CN111769862B

CN111769862B - Joint detection method in spatial modulation network based on deep learning

Info

Publication number: CN111769862B
Application number: CN202010594156.2A
Authority: CN
Inventors: 张文奇; 刘林; 蒋昕怡; 束锋; 邹骏; 夏桂阳
Original assignee: Nanjing University of Science and Technology
Current assignee: Nanjing University of Science and Technology
Priority date: 2020-06-28
Filing date: 2020-06-28
Publication date: 2022-11-11
Anticipated expiration: 2040-06-28
Also published as: CN111769862A

Abstract

The invention discloses a joint detection method in a spatial modulation network based on deep learning. The method comprises the steps of processing a channel and a received signal into real-valued row vectors, stacking the real-valued row vectors and the received signal to generate training samples, forming a training set by a batch of samples, processing original data bits into a single-hot coding form, and generating labels of the training set; then, building a deep neural network, and training the deep neural network by using a training set and labels thereof; and finally, inputting the channel and the received signal to the deep neural network to obtain output, namely a signal detection result, when the channel matrix and the received signal change. As long as the random distribution of channels does not change, no new training is required. The deep learning-based joint detection method has the beneficial effects that the deep learning-based joint detection method provided by the invention does not need repeated and tedious calculation, and obtains better bit error rate performance with acceptable complexity.

Description

Joint detection method in spatial modulation network based on deep learning

Technical Field

The invention relates to the technical field of wireless communication, in particular to a joint detection method in a spatial modulation network based on deep learning.

Background

At present, the fifth generation mobile communication technology has been advanced to the aspects of production, manufacturing and daily life, and the development thereof has entered the final stage by using key technologies such as large-scale Multiple Input Multiple Output (MIMO) to improve efficiency and super-dense networking in key areas. The academic community is increasingly interested in developing next Generation networks, and it is expected that the next decade will be devoted to developing the Sixth Generation mobile communication networks (6G). In 2019, china formally starts 6G research and combines technologies such as mobile ultra wide band, internet of things and artificial intelligence. The large-scale MIMO technology has great advantages in the aspects of simplifying a medium access control layer, improving throughput, reducing delay, increasing radiation energy efficiency, using cheap low-power-consumption components and the like, and has great potential in 6G research.

Conventional MIMO has many disadvantages, such as difficulty in removing inter-channel interference, difficulty in synchronizing between antennas, difficulty in balancing system performance with receiver complexity, and fewer transmit antennas than receive antennas.

To solve the above problem, a new MIMO technique, spatial Modulation (SM), arises, which divides a bit information stream into two sub-information blocks, one mapped to an antenna index and the other mapped to a constellation symbol. Within each time slot, only one transmit antenna is activated. The wireless communication channel has high complexity and randomness, and the channels corresponding to each transmitting antenna of the SM are greatly different, so that the receiver can distinguish and identify data from different antennas in a certain detection mode. The main advantages of SM are: 1) Only one antenna is activated in each time slot, so that the interference between channels is well eliminated, the synchronization between the antennas is avoided, and the radio frequency overhead and the power consumption are reduced. 2) The frequency spectrum efficiency and the energy efficiency are improved to a certain extent by using the ordinal number of the antenna to carry bit information. 3) The SM scheme can still operate efficiently if there are more transmit antennas than receive antennas, so the SM is suitable for low complexity mobile unit downlink setup.

In a conventional SM signal detection method, bit Error Rate (BER) performance of Maximum Likelihood (ML) detection is optimal, and a main idea is to search all possible transmission vectors, that is, all possible antenna indexes and constellation symbol pairs, at a receiving end to find an optimal solution of a Maximum Likelihood function. The computational complexity of maximum likelihood detection is very high, growing exponentially as the number of transmit antennas and the order of the modulation constellation increase. Although researchers have proposed methods with lower complexity, such as sphere decoding detection, zero forcing detection, minimum mean square error detection, matched filtering detection, etc., in succession, the whole process is still relatively complicated, has repeatability, and is not favorable for effectively utilizing resources.

