CN116054887A - Antenna signal modulation method based on neural network model - Google Patents

Abstract

The invention relates to an antenna signal modulation method based on a neural network model, which is characterized by at least comprising the following steps: creating a neural network by a BP neural network training method in artificial intelligence, respectively taking a transmission signal vector x of a device transmitting end (1) and a sparse signal theta determined based on the transmission signal vector x of the device transmitting end (1) as input and output, constructing a first training sample set, and training the neural network to obtain a first neural network model; performing compression dimension reduction processing on the sparse representation signals; low-dimensional signal from compressed sensing reconstruction algorithmObtaining sparse signal vector by medium solution

And finally, reconstructing the received signal by using the second neural network model to obtain an original signal. The invention reduces the dimension of the signal to be processed, multiplexes the parallel data streams exceeding the number of the transmitting antennas to the given transmitting antennas and transmits the parallel data streams, and greatly improves the signal multiplexing gain under the condition of receiving and transmitting the number of the antennas of the given MIMO system.

Description

Antenna signal modulation method based on neural network model

The original basis of the divisional application is a patent application with application number (202080000961.6), application date of 2020, 4 and 7 days and the invention name of 'MIMO multi-antenna signal transmission and detection technology based on artificial intelligence'.

Technical Field

The invention relates to the technical field of mobile communication, in particular to an antenna signal modulation method based on a neural network model.

Background

Communication technology development can be considered as a common wealth for all people, but the global communication standard brought by the development is not just a technical standard, but is related to industrial development and national strategy.

MIMO (Multiple-Input Multiple-Output) is also called as a Multiple-Input Multiple-Output system, and refers to a communication system that uses Multiple antennas at the transmitting end and the receiving end of the device at the same time, so that the capacity and the spectrum utilization rate of the communication system can be improved doubly without increasing the bandwidth. The MIMO system adopts multiple antennas at the transmitting end and the receiving end of the device, the transmission information stream is formed into a plurality of information substreams through space-time coding, the plurality of substreams are simultaneously transmitted to the channel, and each transmission signal occupies the same frequency band, so that the system bandwidth is not increased. If the channel responses between the transmitting and receiving antennas are independent, the mimo system can create multiple parallel spatial channels through which the independent transmission information is transmitted, and the data rate must be improved.

MIMO technology is a great breakthrough in the field of wireless communication, and in 2011, many companies have developed WIFI or WIMAX commercial systems based on MIMO technology. In 2012, MIMO technology was selected as one of the key technologies for all standards (e.g., TD.LTE, LET.A, WIMAX, etc.) of the 4G communication system. MIMO systems have been widely used in wireless communications-mobile devices and networks generally use multiple antennas to enhance connectivity, improve network speed, and user experience. Massive MIMO is also a key factor of 5G ultra-high data rate, and can bring about larger network capacity, wider signal coverage and better user experience, and bring the potential of 5G into a brand new level.

From methodology, the goal of 4G to 5G upgrades are enormous, 4G considers that the connection people and people emphasize increasing the transmission rate, while 5G has jumped out of the previous thought limitations and thus considers the interconnection of people and things, not only emphasizing the increase of the transmission rate, but also considering ultra-large connections and ultra-low latency. From the social category, 4G changes our lives, while 5G changes our social structure. For three KPIs of 5G, namely ultra-high bandwidth, ultra-large connection and ultra-low time delay, the existing network architecture needs to be upgraded for achieving the purposes, and intelligent adjustment is carried out according to requirements, so that artificial intelligence (Artificial Intelligence, AI) can energize the 5G. In turn, 5G technology may bring more application requirements, which in turn accelerates the development of AI.

To date, artificial intelligence has not been well-defined and unified. One definition of textbook style is "AI is a computer program that makes reasonable actions based on the perception of the environment and obtains the greatest benefit", and another more technically oriented definition is "AI is a computer algorithm that will learn". From a technical application perspective, and in particular from a communication technology development perspective, the academic community is more inclined to the second definition, then artificial intelligence can be technically equivalent to machine learning. Machine learning is essentially a key technology to extract knowledge from data, and is the power and engine of artificial intelligence development. Machine learning is currently mainly used for solving classification problems, clustering problems, regression problems and the like, and is widely applied to the fields of character recognition, machine translation, voice recognition, search engines, face recognition, unmanned driving and the like. At present, the size of the learning algorithm is hundreds, and a unified framework for describing the design process of the machine learning algorithm does not exist, but at present, the machine learning has two strongest algorithms, namely deep learning and reinforcement learning.

Deep learning belongs to an artificial neural network system, and is a new generation neural network under the development of a traditional neural network. Neural networks are an intelligent science that mimics the human neural, brain architecture and function. The system has the quick response capability, and is convenient for real-time control and processing of things; excellent self-organizing and self-learning capabilities; is good at being used in complex environments. Fully approximates to any nonlinear system, and rapidly obtains an optimized answer meeting various constraint condition questions; has excellent performance such as high robustness and fault tolerance, and thus has been widely used in communication systems. Neurosurgery is a computer system formed by a plurality of very simple processing units interconnected in some way with each other, which processes information by its state without a dynamic response to externally input information.

The problem of MIMO is solved by applying the neural network, and patent achievements are not invented at home and abroad. The Chinese application number 201510473741.6 (self-complex neural network channel prediction method, western electronic technology university) discloses a complex neural network channel prediction method, which mainly solves the problem of channel fading caused by channel time variation in a MIMO system. The technical scheme is that 1, a base station measures a channel to obtain a channel coefficient training sequence containing estimation errors; 2. obtaining corresponding training samples and expected output according to the obtained channel coefficient sequence; 3. inputting a training sample to perform complex wavelet neural network training to obtain a final network weight; 4. and the base station predicts the channel coefficient by using the trained complex wavelet neural network. The method is simple and easy to implement, has good effect, and is suitable for reducing the influence of channel time variation on the MIMO system channel.

The Chinese application number 201810177829.7 patent (a neural network-based wireless channel modeling method, university of southwest) discloses a neural network-based wireless channel modeling method. Firstly, processing a received signal fed back by a user to obtain estimated channel parameters; and then three-dimensional geographic information of the scatterer is obtained according to the two-dimensional image, the three-dimensional geographic information is clustered, and finally, channel parameters and the geographic information are used as input of the neural network, a received signal is used as output, and a nonlinear time-varying neural network model is obtained through training. The method can obtain a more accurate channel model within acceptable complexity, and can meet the channel modeling requirements of scenes such as large-scale MIMO technology, band expansion, high mobility and the like adopted in a future 5G communication system.

The Chinese application number 201810267976.3 patent (a deep neural network large-scale MIMO system detection method constructed based on BP algorithm, southeast university) provides a deep neural network large-scale MIMO system detection method constructed based on BP algorithm, which constructs a deep neural network for large-scale MIMO system detection by expanding and mapping a belief propagation iterative algorithm factor graph onto a neural network structure; the neurons of the deep neural network correspond to nodes in the iterative algorithm factor graph, and the number of the neurons of each layer is equal to the number of symbol nodes in the iterative algorithm factor graph; the mapping function between the hidden layers is an updating formula of the confidence information in the iterative algorithm, and the number of the hidden layers is equal to the iteration times of the iterative algorithm. Specifically, the invention also provides a MIMO detection method for respectively constructing two deep neural networks based on the damping belief propagation and the maximum and belief propagation iterative algorithm. The invention achieves lower error rate without increasing on-line operation complexity, and has robustness to various channel conditions and antenna matching.

