computers similar to human brain intelligence, and enabling computers to achieve higher-level applications. We firmly believe that in the coming years, AI will continue to make significant strides, creating value for traditional industries and profoundly changing our daily lives, such as in robotics, speech recognition, image recognition, and expert systems. Machine learning is a subset of AI. Currently, machine learning mainly solves classification, clustering, and regression problems and is widely used in character recognition, machine translation, speech recognition, search engines, facial recognition, and autonomous driving. Deep learning is the most crucial of all machine learning algorithms.

[0008] The concept of deep learning originated from the research on artificial neural networks. The concept of deep learning was first proposed by Professor G.E. Hinton of the University of Toronto, a leading figure in the field of machine learning. His two core viewpoints are: (1) Artificial neural networks with multiple hidden layers have excellent feature learning capabilities, providing a deeper representation of the learned feature data and enabling better classification or visualization of the final network data; (2) Deep neural networks can overcome the difficulty of training their own network parameters through "layer-by-layer initialization," which can be achieved through unsupervised learning. A multilayer perceptron with multiple hidden layers is a type of deep learning structure. Deep learning discovers distributed feature representations of data by combining low-level features to form more abstract high-level representations of attribute categories or features. Any neural network can have any number of layers, inputs, or outputs. The layer between the input neurons and the last output neuron is the hidden layer of a deep neural network.

[0009] There are also many invention patents both domestically and internationally related to applying deep learning to solve the MIMO problem.

[0010] Existing technologies, such as the patent document CN105610477B with authorization announcement date of June 19, 2018, propose a method for enhancing signal multiplexing in a multiple-transmitter-multiple-receiver system based on compressed sensing. Building upon existing MIMO technology and related MIMO system signal multiplexing techniques, this method selects a random measurement matrix from compressed sensing as the signal compression multiplexing matrix. It then fully utilizes the sparsity characteristics of the transmitted signal on an overcomplete redundant dictionary, and through a compressed sensing reconstruction algorithm, decomposes the high-dimensional transmitted signal from the low-dimensional received multiplexed signal. This significantly improves the signal multiplexing gain under a given number of transmit and receive antennas in a MIMO system, better meeting the application requirements of MIMO systems for broadband transmission. Furthermore, it has the advantages of ensuring that the receiver can reconstruct the multiple data streams transmitted by the transmitter through compression multiplexing steps with a high probability using a mature optimized reconstruction algorithm in the field of compressed sensing, and requires minimal modification to existing MIMO systems.

[0011] Chinese patent application number 201510473741.6 (Complex Neural Network Channel Prediction Method, Xi'an University of Electronic Science and Technology) discloses a complex neural network channel prediction method, mainly addressing the channel fading problem caused by time-varying channel dynamics in MIMO systems. The technical solution is as follows: 1. The base station measures the channel to obtain a channel coefficient training sequence containing estimation errors; 2. Based on the obtained channel coefficient sequence, corresponding training samples and expected outputs are obtained; 3. The training samples are input to train a complex wavelet neural network to obtain the final network weights; 4. The base station uses the trained complex wavelet neural network to predict channel coefficients. This method is simple, easy to implement, and effective, suitable for reducing the impact of time-varying channel dynamics on the channel of MIMO systems.

[0012] Chinese patent application number 201810177829.7 (A Wireless Channel Modeling Method Based on Neural Networks, Southeast University) discloses a wireless channel modeling method based on neural networks. This invention first processes the received signal fed back by the user to obtain estimated channel parameters; then, it obtains the three-dimensional geographic information of the scatterers from a two-dimensional image and clusters them; finally, it uses the channel parameters and geographic information as input to the neural network and the received signal as output to train a nonlinear time-varying neural network model. This method obtains a more accurate channel model within acceptable complexity, meeting the channel modeling requirements of future 5G communication systems such as massive MIMO technology, bandwidth expansion, and high mobility scenarios.

[0013] Chinese patent application number 201810267976.3 (A Deep Neural Network for Detecting Large-Scale MIMO Systems Based on BP Algorithm, Southeast University) provides a deep neural network method for detecting large-scale MIMO systems based on the BP algorithm. This method constructs a deep neural network for large-scale MIMO system detection by expanding and mapping the factor graph of the confidence propagation iterative algorithm onto a neural network structure. The neurons in the deep neural network correspond to the nodes in the factor graph of the iterative algorithm, and the number of neurons in each layer is equal to the number of symbol nodes in the factor graph. The mapping function between hidden layers is the update formula for confidence information in the iterative algorithm, and the number of hidden layers is equal to the number of iterations of the iterative algorithm. Specifically, this invention also provides MIMO detection methods based on two information propagation iterative algorithms: damped confidence propagation and maximum sum confidence propagation, respectively, to construct two deep neural networks. This invention achieves a lower bit error rate without increasing online computational complexity and is robust to various channel conditions and antenna configurations.

[0014] Chinese patent application number 201910063733.2 (An Optimized MIMO Detection Method Based on Deep Learning, Shanghai University) constructs an end-to-end MIMO transmission model. It uses the signal y(t) received at the MIMO device receiver and estimated imperfect channel state information to obtain the complex time-domain vector of the model as input to a deep neural network (DNN). The DNN is used to obtain an estimate of the bitstream at the device transmitter. Compared to existing technologies that obtain the estimate of the transmitted bitstream through hard decision-making, this invention improves accuracy and detection rate under imperfect channel information, ensuring low bit error rate detection performance with low-complexity algorithms, while also exhibiting good robustness even with inherent channel errors.

[0015] Chinese patent application number 201610327115.0 (A Deep Learning-Based Codebook Selection Method for Large-Scale MIMO, Chongqing University of Posts and Telecommunications) relates to a deep learning-based codebook selection method for large-scale MIMO. The method includes: collecting pilot information from the test area to construct a pilot training sequence, thereby obtaining pilot training samples; performing iterative learning on the pilot training samples using a neural network to obtain the final network weight values; and selecting the optimal codeword from the complete codebook based on the channel output by the learned neural network. Then, the unknown area is matched with the test area to obtain its wireless channel, and thus the codeword corresponding to the wireless channel is obtained. This invention can effectively, accurately, and quickly establish a wireless channel model and codebook lookup, avoiding channel estimation in unknown areas and greatly reducing the complexity of codebook selection in unknown areas.