Disclosure of Invention

In the SM system, a signal detector at the receiving end detects the transmit antenna index and constellation symbol from the received signal, and then demodulates the original data bits. The ML detection method searches all possible transmitting vectors, namely all possible antenna index and constellation symbol pairs, at a receiving end to find the optimal solution of the maximum likelihood function. If the channel state information is known at the receiving end, ML detection can be expressed as:

wherein,

antenna index and constellation symbol index, p, estimated for the receiving end, respectively _y (y|x _j，m And H) is a likelihood function. The calculation needs to be repeated, the steps are complicated, a large amount of resources are consumed, and the improvement of the system performance is not facilitated.

The invention aims to provide a combined detection method based on deep learning, aiming at repeated and tedious calculation in the traditional signal detection in an SM network. Generating training samples by using a channel and a received signal to form a training set, generating a label by using original data bits to train a Deep Neural Network (DNN), and after training, inputting the DNN when a channel matrix and the received signal are changed, so that output can be obtained, namely a detection result. After DNN training is completed, no new training is needed as long as the random distribution of channels does not change.

In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps: training samples are generated using the channel and the received signal to form a training set, and the original data bits are used to generate the labels. The deep neural network is trained using the training set and its labels. And inputting a channel and receiving a signal to the deep neural network to obtain output, namely a signal detection result.

Further, the specific process comprises the following steps: s1, processing a channel into a real-valued row vector, processing a received signal into a real-valued row vector, stacking the real-valued row vector and the received signal to generate training samples, forming a training set by a batch of samples, processing original data bits into a single-hot coding form, and generating a label corresponding to the training set; s2, using a Tensorflow frame, adopting a random gradient descent method, introducing an Adam self-adaptive learning rate method and a regularization method, building DNN, and training a deep neural network by using a training set and labels thereof; and S3, inputting a channel and receiving a signal to the deep neural network to obtain an output, namely a signal detection result. When the channel matrix and the received signal are changed, the channel matrix and the received signal are input into DNN, and output, namely a detection result, can be obtained. After DNN training is completed, no new training is needed as long as the random distribution of channels does not change.

And (3) simulation results: setting the number of transmitting antennas N _t Number of receiving antennas N is 4 _r And 4, adopting normalized 256QAM modulation. The simulation results are shown in fig. 4. The results show that: the BER performance of the detection method based on deep learning is almost the same as that of ML detection when the Signal-to-Noise Ratio (SNR) is high. When SNR is higher, for BER performance, the detection method based on deep learning is obviously superior to zero forcing detection, minimum mean square error detection and matched filtering detection, is inferior to ML detection and sphere decoding detection, but the performance loss compared with ML detection is less than 1dB. For the detection method based on deep learning, after DNN training is well carried out, repeated calculation is not needed as long as the random distribution mode of the channel is not changed, and resources are effectively utilized.

The invention has the beneficial effects that: a training set is constructed through a deep learning method, a deep neural network is trained, after training is completed, each time a channel matrix and a received signal change, the channel matrix and the received signal only need to be input into the neural network to obtain output, namely a detection result, repeated and tedious calculation is not needed, and better bit error rate performance is obtained with acceptable complexity. After the network is trained, no new training is needed as long as the random distribution of the channels is not changed.

Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.

Drawings

Fig. 1 is a flowchart of a joint detection method in an SM network based on deep learning.

Fig. 2 is a flowchart of a training process of a joint detection method in an SM network based on deep learning.

Fig. 3 is a flowchart of a detection process of a joint detection method in an SM network based on deep learning.

Fig. 4 is a graph of BER performance for different detection methods in an SM network.

Detailed Description

The present invention is further illustrated by the following figures and specific examples, which are to be construed as merely illustrative and not limitative of the remainder of the disclosure, and all changes and modifications that would be obvious to those skilled in the art are intended to be included within the scope of the present invention and the appended claims are intended to be embraced therein.

The invention provides a joint detection method in an SM network based on deep learning. Suppose the number of transmit antennas of a SM system is N _t ＝2 ^k (k∈N ₊ ) The number of receiving antennas is N _r . In each time slot, the system maps one part of the data vector into the transmitting antenna index according to the mapping table, and maps the rest part into the constellation point of the M-order constellation diagram, so that the antenna index can carry log ₂ N _y Individual bits, constellation points can carry log ₂ M bits. And the activated transmitting antenna transmits the constellation symbols to the receiving antenna through the channel. And finally, a signal detector at the receiving end detects the transmitting antenna index and the constellation point according to the received signal and recovers the original bit data according to the mapping table. The combined detection method based on deep learning mainly comprises the following steps:

s1, generating training samples by using a channel and a received signal to form a training set, and generating a label by using original data bits.