The invention of China application number 201910063733.2 (optimized MIMO detection method based on deep learning, shanghai university) is characterized in that by constructing an MIMO end-to-end transmission model, obtaining complex time domain vectors of the model according to signals y (t) received by a receiving end of an MIMO device and estimated imperfect channel state information as input of a Deep Neural Network (DNN), obtaining an estimated value of a bit stream at a transmitting end of the device by using the DNN, and comparing the estimated value of the bit stream at the transmitting end with an estimated value of the bit stream obtained according to hard decision in the prior art, the invention can improve accuracy and detection rate under imperfect channel information, ensure detection performance of low bit error rate under low complexity algorithm, and has good robustness under the condition of containing inherent channel errors.

Chinese application No. 201610327115.0 (a codebook selection method based on deep learning under massive MIMO, chongqing postal university) relates to a codebook selection method based on deep learning under massive MIMO. The method comprises the following steps: collecting pilot frequency information of a test area to construct a pilot frequency training sequence, and further obtaining a pilot frequency training sample; performing neural network iterative learning on the pilot training sample to obtain a final network weight value; and selecting the optimal code word from the complete codebook according to the channel output by the neural network after learning. And then, carrying out channel information matching on the unknown area and the test area to obtain a wireless channel of the unknown area, and further obtaining a codeword corresponding to the wireless channel. The invention can effectively, accurately and quickly establish the wireless channel model and the codebook inquiry, avoids the channel estimation of the unknown region and greatly reduces the complexity of the channel selection codebook of the unknown region.

The Chinese application No. 201811626005.X patent (Low complexity MIMO-NOMA system signal detection method based on improved gradient projection method, chongqing university) discloses a low complexity MIMO-NOMA system signal detection method based on improved gradient projection method, which relates to wireless communication technology. According to the sparse characteristics of active users of the system, converting the system model into a strict quadratic programming problem by utilizing a convex optimization algorithm idea; and then carrying out iterative solution on the problem, and carrying out preprocessing operation on the iterative result of each time to achieve effective detection on active users and signals thereof. The invention breaks through the problem of low algorithm convergence speed in the traditional detection method, and carries out preprocessing operation on each iteration result, so that the detection result can be converged rapidly, and the active user set can be detected, thereby having simple implementation process and wide application range.

The Chinese application number 201910014714.0 (a method for designing a MIMO system beamforming matrix based on deep learning, nanjing university of post) discloses a method for designing a MIMO system beamforming matrix based on deep learning, which comprises the following steps of firstly obtaining a training sample set required by a deep learning network by using a known algorithm; then constructing a deep learning neural network model, initializing relevant parameters of the model and training by using a training sample set; and then, the pilot frequency is used for obtaining a channel and sending the channel into a neural network to predict the beamforming matrix coefficient, and finally, the channel and the beamforming matrix coefficient are combined to form a beamforming matrix. The method utilizes the beam forming matrix obtained by the deep learning neural network to simultaneously consider the performance and the algorithm complexity, and can reduce the time delay on the premise of ensuring the performance, so that the MIMO system can provide real-time service.

The Chinese application number 201810182937.3 (a machine learning-based MIMO link adaptive transmission method, dongnan university) discloses a machine learning-based MIMO link adaptive transmission method, which uses an unsupervised learning self-coding algorithm to extract and reduce the dimension of features, introduces the idea of deep learning, and can reduce the feature dimension and the computation complexity on the premise of keeping the state information of main information. The invention utilizes the logistic regression algorithm to construct the mapping relation between the channel state information and the transmission parameters, is different from the traditional fixed parameterized model, can train based on sample data, can better establish the mapping relation between the channel state information and the transmission parameters under the condition that the quality of a data set is better and all states are covered, and can more fully utilize the channel state information compared with the traditional single equivalent signal-to-noise ratio. In addition, the invention also carries out CQI selection based on the channel matrix, and the MIMO link self-adaptive method based on machine learning is not limited by the design of a receiver through the research of the channel matrix and the noise variance, thereby having universality.

The chinese application No. 201710495044.X patent (a method for joint precoding and antenna selection of MIMO system based on deep learning, university of zhejiang technology) discloses a method for joint precoding and antenna selection of MIMO system based on deep learning, comprising the following steps: firstly, generating a training data set required by deep learning through an existing antenna selection method; then, establishing a deep learning model, training the deep learning model by using training data and storing the deep learning model; then, completing antenna selection by using the stored deep learning model; and finally, carrying out optimal precoding design on the selected MIMO subsystem. The invention designs the MIMO system joint precoding and antenna selection by using the deep learning technology, and can realize lower calculation complexity under the condition of obtaining good system signal-to-noise ratio.

The invention patent of China application number 201910242525.9 (a high-speed rail oriented depth signal detection method, shenzhen university) proposes a high-speed rail oriented depth signal detection method, which comprises the steps of firstly, collecting data, and collecting a plurality of sending signals and receiving signals in various scenes along the high-speed rail according to different environment types along the high-speed rail; secondly, dividing scenes, and further dividing each scene into a plurality of areas through data analysis to meet the compatibility of the neural network; thirdly, establishing a deep high-speed rail signal detection neural network model; then, training a high-speed rail signal detection neural network offline; and finally, carrying out online real-time signal detection, determining the position information of the high-speed rail through a GPS (global positioning system) in the running process, judging the area where the high-speed rail is positioned, selecting a corresponding neural network model, inputting the real-time received signal into a trained neural network, and outputting the signal sent by the base station end in real time. The system performance of the invention is greatly improved, the bit error rate of signal detection is reduced, and the algorithm is more robust. The method used by the invention does not need to estimate the channel, and saves pilot frequency overhead.

The invention patent of China application No. 201810279530.2 (visible light communication MIMO anti-interference noise reduction method based on BP neural network, university of national university of middle and south) discloses a visible light communication MIMO anti-interference noise reduction method based on BP neural network, and relates to MIMO antenna technology in the field of visible light communication. The system comprises a system transmitting end, a system device receiving end signal processing part and a BP neural network signal processing part which are sequentially communicated. The method comprises the following steps: 1) The electric signal is loaded on the LED array and emitted out in the form of an optical signal; 2) The photoelectric detector at the receiving end of the device converts the optical signal into an electric signal; 3) The multipath electric signals remove high-frequency interference through a low-pass filter; 4) After training, the BP neural network carries out noise reduction and interference elimination processing on the multipath signals, and finally the multipath signals are converted into binary serial data streams through parallel-serial conversion. The invention improves the transmission performance of the existing MIMO technology; combining the neural network with the visible light MIMO technology, and exerting the advantages of the neural network in the aspect of noise reduction and disturbance elimination in wireless communication; the adoption of the neural network receiving processing technology enables the whole VLC system to be more stable.

The invention patent of China application number 201710213235.2 (a visible light channel joint equalization method based on orthogonal mapping and a probabilistic neural network, university of Zhongshan) discloses a visible light channel joint equalization method based on the orthogonal mapping and the probabilistic neural network, which comprises a device transmitting end and a device receiving end, wherein signals are transmitted from the transmitting end to the device receiving end through a visible light MIMO channel; the visible light MINO channel is a multiple-input multiple-output channel; the combined equalization is that the pre-equalization and the post-equalization are combined; the invention adopts a joint equalization scheme combining a front equalization technology and a rear equalization technology, namely a visible light multi-input multi-output channel joint equalization method based on orthogonal mapping and a probability neural network, which can effectively inhibit interference between channels of a visible light MIMO communication system and improve data transmission reliability.