[0016] Chinese patent application number 201811626005.X (A Signal Detection Method for Low-Complexity MIMO-NOMA Systems Based on Improved Gradient Projection Method, Chongqing University of Posts and Telecommunications) discloses a signal detection method for low-complexity MIMO-NOMA systems based on an improved gradient projection method, relating to wireless communication technology. Based on the sparse characteristics of active users in the system, the method utilizes the idea of ​​convex optimization algorithms to transform the system model into a rigorous quadratic programming problem. This problem is then iteratively solved, and the results of each iteration are preprocessed to achieve effective detection of active users and their signals. This invention overcomes the problem of slow convergence speed in traditional detection methods. Preprocessing the results of each iteration not only enables rapid convergence of detection results but also allows for the detection of the active user set. Its implementation is simple and its application range is wide.

[0017] Chinese patent application number 201910014714.0 (A Beamforming Matrix Design Method for MIMO Systems Based on Deep Learning, Nanjing University of Posts and Telecommunications) discloses a beamforming matrix design method for MIMO systems based on deep learning. The steps are as follows: First, a training sample set required for the deep learning network is obtained using a known algorithm; then, a deep learning neural network model is constructed, its parameters are initialized, and it is trained using the training sample set; next, the channel is acquired using pilot signals and fed into the neural network to predict the beamforming matrix coefficients; finally, the channel and beamforming matrix coefficients are combined to form the beamforming matrix. This method, using a deep learning neural network to obtain a beamforming matrix, can simultaneously balance performance and algorithm complexity, reducing latency while ensuring performance, enabling the MIMO system to provide real-time services.

[0018] Chinese patent application number 201810182937.3 (A Machine Learning-Based Adaptive Transmission Method for MIMO Links, Southeast University) discloses a machine learning-based adaptive transmission method for MIMO links. This method uses an unsupervised learning autoencoder algorithm for feature extraction and dimensionality reduction, incorporating deep learning concepts to reduce feature dimensionality and computational complexity while preserving key state information. This invention utilizes logistic regression to construct a mapping relationship between channel state information and transmission parameters. Unlike previous fixed parameterized models, this method can be trained based on sample data. With high-quality datasets covering all states, it can better establish the mapping relationship between channel state information and transmission parameters, and compared to the traditional single equivalent signal-to-noise ratio, it can more fully utilize channel state information. Furthermore, this invention also performs CQI selection based on the channel matrix. By studying the channel matrix and noise variance, the machine learning-based adaptive transmission method for MIMO links is not constrained by receiver design and has universality.

[0019] Chinese patent application number 201710495044.X (A Deep Learning-Based Joint Precoding and Antenna Selection Method for MIMO Systems, Zhejiang Sci-Tech University) discloses a deep learning-based joint precoding and antenna selection method for MIMO systems, comprising the following steps: First, generating the training dataset required for deep learning using existing antenna selection methods; next, establishing a deep learning model, training the deep learning model using the training data, and saving the model; then, using the saved deep learning model to complete antenna selection; finally, performing optimal precoding design on the selected MIMO subsystem. This invention utilizes deep learning technology to design joint precoding and antenna selection for MIMO systems, achieving low computational complexity while obtaining a good system signal-to-noise ratio.

[0020] Chinese Patent Application No. 201910242525.9 (A Deep Signal Detection Method for High-Speed ​​Rail, Shenzhen University) proposes a deep signal detection method for high-speed rail. First, data is collected by gathering several transmitted and received signals in various scenarios along the high-speed rail line, based on different environmental types. Second, scenarios are divided into multiple regions through data analysis to ensure neural network compatibility. Third, a deep high-speed rail signal detection neural network model is established. Next, the high-speed rail signal detection neural network is trained offline. Finally, online real-time signal detection is performed. During high-speed rail operation, its location information is first determined via GPS to identify its region, and an appropriate neural network model is selected. The received signals are then input into the trained neural network, and the transmitted signals from the base station are output in real time. This invention significantly improves system performance, reduces the signal detection bit error rate, and enhances algorithm robustness. The method used in this invention eliminates the need for channel estimation, saving pilot overhead.

[0021] Chinese Patent Application No. 201810279530.2 (A Method for Anti-interference and Noise Reduction of Visible Light Communication MIMO Based on BP Neural Network, South-Central University for Nationalities) discloses a method for anti-interference and noise reduction of visible light communication MIMO based on BP neural network, relating to MIMO antenna technology in the field of visible light communication. This system includes a transmitting end of a system device, a signal processing section of a receiving end of the system device, and a BP neural network signal processing section connected in sequence. The method is as follows: 1) An electrical signal is loaded onto an LED array and transmitted in the form of an optical signal; 2) A photodetector at the receiving end of the device converts the optical signal into an electrical signal; 3) Multiple electrical signals are filtered by a low-pass filter to remove high-frequency interference; 4) After training, the BP neural network performs noise reduction and interference elimination processing on the multiple signals, and finally converts them into a binary serial data stream through parallel-to-serial conversion. This invention improves the transmission performance of existing MIMO technology; combines neural networks with visible light MIMO technology to leverage the advantages of neural networks in noise reduction and interference elimination in wireless communication; and uses neural network receiving and processing technology to make the entire VLC system more stable.

[0022] Chinese Patent Application No. 201710213235.2 (A Joint Equalization Method for Visible Light Channels Based on Orthogonal Mapping and Probabilistic Neural Networks, Sun Yat-sen University) discloses a joint equalization method for visible light channels based on orthogonal mapping and probabilistic neural networks. The method includes a transmitter and a receiver. Signals are transmitted from the transmitter to the receiver via a visible light MIMO channel. The visible light MIMO channel is a multiple-input multiple-output (MIMO) channel. The joint equalization combines pre-equalization and post-equalization. This invention employs a joint equalization scheme combining pre-equalization and post-equalization techniques, namely, a joint equalization method for visible light MIMO channels based on orthogonal mapping and probabilistic neural networks. This method effectively suppresses interference between channels in visible light MIMO communication systems and improves data transmission reliability.

[0023] Chinese Invention Patent No. 201910125325.5 (A Deep Learning-Based MIMO Decoding Method, Apparatus, and Storage Medium, Shenzhen Baolian Artificial Intelligence Technology Co., Ltd.) discloses a deep learning-based MIMO decoding method, apparatus, and storage medium. The invention involves constructing a training dataset for MIMO decoding, including multiple training data sets; then training a neural network based on this dataset to obtain a trained neural network model; upon receiving a MIMO signal to be decoded, inputting the signal into the neural network model for MIMO decoding, and finally obtaining the MIMO decoding result output by the neural network model. Through the implementation of this invention, a neural network model for joint MIMO detection and channel decoding is designed based on deep learning, treating MIMO detection and channel decoding as a joint decoding process. Furthermore, training improves the approximation of the neural network model's output, ensuring the overall performance of MIMO decoding and achieving higher decoding accuracy and faster decoding speed.