This task essentially comprises two steps: training samples are generated and composed into a training set using the channel and the received signal, and labels are generated using the raw data bits.

And S11, generating training samples by using the channel and the received signal and forming a training set.

The DNN requirements for the training set are: each training sample is a real-valued row vector. The channel matrix and the received signal need to be processed into one real-valued row vector.

First, the channel is processed into real-valued row vectors.

Assuming that a channel matrix corresponding to bit data transmitted in the ith time slot is:

wherein,

will H ⁽ⁱ⁾ Line 2 to N _r Rows, moving to the first row in turn, making them row vectors:

wherein,

each represents H ⁽ⁱ⁾ Line 1, line 2, line N _r And (6) rows. Then H is introduced ⁽ⁱ⁾ The real and imaginary parts of each element are taken out and placed side by side in the row vector:

H ⁽ⁱ⁾ ″＝[real(H ⁽ⁱ⁾ ′)imag(H ⁽ⁱ⁾ ′)]#(4)

wherein,

real(H ⁽ⁱ⁾ ') denotes the pair H ⁽ⁱ⁾ ' Each element takes the real part, imag (H) ⁽ⁱ⁾ ') denotes the pair H ⁽ⁱ⁾ Each element of' takes an imaginary part. Assuming that a training batch has t samples corresponding to t timeslots, the first part of the training set x is:

wherein,

the received signal is then processed into real-valued row vectors.

After the transmission signal of the ith time slot is transmitted in the channel, the receiving signal of the receiving end is:

wherein,

transposing the vector to obtain a row vector:

wherein,

will r is ⁽ⁱ⁾ ' real and imaginary parts of elements are taken out side by side in the row vector:

r ⁽ⁱ⁾ ″＝[real(r ⁽ⁱ⁾ ′)imag(r ^(i)′ )]#(8)

wherein,

real(r ⁽ⁱ⁾ ') denotes the pair r ⁽ⁱ⁾ ' Each element takes the real part, imag (r) ⁽ⁱ⁾ ') denotes the pair r ⁽ⁱ⁾ Each element of' takes an imaginary part. Assuming that a training batch has t samples corresponding to t timeslots, the second part of the training set x is:

wherein,

mixing the sampleFirst part x of x ₁ And a second part x ₂ Stacking, the training set to obtain DNN is:

s12, generating a label by using the original data bit.

The invention regards the signal detection of the spatial modulation network as a classification problem, and the DNN for solving the classification problem has the following requirements on the label: each tag is in a unique thermally encoded form. The original data bits need to be processed into a one-hot encoded form.

In an SM system, if the number of transmitting antennas is N _t If M-order APM modulation is adopted, then one time slot transmits beta = (log) ₂ N _t +log ₂ M) bit data modulated to N _t One of the antenna indices j and one of the M constellation symbols M, for a total of N _t And x M is possible. Assuming that the ith sample corresponds to the nth possibility, the label corresponding to the sample is expressed in a one-hot coded form as:

wherein y is ⁽ⁱ⁾ _n =1 indicates that the sample corresponds to the nth possibility.

Assuming that a training batch has t samples corresponding to t time slots, the label y of the training batch is:

and S2, training the deep neural network by using a training set.

And S21, building a deep neural network.

And (3) using a Tensorflow frame, adopting a random gradient descent method, introducing an Adam self-adaptive learning rate method and a regularization method, and building DNN.

And S22, training a deep neural network.

Having obtained the training set x and labels y that meet the requirements of the DNN classification problem in the foregoing steps, the DNN is now trained using the training set and labels.

And S3, inputting a channel and receiving a signal to the deep neural network to obtain an output, namely a signal detection result.

When the channel matrix and the received signal are changed, the channel matrix and the received signal are input into DNN, and output, namely a detection result, can be obtained. After DNN training is completed, no new training is needed as long as the random distribution of channels does not change.