The embodiment of China 201910125325.5 patent (a MIMO decoding method, device and storage medium based on deep learning, shenzhen Bao-chain artificial intelligence technology Co., ltd.) discloses a MIMO decoding method, device and storage medium based on deep learning, wherein a training data set of MIMO decoding is constructed, and the training data set comprises a plurality of training data; training the neural network based on the training data set to obtain a trained neural network model; when receiving the MIMO signal to be decoded, inputting the MIMO signal to be decoded into the neural network model for MIMO decoding, and then obtaining an MIMO decoding result output by the neural network model. Through the implementation of the invention, a neural network model for joint MIMO detection and channel decoding is designed based on deep learning, the MIMO detection and the channel decoding are regarded as a joint decoding process, the approximation of the output result of the neural network model is improved through training, the overall performance of MIMO decoding is ensured, and the decoding accuracy and the decoding speed are higher.

The Chinese application number 201810757547.4 patent (a machine learning assisted massive MIMO downlink user scheduling method, university of southward) discloses a machine learning assisted massive MIMO downlink user scheduling method, comprising the following steps: s1: the base station acquires a characteristic mode energy coupling matrix in a characteristic direction through an uplink detection signal sent by a user; s2: the base station utilizes the characteristic mode energy coupling matrix to assist in sum rate calculation under various user and beam combinations by a machine learning method; s3: and adopting a greedy algorithm to realize user scheduling with the maximum speed criterion, and obtaining the optimal user beam pairing combination. The invention acquires statistical channel information through the uplink detection signal, and adopts the sum rate maximization criterion to carry out user scheduling. Under the condition that the base station only has statistical channel information, the approximate calculation of the rate is accurately realized through targeted feature extraction and the design of a neural network, the complexity of user scheduling under a large-scale antenna is greatly reduced, the performance is close to optimal, and the method has good applicability and robustness.

The invention patent number 201610353881.4 of China (a modulation recognition method under the MIMO related channel based on a machine learning algorithm, beijing university of post and telecommunications) is a modulation recognition method under the MIMO related channel based on the machine learning algorithm, and belongs to the field of communication; the method comprises the following specific steps: firstly, each data stream of a communication transmitting end is respectively coded by space time, and each codeword is respectively transmitted through Nt transmitting antennas; then, calculating a MIMO channel matrix H according to the correlation matrix of the receiving end of the device and the correlation matrix of the transmitting end; according to the MIMO channel matrix H, calculating a received signal on each receiving antenna and correcting the received signal; finally, each receiving antenna respectively performs feature extraction on the corrected signals, performs training test on the extracted feature values, and calculates a modulation recognition mode to which the sample belongs; the advantages are that: the robustness and generalization capability to the non-Gaussian channel are strong, and the modulation system identification under a more complex environment can be realized through parameter iteration; by extracting the features of the high-order moment and the high-order cumulant, the signal features have obvious difference under the condition of higher signal-to-noise ratio, and the classification of the machine learning algorithm is facilitated.

The application of MIMO technology makes space a resource that can be used to improve performance and can increase the coverage of a wireless system. MIMO technology has become one of the key technologies in the wireless communication field, and has been increasingly applied to various wireless communication systems by the continued development in recent years. As the number of antennas increases, the complexity of implementation of the MIMO technology increases greatly, so that the number of antennas used is limited, and the advantages of the MIMO technology cannot be fully exerted. Meanwhile, since the practical proposal of artificial intelligence in 1956, over 50 years, the development of the artificial intelligence is achieved, and the development of the artificial intelligence is a wide intersection and leading edge science. As an important branch of artificial intelligence, it can be seen from the above description that neural networks, which implement a mathematical network through a computer, are recognized as "out-ways" that solve some of the bottleneck problems encountered in current communications.

By integrating the research of key technologies around the neural network and communication problems at home and abroad, a plurality of research results exist at present, and viable solutions are provided for the results from the angles of MIMO channel estimation, signal detection and the like. How to improve the system capacity and reduce the algorithm complexity and the implementation complexity of the MIMO technology on the basis of ensuring certain system performance becomes a great challenge facing the industry. However, related researches around MIMO have not been conducted to develop schemes for signal transmission and detection from the point of improving the multiplexing gain of the system by combining artificial intelligence technology with a given number of transmitting antennas and receiving antennas.

Furthermore, there are differences in one aspect due to understanding to those skilled in the art; on the other hand, as the inventors studied numerous documents and patents while the present invention was made, the text is not limited to details and contents of all that are listed, but it is by no means the present invention does not have these prior art features, the present invention has all the prior art features, and the applicant remains in the background art to which the rights of the related prior art are added.

Disclosure of Invention

Aiming at the problems of improving the system capacity and reducing the algorithm complexity and the implementation complexity of the MIMO technology on the basis of ensuring certain system performance, the prior related research on the MIMO condition does not research the scheme of signal transmission and detection from the angle of improving the multiplexing gain of the system by combining the artificial intelligence technology under the condition of giving the number of transmitting antennas and receiving antennas.

According to the method, the latest research progress of an artificial intelligence technology and a compressed sensing technology is combined, and a multiplexing transmission and detection scheme of signals in a MIMO system based on the artificial intelligence technology and the compressed sensing technology is provided; on the other hand, the compression multiplexing matrix provided by the invention does not need to depend on channel state information, and the technical scheme adopted by the invention can be used for adding sparsification and compressed sensing multiplexing steps on the transmitting end of the device and adding a mature optimization reconstruction algorithm in the compressed sensing field on the receiving end of the device on the basis of not modifying the prior MIMO technical scheme, so that the compression multiplexing signal can be reconstructed, and the method has the advantages of small modification on the prior MIMO system and convenient realization.

The MIMO device based on the artificial intelligence and the compressed sensing technology mainly comprises a sparsification module and a compressed multiplexing module of a device transmitting end and a de-multiplexing module of a device receiving end. In addition, preferably, a random number generator (or called an original information bit generation module), a bit level processing module and a modulation module are also arranged in the transmitting end of the device, and a channel estimation module is also arranged in the receiving end of the device.

Preferably, as shown in fig. 5, the device transmitting end: the method comprises the steps that original data generated by a random number generator sequentially pass through a bit-level processing module and a modulation module to generate modulated signals, the modulated signals are then expressed as sparse signals based on a first neural network model by a sparsification module, the sparse signals are then subjected to compression multiplexing module, compression reduction and multiplexing are carried out on transmitting signals by the compression multiplexing module, and the compressed and multiplexed signals are transmitted by a transmitting antenna; the device receiving end: and the channel estimation module carries out channel estimation on the received signals, and the estimated channel state information is subjected to a demultiplexing module to reconstruct and output an original transmission data stream x based on the second neural network model.

The process specifically comprises the following steps:

MIMO multi-antenna signal transmission and detection device based on artificial intelligence, the device includes: the sparse module performs sparse representation on the original signal by using a first neural network model; the compression multiplexing module is used for carrying out compression dimension reduction processing on the signals after sparse representation; the device receiving end is used for processing the received signals to reconstruct the target signals, and the device further comprises a demultiplexing module, wherein the demultiplexing module is configured to solve and obtain sparse representation vectors from the low-dimensional signals by using a compressed sensing reconstruction algorithm at the device receiving end

And finally, reconstructing the received signal by using the second neural network model to obtain an original signal x.