[0024] Chinese patent application number 201810757547.4 (A Machine Learning-Assisted Downlink User Scheduling Method for Large-Scale MIMO, Southeast University) discloses a machine learning-assisted downlink user scheduling method for large-scale MIMO, comprising the following steps: S1: The base station obtains the feature mode energy coupling matrix in the feature direction through the uplink probe signal sent by the user; S2: The base station uses the feature mode energy coupling matrix to perform sum rate calculation under various user and beam combinations with the assistance of machine learning methods; S3: A greedy algorithm is used to implement user scheduling based on the sum rate maximization criterion to obtain the optimal user beam pairing combination. This invention obtains statistical channel information through uplink probe signals and uses the sum rate maximization criterion for user scheduling. Even when the base station only has statistical channel information, through targeted feature extraction and neural network design, it accurately achieves approximate calculation of the sum rate, greatly reducing the complexity of user scheduling under large-scale antennas, and achieving near-optimal performance, with good applicability and robustness.

[0025] Chinese Invention Patent No. 201610353881.4 (A Modulation Recognition Method for MIMO Correlated Channels Based on Machine Learning Algorithms, Beijing University of Posts and Telecommunications) is a modulation recognition method for MIMO correlated channels based on machine learning algorithms, belonging to the field of communications. The specific steps are as follows: First, each data stream at the transmitting end of the communication device is space-time coded, and each codeword is transmitted through Nt transmitting antennas. Then, the MIMO channel matrix H is calculated based on the correlation matrix at the receiving end and the correlation matrix at the transmitting end. Based on the MIMO channel matrix H, the received signal on each receiving antenna is calculated and corrected. Finally, features are extracted from the corrected signal at each receiving antenna, and the extracted feature values ​​are used for training and testing to calculate the modulation recognition mode to which the sample belongs. The advantages are: strong robustness and generalization ability to non-Gaussian channels; modulation scheme recognition can be achieved in more complex environments through parameter iteration; and by extracting features of higher-order moments and higher-order cumulants, the signal feature differences are obvious at higher signal-to-noise ratios, facilitating classification by machine learning algorithms.

[0026] The application of MIMO technology makes space a resource that can be used to improve performance and increase the coverage of wireless systems. MIMO technology has become one of the key technologies in the field of wireless communication, and through continuous development in recent years, it has been increasingly applied to various wireless communication systems. However, with the increase in the number of antennas used, the complexity of implementing MIMO technology increases significantly, thus limiting the number of antennas that can be used and preventing the full realization of MIMO's advantages. Meanwhile, since the formal introduction of artificial intelligence as a discipline in 1956, it has made significant progress over the past 50 years, becoming a broad interdisciplinary and cutting-edge science. As an important branch of artificial intelligence, as can be seen from the above description, neural networks, implemented through computers to create mathematical networks, are widely recognized as a solution to some of the bottleneck problems currently encountered in communication.

[0027] Based on the aforementioned research on key technologies related to neural networks and communication issues both domestically and internationally, numerous research results have been achieved, proposing feasible solutions from the perspectives of MIMO channel estimation and signal detection. However, simultaneously, how to improve system capacity and reduce the algorithmic and implementation complexity of MIMO technology while ensuring certain system performance remains a significant challenge for the industry. Existing research on MIMO conditions has not yet explored solutions for improving system multiplexing gain using neural network technology, given a fixed number of transmit and receive antennas.

[0028] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the inventors studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0029] To address the challenge of improving system capacity and reducing the algorithmic and implementation complexity of MIMO technology while maintaining adequate system performance, existing research on MIMO has not yet explored solutions for enhancing signal transmission and detection by incorporating neural network technology to improve system multiplexing gain under given transmit and receive antenna numbers. This application, combining the latest advancements in compressed sensing technology with neural network techniques, proposes a signal multiplexing transmission and detection scheme for MIMO systems based on deep learning and compressed sensing. Firstly, compared to traditional MIMO schemes, this application introduces a compressed multiplexing module and a demultiplexing module at the receiver, reducing the required number of antennas while transmitting the same amount of data, thus improving the multiplexing gain and capacity of the MIMO system. Secondly, unlike existing MIMO spatial multiplexing techniques that only focus on eliminating interference between adjacent data, this application focuses not only on eliminating interference but also on how to multiplex more data streams to the receiver under a given number of transmit antennas, while ensuring the receiver's detection performance, thereby achieving multiplexing gain and transmission capacity exceeding the inherent capabilities of MIMO systems.

[0030] The MIMO device based on deep learning and compressed sensing proposed in this application mainly consists of a compression multiplexing module 104 at the transmitter 1 and a first neural network signal processing module 202 and a second neural network signal processing module 203 at the receiver 2. Preferably, the transmitter 1 also includes a random number generator 101 (or raw information bit generation module), a bit-level processing module 102, and a modulation module 103, while the receiver 2 includes a channel estimation module 201.

[0031] Preferably, such as Figure 5As shown, at the transmitting end 1 of the device: the raw data generated by the random number generator 101 is processed by the bit-level processing module 102 and the modulation module 103 to generate a modulated signal. The modulated signal is then processed by the compression and multiplexing module 104. The compression and multiplexing module 104 performs compression, dimensionality reduction and multiplexing processing on the transmitted signal. The compressed and multiplexed signal is then transmitted through the transmitting antenna. At the receiving end 2 of the device: the channel estimation module 201 performs channel estimation on the received signal. Based on the received signal and the estimated channel state information, the input of the first neural network signal processing module 202 is obtained. Based on the output of the first neural network signal processing module 202, the input of the second neural network signal processing module 203 is obtained. The second neural network model 203 reconstructs and outputs the original transmitted data stream x.

[0032] The above process specifically includes the following:

[0033] A deep learning-based MIMO multi-antenna signal transmission and detection device includes: a compression and multiplexing module for compressing and reducing the dimension of the modulated signal; a transmitter for transmitting the target signal processed by the compression and multiplexing module through the transmitter antenna, given a given number of transmit and receive antennas; and a receiver for processing the received signal to reconstruct the target signal. The receiver includes a first neural network signal processing module and a second neural network signal processing module. The second neural network signal processing module reconstructs the original signal x by inputting the high-dimensional sparse signal θ, which is solved from the low-dimensional target signal by the first neural network signal processing module using its constructed first neural network model, into its constructed second neural network model.