The invention has the beneficial effects that: through a deep learning method, a training set is constructed, a deep neural network is trained, after training is completed, each time a channel matrix and a received signal change, the channel matrix and the received signal only need to be input into the neural network to obtain output, namely a detection result, repeated and tedious calculation is not needed, and better bit error rate performance is obtained with acceptable complexity. After the network has been trained, no new training is needed as long as the random distribution of the channels does not change.

Fig. 1 is a flow chart of a joint detection method in an SM network based on deep learning.

FIG. 4 is a graph of BER performance for different detection methods in an SM network, where the number of transmit antennas N is _t 4, number of receiving antennas N _r For 4, normalized 256QAM modulation is used. The curves in the figure show that: the BER performance of the detection method based on deep learning is almost the same as that of ML detection when the Signal-to-Noise Ratio (SNR) is high. When SNR is higher, for BER performance, the detection method based on deep learning is obviously superior to zero forcing detection, minimum mean square error detection and matched filtering detection, is inferior to ML detection and sphere decoding detection, but the performance loss compared with ML detection is less than 1dB. For the detection method based on deep learning, after DNN training is good, the detection method only needs to be carried outThe random distribution mode of the channel is not changed, repeated calculation is not needed, and resources are effectively utilized.

Claims

1. A joint detection method in a spatial modulation network based on deep learning is characterized in that: the method comprises the steps of generating training samples by using a channel and a received signal to form a training set, generating a label by using original data bits, training a deep neural network by using the training set and the label, inputting the channel and the received signal to the deep neural network after training is finished, and obtaining output, namely a signal detection result; the method comprises the following specific steps:

s1, processing a channel into a real-valued row vector, processing a received signal into a real-valued row vector, stacking the real-valued row vector and the received signal to generate training samples, forming a training set by a batch of samples, processing original data bits into a single-hot coding form, and generating a label corresponding to the training set; the relevant process is as follows:

in the spatial modulation system, if the number of transmitting antennas is N _t With M-order phase amplitude modulation, one slot transmits β = (log) ₂ N _t +log ₂ M) bit data modulated to N _t One of the antenna indices j and one of the M constellation symbols N, for a total of N _t xM is possible; assuming that the ith sample corresponds to the nth possibility, the label corresponding to the sample is expressed in a form of one-hot encoding as follows:

wherein y is ⁽ⁱ⁾ _n =1 indicates that this sample corresponds to the nth possibility;

s2, using a Tensorflow frame, adopting a random gradient descent method, introducing an Adam self-adaptive learning rate method and a regularization method, building DNN, and training a deep neural network by using a training set and labels thereof;

s3, inputting a channel and a receiving signal to the deep neural network to obtain an output, namely a signal detection result; when the channel matrix and the received signal are changed, the channel matrix and the received signal are input into DNN to obtain output, namely a detection result; after DNN training is completed, no new training is needed as long as the random distribution of channels does not change.

2. The joint detection method in the spatial modulation network based on deep learning of claim 1, wherein: generating training samples by using a channel and a received signal to form a training set, and generating a label by using original data bits, wherein the method comprises the following specific steps of: processing a channel into a real-valued row vector, processing a received signal into a real-valued row vector, stacking the real-valued row vector and the received signal to generate training samples, forming a training set by a batch of samples, processing original data bits into a single-hot coding form, and generating a label corresponding to the training set.

3. The joint detection method in the spatial modulation network based on deep learning of claim 1, wherein: training a deep neural network by using a training set, and specifically comprising the following steps: a Tensorflow frame is used, a random gradient descent method is adopted, an Adam self-adaptive learning rate method and a regularization method are introduced, DNN is built, and a training set and labels thereof are used for training a deep neural network.

4. The joint detection method in the spatial modulation network based on deep learning of claim 1, wherein: inputting a channel and receiving a signal to the deep neural network to obtain an output, namely a signal detection result, and specifically comprising the following steps: inputting a channel and a receiving signal to the deep neural network to obtain an output, namely a signal detection result; when the channel matrix and the received signal are changed, the channel matrix and the received signal are input into DNN to obtain output, namely a detection result; after DNN training is completed, no new training is needed as long as the random distribution of channels does not change.