According to a preferred embodiment, the input and output of the second neural network model are opposite to the input and output of the first neural network model for the sparsification module at the transmitting end of the device to solve the sparse signal θ to be transmitted to the compression multiplexing module for processing from the low-dimensional signal.

According to a preferred embodiment, the apparatus receiving end further comprises a channel estimation module configured to perform channel estimation on the signal received by the apparatus receiving end and use a channel parameter matrix obtained by the channel estimation module as an input of the compressed sensing reconstruction algorithm.

According to a preferred embodiment, the demultiplexing module obtains the sparse representation vector by solving based on a compressed sensing multiplexing matrix and a channel parameter matrix which do not need to rely on channel state information

A kind of electronic device.

According to a preferred embodiment, the sparsification module creates a neural network through a BP neural network training method in artificial intelligence, takes a transmission signal vector x of the device transmitting end and a sparse representation θ determined based on the transmission signal vector x of the device transmitting end as input and output respectively, constructs a first training sample set, trains the neural network, and obtains the first neural network model.

According to a preferred embodiment, the demultiplexing module composes the redundant dictionary D by taking all possible combinations of the transmitted signal vectors x as different column vectors of the redundant dictionary, respectively, to achieve a sparse representation θ of the transmitted signal vectors, i.e. x=dθ.

According to a preferred embodiment, the thinned sparse representation θ is compressed into ρl signals by the compression multiplexing module, and then the device transmits the compressed multiplexed signal z via the transmitting antenna.

According to a preferred embodiment, the compression multiplexed signal z is obtained by the formula z=aθ, where a is N _t A compressed dimension-reducing matrix of row m and column,

representing the compression ratio.

The system comprises a device transmitting end, a compression multiplexing module and a device receiving end, wherein the device transmitting end is used for transmitting a low-dimensional signal to the compression multiplexing module for processing by constructing a MIMO end-to-end transmission model, the device transmitting end is used for solving sparse signals before the transmission to the compression multiplexing module for processing by the aid of a first neural network model constructed by the device transmitting end, and a second neural network model constructed by the device receiving end and opposite to the input and output of the first neural network model of the sparse module is used for reconstructing the received signals by the device receiving end to obtain original signals.

According to a preferred embodiment, the input of the second neural network model is a sparse representation vector solved by a compressed perceptual reconstruction algorithm

Drawings

Fig. 1 is a schematic block diagram of a signal processing flow of a signal transmission and detection system of a MIMO multi-antenna system based on artificial intelligence and compressed sensing technology provided by the present invention;

FIG. 2 is a schematic block diagram of a preferred signal compression multiplexing and detection process provided by the present invention;

fig. 3 is a graph of the error rate performance of the classical detection algorithm ZF and the MIMO system multi-antenna signal transmission and detection technique of the present invention under different transceiver antenna configurations;

Fig. 4 is a graph of the error rate performance of the classical detection algorithm ZF and the MIMO system multi-antenna signal transmission and detection technique of the present invention under different transceiver antenna configurations; and

fig. 5 is a schematic diagram of module connection of a signal transmission and detection system of a MIMO multi-antenna system based on artificial intelligence and compressed sensing technology provided by the present invention.

List of reference numerals

1: device transmitting end 101: random number generator

102: bit level processing module 103: modulation module

104: sparsification module 105: compression multiplexing module

2: device receiving end 201: channel estimation module

202: demultiplexing module

Detailed Description

The present invention will be described in detail with reference to the accompanying drawings.

The MIMO device based on artificial intelligence and compressed sensing technology provided in the present application mainly comprises a thinning module 104 and a compression multiplexing module 105 of a device transmitting end 1, and a demultiplexing module 202 of a device receiving end 2. In addition, preferably, the device transmitting end 1 is further provided with a random number generator 101 (or called an original information bit generation module), a bit level processing module 102 and a modulation module 103, and the device receiving end 2 is further provided with a channel estimation module 201.

Preferably, as shown in fig. 5, the device transmitting end 1: the raw data generated by the random number generator 101 is passed through the bit-level processing module 102 and the modulation module 103 in sequence, generates a modulated signal,the modulated signal is then expressed as a sparse signal based on the first neural network model by the sparse module 104, the sparse signal is then subjected to compression multiplexing module 105, the compression multiplexing module 105 performs compression reduction and multiplexing on the transmission signal, and the compressed and multiplexed signal is transmitted by the transmitting antenna; device receiving end 2: channel estimation is performed on the received signal by the channel estimation module 201, the estimated channel state information is subjected to a demultiplexing module 202, and the original transmission data stream is reconstructed and output based on the second neural network model

As shown in fig. 2, the technical scheme adopted by the invention is a signal sparsification design, a compression multiplexing matrix design and a signal detection method based on compressed sensing in a MIMO system, and the technical scheme is as follows:

1. device transmitting end 1 signal processing:

for a MIMO communication system with one transmit antenna and one receive antenna. And the system transmitting end obtains a signal x of one path after channel coding and signal modulation. The modulated l-path signal x is mapped into a sparse signal theta through a first neural network model generated by training in advance ^m×1 . The first neural network model training algorithm adopts a BP algorithm.

2. Compression multiplexing processing of the device transmitting end 1:

sparse signal theta obtained after sparsification ^m×1 Compressed and multiplexed by the compression multiplexing module 105 into ρl signals, and then the compressed and multiplexed signal z is transmitted by the transmitting antenna. The compression process of the input signal by the compression multiplexing module 105 can be expressed as: z=aθ, where a is N _t A compressed dimension-reducing matrix of row m and column,

representing the compression ratio. The compression ratio ρ is determined by the size of the measurement matrix a in the compressed sensing technique. Preferably, a Gaussian random matrix is selected as the compressed sensing multiplexing matrix/compressed dimension-reducing matrix A in the device transmitting end 1.

3. Signal detection at the device receiving end 2:

the device receiving end 2 link is approximately the reverse process of the device transmitting end 1 link, and the signal received by the device receiving end 2 is: y=hz+n=ha θ+n. Wherein y is a received signal vector of nr×1, that is, ρl path modulation symbols after compression multiplexing, which are transmitted by n transmitting antennas and received by the device receiving end 2; n is Nr×1 Gaussian white noise vector, and the element is independent co-distributed complex Gaussian variable with mean value of 0 and variance of 1; h is the nr×nt channel propagation matrix and is a deterministic matrix that remains unchanged for a coherence time interval. The apparatus receiving end 2 may estimate the channel propagation matrix H from pilot signals inserted in the transmission data.

Assuming that the channel matrix H is gaussian distributed, it can be proved that the new matrix HA is still gaussian distributed, and meets the conditions that the compressed sensing measurement matrix needs to meet. According to the compressed sensing multiplexing matrix HA, a decompression multiplexing process is realized by solving the following optimization problems, and a transmitted sparse vector is calculated and determined: min θ ₀ s.t.y=h.a.θ. Here, a Bayes-based compressed sensing (Bayesian Compressive Sensing, BCS) reconstruction algorithm in the compressed sensing technology can be adopted to solve and obtain sparse signals

4. Training to generate a second neural network model and sparse signals according to a training method of deep learning

Is mapped as

Finally, pair->

Demodulation is performed to restore the original transmission data stream.