[0034] According to a preferred embodiment, the receiving end of the device further includes a channel estimation module, which is configured to perform channel estimation based on the low-dimensional target signal that has undergone compression and dimensionality reduction processing received by the receiving end of the device and use the obtained channel parameter matrix as input to a first neural network model.

[0035] According to a preferred embodiment, the first neural network signal processing module creates a neural network using the backpropagation algorithm of deep learning. It constructs a first set of training samples using the received signal vector y from the receiving end of the device and the sparse representation θ determined based on the transmitted signal vector x from the transmitting end of the device as samples, and trains the neural network to obtain the first neural network model.

[0036] According to a preferred embodiment, the sparse representation θ of the transmitted signal vector is achieved by forming a redundant dictionary D by using all possible combinations of the transmitted signal vector x as different column vectors of the redundant dictionary.

[0037] According to a preferred embodiment, the second neural network signal processing module creates a neural network using the backpropagation algorithm of deep learning, constructs a second set of training samples using the transmitted signal vector x from the transmitter of the device and the sparse representation θ determined based on the transmitted signal vector x from the transmitter of the device as samples, trains the neural network, and obtains the second neural network model.

[0038] According to a preferred embodiment, 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 then modulated by BPSK to generate the modulated signal x.

[0039] According to a preferred embodiment, the modulated l-channel signal x is compressed into p1-channel signal by the compression and multiplexing module, and then the device transmits the compressed and multiplexed signal z through the transmitting antenna.

[0040] According to a preferred embodiment, the compressed and multiplexed ρ1-channel signal z is obtained by calculating z = Ax, where A is N. t A compressed dimension reduction matrix with row l and column l, This represents the compression ratio.

[0041] The system is a deep learning-based MIMO multi-antenna signal transmission and detection system. The system constructs a MIMO end-to-end transmission model, obtains the input of neural network signal processing based on the target signal that has been compressed and reduced in dimensionality and the estimated channel state information received by the receiving end of the device, and reconstructs the original signal using neural network signal processing.

[0042] A deep learning-based MIMO multi-antenna signal transmission and detection method includes at least one step: transmitting the target signal after compression and multiplexing through the transmit antenna, given a given number of transmit and receive antennas; obtaining a first neural network model and a second neural network model in advance using a deep learning BP training algorithm; extracting a high-dimensional sparse signal θ from the low-dimensional target signal using the constructed first neural network model; and reconstructing the original signal x by inputting the sparse signal θ into the constructed second neural network model. Attached Figure Description

[0043] Figure 1 This is a schematic block diagram of the signal processing flow of the deep learning-based MIMO multi-antenna signal transmission and detection system provided by the present invention.

[0044] Figure 2 This is a schematic block diagram of the preferred signal compression multiplexing and detection processing procedure provided by the present invention;

[0045] Figure 3These are the bit error rate performance curves of the classic detection algorithm ZF and the multi-antenna signal transmission and detection technology of the MIMO system of this invention under different transmit and receive antenna configurations;

[0046] Figure 4 These are the bit error rate performance curves of the classic detection algorithm ZF and the multi-antenna signal transmission and detection technology of the MIMO system of this invention under different transmit and receive antenna configurations; and

[0047] Figure 5 This 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.

[0048] List of reference numerals

[0049] 1: Device transmitter 101: Random number generator

[0050] 102: Bit-level processing module; 103: Modulation module

[0051] 104: Compression and Multiplexing Module 2: Device Receiver

[0052] 201: Channel estimation module; 202: First neural network signal processing module

[0053] 203: Second Neural Network Signal Processing Module Detailed Implementation

[0054] The present invention will now be described in detail with reference to the accompanying drawings.

[0055] like Figure 2 As shown, the technical solution adopted in this invention is a signal transmission and detection technology and method for a MIMO multi-antenna system based on artificial intelligence and compressed sensing technology in a MIMO system. The technical solution is as follows:

[0056] 1. Signal processing at the transmitter end of the device: For a device equipped with N t root transmitting antenna and N r A MIMO communication system with a single receiving antenna. The system's transmitting end receives l-channel signals x after channel coding and signal modulation, which are then compressed into pl-channel signals by a compression and multiplexing module 104. The compressed and multiplexed signals z are then transmitted via the transmitting antenna. The compression processing of the input signals by the compression and multiplexing module 104 can be expressed as: z = Ax, where x = [x1, x2, ..., xn]. l ] T Represents the coded modulation symbol of the l-channel, where A is N. t A compressed dimension reduction matrix with row l and column l, The compression ratio ρ represents the compression ratio. Preferably, in this embodiment, a Gaussian random matrix is ​​selected as the compressed sensing multiplexing matrix / compression dimensionality reduction matrix A. The compression ratio ρ is determined by the size of the compression dimensionality reduction matrix A in the compressed sensing technology.

[0057] 2. Signal detection at receiver 2: The link at receiver 2 is roughly the reverse process of the link at transmitter 1. The signal received by receiver 2 is: y = Hz + n = HAx + n. Where y is an Nr×1 received signal vector, i.e., the signal received by receiver 2 at N... t The transmitting antenna transmits l modulation symbols; z is an Nt×1 transmitted signal vector; n is an Nr×1 Gaussian white noise vector; H is an Nr×Nt channel propagation matrix, which is a deterministic matrix that remains unchanged over a coherent time interval. The receiving end 2 of the device can estimate the channel propagation matrix H based on the pilot signals inserted into the transmitted data.

[0058] The aforementioned channel estimation refers to the process by which the receiver 2 determines the state (uncertainty) of the wireless transmission channel by processing the data from the transmitter 1. A common method is non-blind channel estimation based on pilot symbols, where the transmitter 1 sends known pilot information, and the receiver 2 processes this information to derive the channel state.

[0059] 3. Reconstruct the original transmitted data stream based on the first and second neural network models obtained from training. In the first neural network model, the input signal to the training set is y, and the output signal is θ, the sparse representation of the original signal x on the overcomplete redundant dictionary D. In the second neural network model, the input signal to the training set is θ, and the output signal is the original transmitted data stream. That is, the transmitting end N of the device t The compressed and multiplexed l-channel modulated symbols transmitted by the root transmit antenna. θ is the encoded l-channel modulated symbol x = [x1, x2, ..., x...]. l ] T Sparse representation on an overcomplete redundant dictionary D.