Compared with the prior art, the invention has the beneficial effects that:

1. the method has the greatest advantages that the method adopts the technical scheme that the method passes through the sparsification module 104 and the compressed sensing multiplexing module on the basis of the existing MIMO system, the traditional signal transmission flow is that code words are modulated and then are sent, and then the device receiving end 2 carries out signal detection to reconstruct original signals. Meanwhile, by combining with a BP neural network training method in artificial intelligence, sparse representation and reconstruction of signals are realized.

2. The compression multiplexing matrix provided by the invention does not need to depend on channel state information, and the technical scheme adopted by the invention can reconstruct the compression multiplexing signal on the basis that the device transmitting end 1 is added with sparsification and compression sensing multiplexing steps and the device receiving end 2 is added with a mature optimization reconstruction algorithm in the field of compression sensing on the basis that the prior MIMO technical scheme is not modified, thus the modification to the prior MIMO system is small and the realization is convenient.

3. The artificial intelligence method adopted by the invention sparsifies the representation signal, and reconstructs the original signal according to the sparse signal, thereby ensuring the feasibility of compressing and transmitting the multipath data stream and considering the bit error rate.

The specific implementation steps of the above process are described as follows:

sparse representation for "signal": compressed sensing (Compressed sensing), also known as compressed sampling or sparse sampling, is a method of finding sparse solutions for underdetermined linear systems. Compressed sensing is one of the most glaring achievements obtained by the signal processing field entering the 21 st century, and is effectively applied to the fields of magnetic resonance imaging, image processing, wireless communication systems and the like.

Converting the analog signal into a digital signal that can be processed by a computer entails a sampling process. To ensure signal integrity, nyquist gives an answer-the sampling frequency should be twice the highest frequency of the signal. In the past, nyquist sampling law has been regarded as the gold and jade law in the field of digital signal processing. The candles was the earliest to be aware of the possibility of breakthrough and with the help of Tao Zhexuan and doncho, a compressed sensing theory was proposed that considers: if the signal is sparse, it can be recovered by a sample point reconstruction that is far below the sampling theorem requirement.

Unlike the traditional nyquist sampling theorem, as long as the signal x is compressible or sparse in some transform domain D, the high-dimensional sparse signal obtained by the transform can be projected onto a low-dimensional space using an observation matrix a that is uncorrelated with the transform basis D, and then the original signal can be reconstructed from these small projections with high probability by solving an optimization problem. Within this theoretical framework, the sampling rate is not dependent on the bandwidth of the signal, but on the structure and content of the information in the signal. The compressed sensing theory mainly comprises three aspects of sparse representation of signals, coded sampling and a reconstruction algorithm. Since signals commonly existing in nature are generally not sparse, when signals are projected to a certain transform domain D, only a few elements are nonzero, the obtained transform vector is said to be sparse or approximately sparse, i.e., x=dθ, θ is a concise expression of the original signal x, which is a priori condition of compressed sensing.

How to find the optimal sparse domain of the signal is the basis and the premise of the application of the compressed sensing theory, and the sparseness of the signal can be guaranteed only by selecting a proper basis to represent the signal, so that the recovery precision of the signal is guaranteed. In studying sparse representation of signals, the sparse representation capability of the transform basis can be measured by the transform coefficient decay rate. Candles et al, document "Near optimal signal recovery from random projections: universal encoding strategies, "IEEE Transmission information Theory, vol.52, no.12, pp.5406-5425, 2006 states that satisfying signals with power-of-speed attenuation can be recovered using compressed sensing Theory. In recent years, a hotspot in sparse representation research is sparse decomposition of signals under redundant dictionaries. This is a completely new signal representation theory: the base functions are replaced by an overcomplete library of redundant functions, referred to as a redundant dictionary D, with the elements in the dictionary being referred to as atoms. Temlyakov in documents "Nonlinear Methods of Approximation, IMI Research Reports, dept. Of Mathematics, university of South Carolina,2001 indicates that the dictionary D should be selected to fit the structure of the approximated signal as well as possible, and its construction may be without any limitation. A small number of atoms with the best linear combination are found from the redundant dictionary to represent a signal x=dθ, known as a sparse or highly nonlinear approximation of the signal, where θ is non-zero for only K term elements. The overcomplete redundant dictionary D should be constructed to conform as much as possible to the inherent characteristics of the signal itself, which is important for sparse representation of the signal. The more the structure of the overcomplete redundant dictionary approximates the characteristics of the signal, the fewer atoms needed, the more sparse the θ, the fewer measurements needed, and the more accurate the reconstruction performance. Theoretically, a transform domain D can always be found, so as to realize sparse representation of signals.

The concept of deep learning is derived from the study of artificial neural networks. The neural network is a concept proposed by scientists, beginning with Rumelhart and McClellland, in 1986, and is a multi-layer feedforward neural network trained according to an error back propagation algorithm. At present, in the practical application of the artificial neural network, most neural network models adopt BP networks and variation forms thereof. The artificial neural network is also a core part of the forward network, and the essence of the artificial neural network is embodied. Here, we train out the first neural network model (Neural Network Model, NN 1) using BP algorithm to achieve sparse representation of the signals.

For "observation matrix in compressed sampling system": next, in the compressed sensing theory, an observation matrix a of the compressed sampling system needs to be designed, how to sample to obtain a small amount of observation values, and the original signal can be reconstructed from the small amount of observation values. Obviously, if the observation process destroys the information in the original signal, the quality of the reconstruction cannot be guaranteed. To ensure that the linear projection of the signal can preserve the original structure of the signal, the projection matrix must satisfy the constraint equidistance (Restricted Isometry Property, RIP) condition, and then a linear projection measurement of the original signal is obtained by the product of the original signal and the measurement matrix. RIP conditions are defined as follows: if a constant delta exists _K ∈(0，1]For all sparsity ofThe signal θ of K, matrix a satisfies the following equation:

the matrix a is said to satisfy the constrained equidistant property of order K, where sparsity K refers to the number of non-zero elements of the signal θ. A is N _t A compressed dimension-reducing matrix of rows and columns. The advantage of compressed sensing techniques is that even N _t < l (l refers to the length of the signal), we can still go from N _r (N _r ＝N _t ) The primary data of length l is recovered from the secondary measurements. Let->

Representing the compression ratio. According to the compressed sensing principle, as long as the measurement A meets RIP, even if the A is a matrix with the number of rows being far smaller than the number of columns, the signal theta is projected to a space with one dimension reduced, and the original signal can still be completely recovered from the measurement number far smaller than the signal dimension through a compressed sensing reconstruction algorithm. ρ determines the number of transmit and receive antennas that can be reduced and the performance of the device receiver 2 reconstruction. Davenport in its doctor paper "Random observation on random observations: sparse signal acquisition and processing "theorem 3.5 states that: a is the sum of 2K-order RIP constants delta _2K ∈(0，1]As long as->

C is a constant approximately equal to 0.28, the original signal can be recovered. Meanwhile, the document "Extensions of compressed sensing," Signal Processing, vol.86, no.3, pp.533-548, 2006 by Donoho gives three conditions necessary for observing the matrix, and indicates that most of uniformly distributed random matrices have these three conditions and can be used as the observation matrix, for example: a partial Fourier set, a partial Hadamard set, a uniformly distributed random projection (uniform Random Projection) set, etc. Document "decoding 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 demonstrates that when the measurement matrix A is a Gaussian random matrix, A can meet RIP with a high probability. A gaussian random matrix is chosen as the compressed sensing multiplexing matrix a in this application.

For "signal reconstruction at device receiving end 2": and finally solving the underdetermined problem through a reconstruction algorithm in compressed sensing. Donoho, document "For most large underdetermined systems of linear equations, the minimum ₀ Norm solution is also the sparsest solution, "Communications on Pure and Applied Mathematics, vol.59, no.6, pp.797-829, 2006 indicates the minimum l ₀ The norm problem is an NP-hard problem that requires an exhaustive list of all permutations of non-zero values in theta and is therefore not solved. In view of this, a series of algorithms for finding sub-optimal solutions are proposed by researchers in the field, and mainly include the following four classes:

(1) Greedy tracking algorithm: such methods approach the original signal by choosing a locally optimal solution at each iteration. These algorithms include the "Sparse solution of underdo-termined linear equations by stagewise orthogonal matching pursuit" Technical Report,2006 segmented OMP algorithm, etc. proposed by the document Donoho;

(2) Convex relaxation method: such a method is achieved by solving for l ₀ Conversion of the non-convex problem of norms to convex problem solution finds an approximation of the signal, as suggested by the interior point method of literature "A method for large scale regularized least squares," IEEE Journal on Selected Topics in Signal Processing, vol.4, no.1, pp.606-617, 2007, "Gradient projection for sparse reconstruction: application to compressed sensing and other inverse problems," Journal of Selected Topics in Signal Processing: special Issue on Convex Optimization Methods for Signal Processing, vol.1, no.4, pp.586-598, gradient projection method of 2007, "Daubechies in" An iterative thresholding algorithm for linear inverse problems with a sparsity constraint, "Comm. Pure appl. Math., vol.57, no.11, pp.1413-1457, iterative thresholding method of 2004, etc.;

(3) Bayesian compressed sensing reconstruction BCS algorithm: the method uses Bayesian prior to give a reasonable prior distribution to the solving signal, and then deduces the original signal, such as BCS (Bayesian Compressive Sensing) algorithm proposed by 'Bayesian compressive sensing using laplace priors,' IEEE Trans. Image Process, vol.19, no.1, pp.53-63, 20, etc.;

(4) Combining algorithm: such methods require that the sampling of the signal support a fast reconstruction by packet testing, such as fourier sampling as proposed in documents "Improved time bounds for near optimal sparse Fourier representation," Proceedings of SPIE, wavelets XI, bellingham WA: international Society for Optical Engineering,2005, "One sketch for all: fast algorithms for compressed sensing," Proceedings of the 39th Annual ACM Symposium on Theory of Computing,New York:Association for Computing Machiner,pp.237-246, HHS (Heavg Hitters on Steroids) tracking as proposed in 2007, etc.

As can be seen above, each algorithm has its inherent disadvantages. The convex relaxation method requires a minimum number of observations to reconstruct the signal, but tends to be computationally burdensome. The greedy tracking algorithm is located between these several classes of algorithms in terms of both run time and sampling efficiency, and noise immunity is unstable. An appropriate reconstruction algorithm can be selected according to different environments, and once the sparse representation vector is obtained, the original signal can be recovered.

For "signal reconstruction at device receiving end 2": as shown in fig. 1, the conventional signal transmission flow: the code word is transmitted after modulation, and compared with the original signal reconstructed by signal detection performed by the device receiving end 2, the thinning module 104 and the compression multiplexing module 105 are added at the device transmitting end 1. Firstly, expressing a modulated signal as a sparse signal through deep learning, and realizing nonlinear expression of the signal; and then carrying out compression dimension reduction processing on the sparse signals, wherein channel state information is not needed for selecting a compression dimension reduction matrix, and a measurement matrix in a compressed sensing technology is selected to be used as a signal compression matrix, so that compression dimension reduction and multiplexing processing on the transmitted signals are completed. The device receiving end 2 is divided into two steps to reconstruct signals: (1) Sensing by compression Reconstruction algorithm for solving high-dimensional sparse signals from low-dimensional received signals

(2) A second neural network model (Neural Network Model, NN 2) trained by a deep learning algorithm, reconstructing the original signal +.>

The invention can simultaneously transmit the same data quantity on the basis of reducing the number of required antennas and improving the multiplexing gain and capacity of the MIMO system by introducing the thinning module 104, the compression multiplexing module 105 and the demultiplexing module 202 of the device receiving end 2 on the basis of the existing related MIMO system signal multiplexing technology. Compared with the existing MIMO space multiplexing technical scheme, we do not only pay attention to eliminating the interference of adjacent data, but also pay attention to how to multiplex and transmit more data streams to the device receiving end 2 on the basis of guaranteeing the detection performance of the device receiving end 2 under the condition of given number of transmitting antennas, and obtain multiplexing gain and transmission capacity which exceed the inherent MIMO system.

Examples

The embodiment fuses the MIMO multi-antenna signal transmission and detection technology based on artificial intelligence and compressed sensing, and the specific implementation steps of the invention are illustrated in detail.

First, the information source is generated by generating a 0,1 bit sequence using a random number generator 101.

The modulation is to modulate bit data including BPSK, QPSK, 16QAM, 64QAM, and the like.

This embodiment will be described using BPSK modulation as an example.

According to the processing flow of the signal shown in fig. 2 at the device transmitting end 1 and the device receiving end 2, the specific steps are as follows:

s1: signal processing at the transmitting end 1 of the device.

S11: the 0,1 bit sequence is generated by a random number generator 101 to constitute the original data.

S12: the signal x is generated by BPSK modulation. The transmission data of each group may be different.

S13: training a neural network model. Creating a neural network through a BP neural network training method in artificial intelligence, respectively taking a transmission signal vector x of the device transmitting end 1 and a sparse representation theta determined based on the transmission signal vector x of the device transmitting end 1 as input and output, constructing a first group of training sample set, and training the neural network to obtain the first neural network model.

When the transmitted signal is x, the signal of the device receiving end 2 is y, and meanwhile, in theory, a proper base can be found, so that sparse representation of the signal is realized. Temlyakov in documents "Nonlinear Methods of Approximation, IMI Research Reports, dept. Of Mathematics, university of South Carolina,2001 indicates that the dictionary D should be selected to fit the structure of the approximated signal as well as possible, and its construction may be without any limitation. All possible combinations of x are used as different column vectors of the redundant dictionary D to form the redundant dictionary D, and sparse representation corresponding to x is found: the sparse vector θ with all positions 0 except for the corresponding index position 1.

S14: sparse representation of the signal. And obtaining a first neural network model through a BP neural network training method in artificial intelligence, and outputting a corresponding sparse signal theta when the input is the signal x.

S15: and (3) performing compression multiplexing processing, namely multiplying the sparse signal theta transmission data by the compression multiplexing matrix A to obtain a data vector z, namely z=Aθ.

The compressed multiplexing matrix a referred to herein is chosen to be a gaussian matrix. Wherein, the liquid crystal display device comprises a liquid crystal display device,

representing the compression ratio, represents the ratio of the number of antenna reductions.

S16: the data is transmitted over the channel.