[0060] The specific implementation steps of the above process are explained below:

[0061] Regarding the construction of compressed multiplexing matrices: Compressed sensing, also known as compressed sampling or sparse sampling, is a method for finding sparse solutions to underdetermined linear systems. This method has existed for at least forty years, and recently, thanks to the work of David Donoho, Emmanuel Candès, and Terence Tao, the field has seen significant development. In recent years, compressed sensing technology has been widely applied in fifth-generation mobile communication systems, attracting considerable attention and research.

[0062] Compressed sensing originates from acquiring and reconstructing sparse or compressible signals. Candès and Donoho formally proposed the concept of compressed sensing in their papers "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. Utilizing the sparsity of the original signal, compared to Nyquist theory, it can reconstruct the entire high-dimensional signal from fewer measurements. Its core idea is to combine compression and sampling. First, a non-adaptive linear projection (measurement value) of the signal is acquired, and then the original signal is reconstructed from the measurement value using a corresponding reconstruction algorithm. Traditional signal acquisition and processing 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 less than twice the highest frequency in the analog signal's spectrum.

[0063] Compressed sensing theory differs from the traditional Nyquist sampling theorem. If a signal x is compressible or sparse in a certain transform domain D, then a high-dimensional sparse signal obtained from the transformation can be projected onto a low-dimensional space using an observation matrix A independent of the transform domain D. The original signal can then be reconstructed with high probability from these few projections by solving an optimization problem. Within this theoretical framework, the sampling rate is not determined by the signal bandwidth, but rather by the structure and content of the information in the signal. Compressed sensing theory mainly includes three aspects: sparse representation of signals, coded sampling, and reconstruction algorithms.

[0064] Sparse representation of a signal refers to representing the original signal as a sparse linear combination over a suitably chosen overcomplete basis (dictionary D = [d1, d2, ..., dp], or transform domain), where d1, d2, ..., dp are atoms in the dictionary. "Overcomplete basis" means that the number of atoms in the dictionary greatly exceeds the dimension of the original signal. Since signals in nature are generally not sparse, sparse representation of a signal means that when the signal is projected onto a transform domain D, only a few elements are non-zero. The resulting transform vector is then said to be sparse or approximately sparse, i.e., x = Dθ, where θ is a concise expression for the original signal x. This is a prerequisite for compressed sensing, meaning the signal must be sparsely representable under some transform. Theoretically, a transform domain D can always be found to achieve sparse representation of the signal. If the original signal x is itself sparse, then x = θ. Finding the optimal linear combination of multiple atoms from a redundant dictionary to represent a signal is called sparse approximation or highly nonlinear approximation of the signal.

[0065] Next, in compressed sensing theory, it is necessary to design the observation matrix A of the compressed sampling system, and how to sample a small number of observations while ensuring that the original signal can be reconstructed from them. Clearly, if the observation process destroys information in the original signal, the reconstruction quality cannot be guaranteed. To ensure that the linear projection of the signal preserves its original structure, the projection matrix must satisfy the Restricted Isometry Property (RIP) condition. Then, the linear projection measurement of the original signal is obtained by multiplying the original signal and the measurement matrix. The RIP condition is defined as follows: if there exists a constant δ... K For all signals θ of sparsity K ∈ (0, 1), matrix A satisfies the following equation:

[0066]

[0067] Then matrix A is said to satisfy the constrained isometry property of order K, where sparsity K refers to the number of non-zero elements in the signal θ. A is N t A compressed dimension reduction matrix with row 1 and column 1. The advantage of compressed sensing technology is that even N... t >l, (l refers to the length of the signal), can still be obtained from N r (N r =N t The original data of length l is recovered from the measurements. Let... This represents the compression ratio. According to the principle of compressed sensing, as long as the measurement matrix A satisfies the RIP condition, even if A is a matrix with a number of rows much smaller than its number of columns, projecting the signal θ onto a space with a reduced dimension, the original signal can still be completely recovered from the number of measurements much smaller than the signal dimension using the compressed sensing reconstruction algorithm. The compression ratio ρ determines the number of transmit and receive antennas that can be reduced and the reconstruction performance of the receiver. M. Davenport, in Theorem 3.5 of his doctoral dissertation "Random observation on random observations: Sparse signal acquisition and processing," states that A satisfies the 2K-order RIP constant δ. 2K A matrix ∈ (0, 1] only needs to If C is a constant approximately equal to 0.28, the original signal can be recovered. Meanwhile, Donoho, in his paper "Extensions of compressed sensing," *Signal Processing*, vol. 86, no. 3, pp. 533-548, 2006, gives three necessary conditions for an observation matrix, pointing out that most uniformly distributed random matrices possess these three conditions and can be used as observation matrices, such as partial Fourier sets, partial Hadamard sets, and uniformly distributed random projection sets. His papers "ecoding by linear programming," *IEEE Transactions on Information Theory*, vol. 51, no. 12, pp. 4201-4215, 2005, and "Stable signal recovery from incomplete and inaccurate measurements," *Communications on Pure and Applied Mathematics*, vol. 59, no. 8, pp. 1207-1223, 2006, prove that when the measurement matrix A is a Gaussian random matrix, A can satisfy the RIP condition with a relatively high probability. Therefore, this application selects a Gaussian random matrix as the compressed sensing multiplexing matrix A.

[0068] Regarding the "signal reconstruction at receiver 2": the underdetermined problem can be solved using reconstruction algorithms in compressed sensing. Solving this problem requires exhaustively enumerating all possible permutations of non-zero values ​​in the sparse vector θ, making it difficult. Therefore, researchers have proposed a series of algorithms to obtain suboptimal solutions, mainly including: greedy tracking algorithm, convex relaxation method, Bayesian algorithm, and combinatorial algorithm. Each algorithm has its advantages and disadvantages. The convex relaxation method requires the fewest observations to reconstruct the signal, but often has a heavy computational burden. The greedy tracking algorithm falls between these algorithms in terms of running time and sampling efficiency, but its noise resistance is unstable. A suitable reconstruction algorithm can be selected according to different environments; once the sparse representation vector is obtained, the original signal can be recovered.

[0069] The traditional MIMO signal transmission process is as follows: The transmitted data stream 's' undergoes space-time coding, digital-to-analog conversion, and analog module processing, and is separated into Nt sub-data streams, which are simultaneously transmitted at the same frequency through Nt transmit antennas. The transmitted signals propagate through the wireless channel via reflection and scattering, and these parallel sub-signals arrive at the receiver 2 at different times along different paths, where they are received by Nr antennas. The receiver 2 uses signal processing technology to jointly process the signals received by each antenna, thereby recovering the original data stream. The codewords are modulated and transmitted, and the receiver 2 performs signal detection to reconstruct the original signal.

[0070] Compared to the traditional MIMO signal transmission process described above, we propose a signal transmission and detection scheme based on a combination of deep learning and compressed sensing technologies. For example... Figure 1 As shown, this application adds a compression and multiplexing module 104 to the transmitting end 1 of the device. First, the modulated signal is compressed and reduced in dimensionality. The selection of the compression and dimensionality reduction matrix does not require channel state information. The measurement matrix in compressed sensing technology can be selected as the signal compression matrix to complete the compression, dimensionality reduction and multiplexing of the transmitted signal, thereby reducing the amount of data. The receiving end 2 of this application reconstructs the signal in the following two steps: (1) The first neural network model (NN1) is trained by the BP (back propagation) algorithm in deep learning to solve the high-dimensional sparse signal θ from the low-dimensional received signal. (2) The second neural network model (NN2) is trained by the BP algorithm in deep learning to reconstruct the original signal x.

[0071] Backpropagation (BP) neural networks, a concept proposed in 1986 by scientists led by Rumelhart and McClelland, are multi-layer feedforward neural networks trained using an error backpropagation algorithm. The BP algorithm uses the squared network error as the objective function and employs gradient descent to calculate its minimum. A BP network adds several layers (one or more) of neurons between the input and output layers; these neurons are called hidden units. They have no direct connection to the outside world, but changes in their states affect the relationship between input and output. Each layer can have several nodes. The computation process of a BP neural network consists of forward and backward computation. In the forward propagation process, the input pattern is processed layer by layer from the input layer through the hidden unit layers and then forward to the output layer. The state of each neuron only affects the state of the next layer's neurons. If the desired output cannot be obtained at the output layer, backpropagation is initiated, returning the error signal along the original connection path. By modifying the weights of each neuron, the error signal is minimized. The aforementioned "backpropagation" is an algorithm used in neural networks to efficiently compute gradients, or more generally, a feedforward computational graph. It can be summarized as applying a chain rule of differentiation starting from the network output and then propagating the gradient backward. The first application of backpropagation can be traced back to Vapnik et al. in the 1960s, but the paper "Learning representations by back-propagating errors" is frequently cited. Currently, in practical applications of artificial neural networks, the vast majority of neural network models use BP networks and their variations. It is also the core component of feedforward networks, embodying the essence of artificial neural networks.

[0072] Backpropagation (BP) networks are mainly used in the following four aspects: 1) Function approximation: training a network to approximate a function using an input vector and a corresponding output vector; 2) Pattern recognition: associating a given output vector with an input vector; 3) Classification: classifying input vectors according to a suitable method; 4) Data compression: reducing the dimensionality of the output vector for easier transmission or storage. Here, this application uses the BP algorithm to train a neural network to achieve function approximation.

[0073] This invention, based on existing MIMO system signal multiplexing technologies, improves the multiplexing gain and capacity of the MIMO system by introducing a compression multiplexing module 104 and a demultiplexing module at the receiving end 2, compared to traditional MIMO schemes. This reduces the required number of antennas while simultaneously transmitting the same amount of data. Compared to existing MIMO spatial multiplexing technologies, this application focuses not only on eliminating interference between adjacent data, but also on how to multiplex more data streams to the receiving end 2 under a given number of transmitting antennas, while ensuring the detection performance of the receiving end 2, thereby achieving multiplexing gain and transmission capacity exceeding the inherent multiplexing gain of the MIMO system.

[0074] Example

[0075] This embodiment integrates the MIMO multi-antenna signal transmission and detection technology based on deep learning and compressed sensing proposed in this application, and provides a detailed example of the specific implementation steps of the present invention.

[0076] First, the information source is generated by using random number generator 2 to produce a sequence of 0 and 1 bits.

[0077] Modulation is the modulation of bit data, including BPSK, QPSK, 16QAM and 64QAM.

[0078] This embodiment uses BPSK modulation as an example for illustration.

[0079] According to such Figure 2 The signal processing flow at the transmitting end 1 and receiving end 2 of the device shown is as follows:

[0080] S1: Signal processing at device transmitter 1.

[0081] S11: Use random number generator 101 to generate a sequence of 0 and 1 bits to form the original data.

[0082] S12: After BPSK modulation, signal x is generated. The transmitted data for each group can be different.

[0083] S13: Compression and multiplexing processing, i.e., multiplying the sparse vector transmitted data by the compression and multiplexing matrix A to obtain the data vector z. A represents the measurement matrix in compressed sensing, which is chosen as N here. t A Gaussian matrix of size ×l. Wherein, This represents the compression ratio, indicating the percentage by which the number of antennas is reduced. Finally, the data is transmitted through the channel.

[0084] S2: Signal detection at receiver 2 of the device.

[0085] S21: The received signal is y = Hz + n = HAx + n, where n represents noise. The signal matrix H is estimated using the channel estimation module 201.

[0086] S22: Use the backpropagation (BP) training algorithm of deep learning in advance to obtain the first neural network model and the second neural network model.

[0087] Specifically, when the transmitted signal is x, the signal at receiver 2 is y. Theoretically, we can always find a suitable basis to achieve a sparse representation of the signal. Temlyakov, in his paper *Nonlinear Methods of Approximation*, *IMI Research Reports*, *Dept. of Mathematics*, *University of South Carolina*, 2001, points out that the choice of dictionary D should conform as well as possible to the structure of the approximated signal, and its composition can be unrestricted. Here, all possible combinations of x are used as different column vectors of the redundant dictionary D to form a sparse representation of x: a sparse vector θ where all positions except the corresponding index position are 1. Using y and θ, the first set of training samples is formed; θ and x form the second set of training samples. Through the backpropagation (BP) training algorithm of deep learning, the first set of training samples yields the first neural network model, and the second set of training samples yields the second neural network model. When the input is signal y, the reconstructed transmitted data is obtained.

[0088] S23: The first neural network signal processing module 202 uses its constructed first neural network model to extract a high-dimensional sparse signal θ from the low-dimensional target signal. The input signal of the first neural network model is y, and the output signal is the sparse representation θ of the original signal x on the overcomplete redundant dictionary D.