S2: signal detection at the receiving end 2 of the device.

S21: the received signal is y=hz+n= HAx +n, where n represents noise. The signal matrix H is estimated according to the channel estimation module 201.

S22: according to the known multiplexing matrix A and channel matrix H, solving the following optimization problem, and calculating a sparse vector theta:

min||θ|| ₀ s.t.y＝HA·θ

in the above description, |·|| ₀ Is l of vector ₀ And the norm represents the number of non-zero elements in the sparse vector theta.

Assuming that the channel matrix is subjected to Gaussian distribution, the new matrix HA can be proved to still be subjected to Gaussian distribution, and the condition that the compressed sensing measurement matrix needs to meet is met. The underdetermined problem of compressed sensing can be solved by a reconstruction algorithm in compressed sensing. In this embodiment, a bayesian compressed sensing BCS reconstruction algorithm is adopted, and a sparse representation vector is obtained by solving according to a multiplexing matrix a and a channel matrix H known by the device receiving end 2

S23: training a neural network model. The input and output of the first set of training samples are swapped sequentially to form a second set of training samples. Creating a neural network through a BP neural network training method in artificial intelligence, constructing a second training sample set based on the sparse representation theta and the transmission signal vector x as input and output respectively, and training the neural network to obtain the second neural network model.

S24: according to the second neural network model, when the input is a signal

When a reconstructed transmission data is obtained +.>

Steps S13 and S23 are further described as follows: the neural network model consists of three parts, namely an input layer (layer 1), a middle layer (layer 2, the layer of the first layer, L-1) and an output layer (layerL), wherein the input layer plays a role in signal transmission and is responsible for receiving external input information, and each unit of the input layer represents a characteristic; the intermediate layer can be a single intermediate layer or multiple intermediate layers, which play a role inA part of information processing function, which is responsible for information transformation; the output layer functions to output information to the outside, and each unit of the output layer represents a category. In the application, a BP neural network is utilized to simulate a mapping function, and the mapping function can map input space data to output space; the BP neural network fits as much as possible a function between a transmission signal vector x, which is input and output to the device transmitting terminal 1, and a sparse representation theta determined based on the transmission signal vector x of the device transmitting terminal 1, respectively, and the input and output are the sparse representation theta and the original transmission data stream, respectively

A function therebetween; based on a mapping function model generated by a trained BP neural network, sparse representation theta of the sparse original signal x on the overcomplete redundant dictionary D can be obtained according to the modulated transmission signal vector x of the device transmitting end 1, and original transmission data x can be reconstructed according to the sparse representation theta deduced by a Bayesian reconstruction algorithm.

In the method, the cost function is used for measuring the difference between the BP neural network output and the real output, and the BP neural network is trained so that the output of the input (the signal received by the device receiving end 2) of the network after passing through the neural network can be as close to the theoretical output as possible. In order to minimize the cost function, the gradient descent method is used for solving the neural network parameters, and when the optimal neural network weight is solved, the first neural network model or the second neural network model is built. A BP neural network is created, a large amount of sample data is collected, and the correct classification results are artificially labeled, and then the created neural network is trained with the labeled data. In this process, each layer in the neural network continuously adjusts its own weight and bias based on the difference between the current output value and the labeled correct target value until the target value can be accurately output.

For two parameters that are needed in training the neural network-weight and bias further description: in the present application, a weight parameter matrix between layers of a neural network is used

And (3) representing the number of layers by the superscript of the weight parameter w, and the number of nodes of each of two adjacent layers of the subscript. For example, a->

The weights of the line segments representing the 1 st node of Layer1 and the 2 nd node of Layer2 are input. These weights determine the role of the model, and the goal of the neural network is to compute the weights from the samples. The nodes of each intermediate layer and output layer are a Logistic function g (z) =a. For example->

The input value representing node 1 of Layer2 is brought into the Logistic function to obtain the output +.>

The bias parameter matrix between each layer of the neural network is as follows: b= [ B ] ₁ b ₂ ...b _n ] ^T Inputs to a neural network are known as: y= [ Y ] ₁ y ₂ ...y _n ] ^T The output of the neural network is:

introducing a nonlinear operator:

it can be deduced that: />

Initializing weight parameters: the weight parameter w is randomly initialized to lie between [ -epsilon, epsilon ] which is a preset, sufficiently small value.

Training a neural network model: the process of training the neural network model is mainly divided into two steps, namely, calculating a cost function J (theta), and adjusting a parameter theta to enable the cost function J (theta) to be as small as possible. The forward propagation algorithm is adopted as follows, the output of each sample under the current neural network model is calculated for the sample, the cost function is obtained, and the weight parameters are updated according to the output. Defining a cost function J (theta), wherein m is the number of samples, and since the neural network has K outputs, the cost function correspondingly calculates the cost of the K outputs, and the calculation formula is as follows:

The back propagation algorithm is used to adjust the parameter θ such that the cost function value J (θ) is as small as possible as follows. The back propagation algorithm updates each weight coefficient by taking the partial derivative of the cost function with respect to each weight coefficient. For example, first, the gradient of the last layer is calculated: (1) Calculating the gradient of the cost function value to the nonlinear operator, (2) calculating the gradient of the output of the neural network to the bias and the weight between adjacent layers. Updating the gradient according to the negative direction of the gradient; next, the gradient of the penultimate layer is calculated: (1) Calculating the gradient of the error returned by the upper layer to the nonlinear operator, (2) calculating the gradient of Hn-1 (H is the output of each layer after the activation function) to the bias and the weight between adjacent layers, and updating the gradient according to the negative direction of the gradient; finally, after passing back layer by layer, the gradient of the first layer is finally calculated: (1) Calculating the gradient of the error returned by the second layer to the nonlinear operator; (2) calculating the gradient of H1 to bias and adjacent inter-layer weight. And updates the gradient in the negative direction of the gradient. And after the cycle of the first back propagation process is finished, continuing the forward propagation to obtain output, and backward propagation to update parameters until the mean square error is minimum, thereby completing the training process of the neural network model.

Fig. 3 shows bit error rate performance of different transceiver antenna configurations of the MIMO system after the signal transmission and detection techniques of compressed sensing and neural networks are adopted. Here a flat fading channel is assumed. Under the conventional scheme, 4 transmit antennas can only transmit 4 data symbols at the same time. By applying the proposal provided by the application, BPSK modulation signals are firstly adopted to obtain the original signal x ^{4 ×1} Obtaining a sparse signal theta through a first neural network model ^16×1 . Random Gaussian matrix A ^4ρ×16 As a compressed dimension-reducing matrix, z is obtained. If ρ=0.5, only 2 is neededThe transmission of the original data can be realized by the root transmitting antenna. If ρ=0.75, only 3 transmitting antennas are needed to realize the transmission of the original data. The device receiving end 2 obtains a reconstructed transmitting signal by adopting a BCS reconstruction algorithm and a second neural network model trained by the neural network through 2 or 3 receiving antennas

The error rate performance of the scheme is shown as (2 x 2) -4 in the figure, the first number in brackets indicates the number of transmitting antennas, the second number indicates the number of receiving antennas, and the last number indicates the original data length. The number of receiving antennas is increased and this scheme is denoted as (3 x 3) -4. Compared with a classical detection algorithm ZF (Zero Forcing), the scheme provided by the application can reduce the number of required receiving and transmitting antennas while guaranteeing the error rate under the condition of high SNR.