[0089] S24: The second neural network signal processing module reconstructs the original signal x by inputting the sparse signal θ into its constructed second neural network model. The input signal of the second neural network model is the sparse representation θ of the original signal x on the overcomplete redundant dictionary D, and the output signal is the original transmitted data stream.

[0090] The following further explains steps S22 to S24: The neural network model consists of three parts: an input layer (layer1), intermediate layers (layer2, ..., L-1), and an output layer (layerL). The input layer is responsible for signal transmission, receiving external input information, and each unit of the input layer represents a feature. The intermediate layers can be single or multiple, and they are responsible for internal information processing and information transformation. The output layer is responsible for outputting information to the outside, and each unit of the output layer represents a category. In this application, a BP neural network is used to simulate a mapping function that maps the input space data to the output space. The BP neural network will try to fit the function between the original device receiver signal y and the original signal x in the sparse representation θ of the overcomplete redundant dictionary D, as well as the sparse representation θ and the original transmitted data stream. The function between them: Based on the mapping function model generated by the trained BP neural network, it can restore the sparse representation θ of the theoretically calculated original signal x on the overcomplete redundant dictionary D according to the signal y received by the receiver 2 of the device, and reconstruct the original transmitted data x according to the sparse representation θ calculated above.

[0091] This application uses a cost function to measure the difference between the output of the BP neural network and the true output. The BP neural network is trained so that the output of the network's input (the signal received by receiver 2) is as close as possible to the theoretical output. To minimize the cost function, gradient descent is used to solve for the neural network parameters. Once the optimal neural network weights are found, a first or second neural network model is established. The BP neural network is created by collecting a large amount of sample data and manually labeling the correct classification results. This labeled data is then used to train the created neural network. During this process, each layer in the neural network continuously adjusts its weights and biases based on the difference between the current output value and the correctly labeled target value until it can accurately output the target value.

[0092] Further explanation regarding the two parameters required for training a neural network—weights and biases: In this application, the weight parameter matrix between each layer of the neural network is used... This indicates that the superscript of the weight parameter w represents the layer number, and the subscript represents the nth node in each of the two adjacent layers. For example, This represents the weight of the line segment between the first node of Layer 1 and the second node of Layer 2. These weights determine the model's function; the goal of the neural network is to calculate these weights using samples. Each node in the intermediate and output layers represents a Logistic function g(z) = a. For example... This represents the input value of the first node in Layer 2, which is then substituted into the Logistic function to obtain the output.

[0093] The bias parameter matrix between each layer of the neural network is: B = [b1b2...b n ] T Given that the input to the neural network is: Y = [y1y2...y...] n ] T The output of the neural network is: Introduce a nonlinear operator: Then we can deduce that:

[0094]

[0095] Initialize weight parameters: Randomly initialize the weight parameters w to a value between [-ε, ε], where ε is a preset, sufficiently small value.

[0096] Training a neural network model: The process of training a neural network model mainly consists of two steps: first, calculating the cost function J(θ); and second, adjusting the parameters θ to minimize the cost function value J(θ). The forward propagation algorithm is used below to calculate the output of each sample under the current neural network model, obtain the cost function, and then update the weight parameters based on the output. The cost function J(θ) is defined, 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 as follows:

[0097]

[0098] The backpropagation algorithm is used to adjust the parameter θ to minimize the cost function value J(θ). The backpropagation algorithm updates the weight coefficients by calculating the partial derivatives of the cost function with respect to each weight coefficient. For example, first, the gradient of the last layer is calculated: (1) the gradient of the cost function value with respect to the nonlinear operator, (2) the gradient of the neural network output with respect to the bias and the weights between adjacent layers. The gradient is updated in the negative direction of the gradient. Second, the gradient of the penultimate layer is calculated: (1) the gradient of the backpropagation error of the upper layer with respect to the nonlinear operator, (2) the gradient of Hn-1 (H is the output of each layer after the activation function) with respect to the bias and the weights between adjacent layers, and the gradient is updated in the negative direction of the gradient. Finally, after backpropagation layer by layer, the gradient of the first layer is calculated: (1) the gradient of the backpropagation error of the second layer with respect to the nonlinear operator, (2) the gradient of H1 with respect to the bias and the weights between adjacent layers. The gradient is updated in the negative direction of the gradient. Thus, after the first backpropagation process is completed, the above two steps of forward propagation to obtain the output and backpropagation to update the parameters are continued until the mean square error is minimized, which completes the training process of the neural network model.

[0099] Figure 3The bit error rate (BER) performance of a MIMO system with different transmit and receive antenna configurations after employing compressed sensing and neural network signal transmission and detection techniques is presented. A flat fading channel is assumed here. In the conventional scheme, only four transmit antennas can transmit four data symbols simultaneously. Using the scheme of this application, the signal is first modulated using BPSK to obtain the original signal x. 4×1 A random Gaussian matrix 4ρ×4 As a compression and dimensionality reduction matrix, z is obtained. If ρ = 0.5, only 2 transmitting antennas are needed to transmit the original data. If ρ = 0.75, only 3 transmitting antennas are needed. The receiving end 2 of the device, also using 2 or 3 receiving antennas, employs models 1 and 2 trained by the neural network to obtain the reconstructed transmitted signal. The bit error rate performance of the proposed scheme is shown in Figure (2×2)-4. The first number in parentheses represents the number of transmit antennas, the second number represents the number of receive antennas, and the last number represents the original data length. With an increased number of receive antennas, this scheme is denoted as (3×3)-4. Compared to the traditional signal detection algorithm for MIMO systems—zero-forcing (ZF)—the proposed scheme can reduce the required number of transmit and receive antennas while maintaining a high SNR.

[0100] The zero-forcing detection algorithm described above uses a filter matrix WZF multiplied by the received signal y to eliminate interference between transmitted signals, thereby estimating each transmitted symbol. The filter matrix is: W ZF =H -1 =(H H H) -1 H H Therefore, the estimated signal vector is: After obtaining the estimated signal vector, it is mapped onto the constellation point with the nearest Euclidean distance in the constellation diagram. This constellation point is the optimal solution, thus recovering the final symbol vector XZF.

[0101] Figure 4 The bit error rate (BER) performance of a MIMO system with different transmit and receive antenna configurations after employing compressed sensing and neural network signal transmission and detection techniques is presented. A flat fading channel is assumed here. In the conventional scheme, 20 transmit antennas can only transmit 20 data symbols simultaneously. Using the scheme of this application, BPSK modulation of the signal is first employed, and then x... 20×1 Divided into 5 groups, the vector x in each group i The length of the sequence (i = 1, 2, 3, 4, 5) is 4. The random Gaussian matrix A... 4ρ×4 As a compression and dimensionality reduction matrix, z is obtained. i z i Concatenating them yields the vector to be sent. If ρ = 0.5, only 10 transmit antennas are needed to transmit the original data. If ρ = 0.75, 15 transmit antennas are needed. Receiver 2 also uses 10 receive antennas. The bit error rate performance of this scheme is shown in Figure (10×10)-20. The first number in parentheses represents the number of transmit antennas, the second number represents the number of receive antennas, and the last number represents the length of the original data. For the same data and the same number of packets, ρ = 0.75, the compression and dimensionality reduction matrix is ​​A. 3×4 At this point, the number of receiving antennas increases, and this scheme is denoted as (15×15)-20. Compared with the classic detection algorithm ZF (Zero Forcing), our proposed scheme can reduce the number of transceiver antennas required while maintaining a high SNR. Therefore, the scheme proposed in this application can reduce the number of transceiver antennas required while maintaining a high SNR.

[0102] As described above, the enhanced spatial multiplexing method proposed in this invention can, based on existing MIMO systems, combine neural network technology to transmit the same amount of data while reducing the number of antennas, thereby reducing the required number of antennas and improving multiplexing gain and system capacity.

[0103] It should be noted that the specific embodiments described above are exemplary, and those skilled in the art can devise various solutions inspired by the disclosure of this invention. These solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents.

Claims

1. A method for reconstructing the original signal of an antenna signal, characterized in that, The method includes at least: Using the backpropagation (BP) training algorithm of deep learning, the first neural network model and the second neural network model are obtained. The constructed first neural network model is used to extract high-dimensional sparse signals from low-dimensional target signals. ; By the sparse signal The original signal is reconstructed by inputting the constructed second neural network model; The steps for obtaining the first and second neural network models using the backpropagation (BP) training algorithm in deep learning include: use and This constitutes the first set of training samples. This indicates the signal received by the original device receiver. use and This constitutes the second set of training samples; Using the backpropagation (BP) training algorithm in deep learning, the first set of training samples yields the first neural network model, and the second set of training samples yields the second neural network model. The weight parameter matrices between the layers of the neural network are used... This indicates that the weight parameters are... The superscript indicates the layer number, and the subscript indicates the nth node of each of two adjacent layers; The bias parameter matrix between each layer of the neural network is as follows: The input to the neural network is known to be: The output of the neural network is: By introducing a nonlinear operator, we can derive: ;in, Represents the bias vector of each node; The dimension is The input vector; The dimension is The output vector; F represents the activation function; This represents the weight parameter between the k-th node and the i-th node; Initialize weight parameters: Set the weight parameters Randomly initialized to a value between [-ε, ε], where ε is a preset value; When the input signal is a signal At that time, the original transmitted data stream of the reconstructed output is obtained. ; This represents the input signal of the first neural network model. Indicates a sparse signal; Represents the transmitted signal vector. This indicates the original data stream sent as output.

2. The method for reconstructing the original antenna signal according to claim 1, characterized in that, The method further includes: Channel estimation is performed on the low-dimensional target signal, and the resulting channel parameter matrix is ​​used as the input to the first neural network model.

3. The method for reconstructing the original antenna signal according to claim 1, characterized in that, The method further includes: Send signal vector The redundant dictionary D is constructed by using all possible combinations as different column vectors of the redundant dictionary, thus realizing the transmission signal vector. sparse signal .

4. The method for reconstructing the original antenna signal according to claim 1, characterized in that, The process of training a neural network model mainly includes: Calculate the cost function J(θ). Adjust the parameter θ to minimize the cost function value J(θ). The cost function J(θ) is obtained by using the forward propagation algorithm to calculate the output of each sample under the current neural network model.

5. The method for reconstructing the original antenna signal according to claim 4, characterized in that, The backpropagation algorithm is used to adjust the parameter θ so that the cost function value J(θ) is minimized. The backpropagation algorithm updates each weight coefficient by calculating the partial derivative of the cost function J(θ) with respect to each weight coefficient.

6. A device for reconstructing the original signal of an antenna signal, characterized in that, It includes at least a first neural network signal processing module (202) and a second neural network signal processing module (203). The second neural network signal processing module (203) extracts the high-dimensional sparse signal from the low-dimensional target signal by using the first neural network model constructed by the first neural network signal processing module (202). The original signal is reconstructed by inputting the second neural network model it constructed. The first neural network signal processing module (202) utilizes and This constitutes the first set of training samples. This indicates the signal received by the original device receiver. The second neural network signal processing module (203) utilizes and This constitutes the second set of training samples; Using the backpropagation (BP) training algorithm in deep learning, the first set of training samples yields the first neural network model, and the second set of training samples yields the second neural network model. The weight parameter matrices between the layers of the neural network are used... This indicates that the weight parameters are... The superscript indicates the layer number, and the subscript indicates the nth node of each of two adjacent layers; The bias parameter matrix between each layer of the neural network is as follows: The input to the neural network is known to be: The output of the neural network is: By introducing a nonlinear operator, we can derive: ;in, Represents the bias vector of each node; The dimension is The input vector; The dimension is The output vector; F represents the activation function; This represents the weight parameter between the k-th node and the i-th node; Initialize weight parameters: Set the weight parameters Randomly initialized to a value between [-ε, ε], where ε is a preset value; When the input signal is a signal At that time, the original transmitted data stream of the reconstructed output is obtained. ; This represents the input signal of the first neural network model. Indicates a sparse signal; Represents the transmitted signal vector. This indicates the original data stream sent as output.

7. The original signal reconstruction apparatus for antenna signals according to claim 6, characterized in that, It also includes a channel estimation module (201), which is configured to perform channel estimation based on the received low-dimensional target signal after compression and dimensionality reduction and use the resulting channel parameter matrix as input to the first neural network model.

8. The original signal reconstruction apparatus for antenna signals according to claim 6 or 7, characterized in that, The training process of the neural network model by the first neural network signal processing module (202) and the second neural network signal processing module (203) mainly includes: Calculate the cost function J(θ). Adjust the parameter θ to minimize the cost function value J(θ). The cost function J(θ) is obtained by using the forward propagation algorithm to calculate the output of each sample under the current neural network model.

Images