Fig. 4 shows bit error rate performance of different transceiver antenna configurations of a MIMO system after signal transmission and detection techniques using compressed sensing and neural networks. Here a flat fading channel is assumed. Under the conventional scheme, only 20 data symbols can be simultaneously transmitted by 20 transmit antennas. By applying the scheme of the application, BPSK modulation signals are adopted first, and then x is calculated ^20×1 Divided into 5 groups, the vector x of each group _i (i=1, 2,3,4, 5) is equal to 4. Each group of signals passes through a first neural network model to obtain sparse signals theta _i ^16×1 (i=1, 2,3,4, 5). Selecting a random Gaussian matrix A ^4ρ×16 As a compressed dimension-reducing matrix, thereby obtaining z _i . Will z _i Cascading to obtain the vector to be sent

If ρ=0.5, only 10 transmitting antennas are needed to realize the transmission of the original data. If ρ=0.75, 15 transmitting antennas are needed to realize the transmission of the original data. The receiving end 2 of the device also passes through 10 receiving antennas, the error rate performance of the scheme is shown as (10 multiplied by 10) -20 in the figure, the first number of transmitting antennas is arranged in brackets, the second number represents the number of receiving antennas, and the last number represents the original dataLength. The same data, the same packet number, ρ=0.75, the compressed dimension-reduction matrix is a ^3×4 The number of receiving antennas increases, and this scheme is denoted as (15×15) -20. Compared with a classical detection algorithm ZF (Zero Forcing), the scheme provided by the application can reduce the number of required receiving and transmitting antennas while guaranteeing the error rate under the condition of high SNR. Therefore, the proposal provided by the proposal can reduce the number of the needed receiving and transmitting antennas while ensuring the error rate.

As described above, the enhanced spatial multiplexing method provided by the invention can combine artificial intelligence on the basis of the existing MIMO system, transmit the same data volume on the basis of reducing the number of antennas, reduce the number of required antennas and improve multiplexing gain and system capacity.

It should be noted that the above-described embodiments are exemplary, and that a person skilled in the art, in light of the present disclosure, may devise various solutions that fall within the scope of the present disclosure and fall within the scope of the present disclosure. It should be understood by those skilled in the art that the present description and drawings are illustrative and not limiting to the claims. The scope of the invention is defined by the claims and their equivalents.

Claims (10)

1. An antenna signal modulation method based on a neural network model, which is characterized in that the method at least comprises the following steps:

creating a neural network by a BP neural network training method in artificial intelligence, respectively taking a transmission signal vector x of a device transmitting end (1) and a sparse signal theta determined based on the transmission signal vector x of the device transmitting end (1) as input and output, constructing a first training sample set, and training the neural network to obtain a first neural network model;

performing compression dimension reduction processing on the sparse representation signals;

solving and obtaining sparse signal vector from low-dimensional signal by using compressed sensing reconstruction algorithm

And finally, reconstructing the received signal by using the second neural network model to obtain an original signal.

2. The neural network model-based antenna signal modulation method of claim 1, further comprising:

training to generate a second neural network model and sparse signals according to a training method of deep learning

Is mapped as +.>

Finally, pair->

Demodulation is performed to restore the original transmission data stream.

3. The antenna signal modulation method based on the neural network model according to claim 1 or 2, characterized in that the method further comprises:

Obtaining a signal x of a channel after channel coding and signal modulation; the modulated l-path signal x is mapped into a sparse signal theta through a first neural network model generated by training in advance ^m×1 。

4. A method for modulating an antenna signal based on a neural network model according to any one of claims 1 to 3, wherein the method for performing the compressed dimension reduction processing on the sparsely represented signal further comprises:

sparse signal theta obtained after sparsification ^m×1 Compression multiplexing module compresses into ρ1 path signal, and then the compressed multiplexed signal z is transmitted by transmitting antenna;

the compression processing of the input signal by the compression multiplexing module can be expressed as: z=aθ, where a is N _t Compression dimension-reducing moment of row m and columnThe array of which is arranged in a row,

representing the compression ratio.

5. The method for modulating antenna signals based on neural network model according to any one of claims 1-4, wherein sparse signal vectors are obtained by solving from low-dimensional signals using a compressed sensing reconstruction algorithm

The method of (1) comprises:

setting the received signal as: y=hz+n=ha θ+n;

wherein y is a received signal vector of Nr×1, namely a ρ1 path modulation symbol after compression multiplexing, which is transmitted by n transmitting antennas; n is Nr multiplied by 1 Gaussian white noise vector, and the element is an independent co-distributed complex Gaussian variable with the mean value of O and the variance of 1; h is the nr×nt channel propagation matrix;

According to the compressed sensing multiplexing matrix HA, a decompression multiplexing process is realized by solving the following optimization problems, and a transmitted sparse vector is calculated and determined: min θ ₀ s.t.y＝H·A·θ；||·|| ₀ Is l of vector ₀ A norm representing the number of non-zero elements in the sparse vector θ;

bayes-based compressed sensing reconstruction algorithm solution to obtain sparse signals

6. The method for modulating an antenna signal based on a neural network model according to any one of claims 1 to 5, further comprising:

and carrying out channel estimation on the signal received by the receiving end (2) of the device, and taking the obtained channel parameter matrix as the input of the compressed sensing reconstruction algorithm.

7. The antenna signal modulation method based on the neural network model according to any one of claims 1 to 6, characterized in that the method further comprises:

and forming a redundant dictionary D by taking all possible combinations of the transmitted signal vectors x as different column vectors of the redundant dictionary respectively, so as to realize sparse signals theta of the transmitted signal vectors, namely x=Dtheta.

8. The antenna signal modulation method based on the neural network model according to any one of claims 1 to 7, wherein the training method of the second neural network model comprises:

Exchanging the input and output of the first training sample set for a sequence to form a second training sample set;

creating a neural network by a BP neural network training method in artificial intelligence, and constructing a second training sample set by taking the sparse representation theta and the transmission signal vector x as input and output respectively;

training the neural network to obtain the second neural network model.

9. The method for modulating an antenna signal based on a neural network model according to any one of claims 1 to 8, wherein the method for training the neural network model further comprises:

calculating a cost function J (θ): calculating the output of each sample under the current neural network model by adopting a forward propagation algorithm, and solving a cost function;

and adjusting the parameter theta to make the cost function value J (theta) as small as possible, wherein the back propagation algorithm is adopted to update each weight coefficient by solving the partial derivative of the cost function on each weight coefficient.

10. The method for modulating an antenna signal based on a neural network model according to any one of claims 1 to 9, wherein the step of using a back propagation algorithm to calculate partial derivatives of the cost function with respect to the respective weight coefficients comprises:

The gradient of the last layer is calculated: calculating the gradient of the cost function value to the nonlinear operator, and calculating the gradient of the output pair bias of the neural network and the weight between adjacent layers;

calculating the gradient of the penultimate layer: calculating the gradient of the error pair nonlinear operator returned by the upper layer, calculating the gradient of Hn-1 (H is the output of each layer after the activation function) pair bias and adjacent layer weight, and updating the gradient according to the negative direction of the gradient;

after passing back layer by layer, the gradient of the first layer is finally calculated: calculating the gradient of the error returned by the second layer to the nonlinear operator; calculate H ₁ Gradient to bias and adjacent inter-layer weights.

Images