阅读说明：本技术 一种基于神经网络的智能反射表面反射相位配置方法 (Intelligent reflection surface reflection phase configuration method based on neural network ) 是由 党建 鲁文韬 朱永东 郭荣斌 张在琛 吴亮 于 2021-01-26 设计创作，主要内容包括：本发明公开了一种基于神经网络的智能反射表面反射相位配置方法,包括：由发送端向接收端发射导频信号,使接收端获取信道状态信息；基站端根据获取的统计信道状态信息和神经网络预测最优的反射相位,并通过有线或无线链路控制智能反射表面,使其反射阵列的反射相位得到优化,更新神经网络数据集并继续训练神经网络参数,当发送端或接收端的载波频率、相对空间位置,或者它们之间的无线传输环境发生改变而导致统计信道状态信息发生改变后,重复步骤。本发明使用信道状态信息来配置反射相位,通过不断优化神经网络模型参数,来快速预测优化的智能反射表面优化相位配置,使得接收信号的功率得到优化。(The invention discloses an intelligent reflection surface reflection phase configuration method based on a neural network, which comprises the following steps: transmitting a pilot signal to a receiving end by a transmitting end to enable the receiving end to acquire channel state information; the base station terminal predicts the optimal reflection phase according to the acquired statistical channel state information and the neural network, controls the intelligent reflection surface through a wired or wireless link, optimizes the reflection phase of the reflection array, updates the neural network data set and continues training neural network parameters, and repeats the steps after the statistical channel state information is changed due to the change of the carrier frequency and the relative spatial position of the transmitting terminal or the receiving terminal or the wireless transmission environment between the transmitting terminal and the receiving terminal. The invention uses the channel state information to configure the reflection phase, and the optimized intelligent reflection surface optimized phase configuration is rapidly predicted by continuously optimizing the neural network model parameters, so that the power of the received signal is optimized.)

1. An intelligent reflection surface reflection phase configuration method based on a neural network is characterized in that uplink data transmission comprises the following steps:

step 1, sending data frame from sending end to receiving end, where the sending end has N antennas, the receiving end has M antennas, each data frame contains L time slots, where the first L time slots_pOne time slot is a pilot signal, and L-L is remained_pEach time slot is a data signal, and partial power of a data frame is reflected by an intelligent reflection surface with K reflection units and reaches a receiving end;

step 2, receiving the pilot frequency part in the current data frame and utilizing the received L_pPilot frequency time slot, obtaining channel state information matrix H_t∈C^M×NT is the sequence number of the current data frame, C represents the complex field, and meanwhile, the receiving end measures the power of the pilot signal, and the measurement result is marked as P_r；

Step 3, at the receiving end, for the t-th data frame, H_t、H_t-1、θ_(t-1,n)、θ_(t-2,n)The intelligent reflection surface reflection phase theta in the t data frame time is predicted as the input of the initial neural network, and the prediction result is recorded as theta_(t,n)N is 1,2, and K, n is the serial number of the current reflection unit;

step 4, the base station adjusts the phase configuration of the intelligent reflection surface according to the predicted phase configuration, and the intelligent reflection surface is used for enhancing and receiving the data part in the data frame;

step 5, the receiving end receives the data part of the t-th data frame, carries out signal detection and measures the dataSignal power, measurement P_cCalculating the power gainWherein G (x) is a monotonically increasing function with respect to x;

step 6, the receiving end sends the channel state information H_t，H_t-1And reflection phase configuration theta of the intelligent reflection surface when receiving pilot frequency_(t-1,n)，θ_(t-2,n)As input, the prediction result θ obtained_(t,n)As an output, a new sample is created, denoted M_t；

Step 7, the receiving end constructs the sample M according to the new structure_tConstructing a new sample set R_t(ii) a For all constructed samples M_tPerforming extraction for W times according to probability, and constructing a sample set with W samples;

step 8, the receiving end uses the constructed sample set R_tFurther training the existing neural network model, wherein the maximum iteration number of the training isWherein F (x) is a monotone decreasing function for x and stores the trained neural network for prediction of the reflection phase in the next data frame;

and 9, repeating the steps 1-8 when receiving the next data frame.

2. The method for configuring the reflection phase of the intelligent reflection surface based on the neural network as claimed in claim 1, wherein: the probability that the kth sample is extracted is

Where τ (T-k) is a decay function, a monotonous decreasing function with respect to (T-k), G_kIs the power gain of the corresponding sample.

3. The method for configuring the reflection phase of the intelligent reflection surface based on the neural network as claimed in claim 1, wherein: in step 3, the initial neural network is a neural network which is pre-trained or uses random parameters; the neural network model is trained by adopting a BP or CNN neural network model.

4. The method for configuring the reflection phase of the intelligent reflection surface based on the neural network as claimed in claim 1, wherein: the intelligent reflection surface is composed of a plurality of basic reflection units which are densely arranged, and each basic reflection unit can generate independent and controllable amplitude attenuation, phase offset and frequency offset on incident electromagnetic waves; the intelligent reflection surface is connected with the base station and controlled by the base station, and the power of the received signal reaches the maximum value by adjusting the phase offset of each basic reflection unit.

5. The method for configuring the reflection phase of the intelligent reflection surface based on the neural network as claimed in claim 1, wherein: in step 8, the monotone decreasing function F (x) is required to ensure the signal power gainThe smaller the maximum number of iterations of the neural network, the greater the set

Technical Field

The invention relates to an intelligent reflection surface phase configuration method based on a neural network, and belongs to the technical field of wireless communication.

Background

Recently, with the advance of new artificial electromagnetic materials, intelligent reflective surfaces have been proposed as a promising low-cost solution, which enables optimization of signal transmission by modifying the propagation environment of a wireless communication system so that the propagation environment between transceivers can be controlled.

A smart transmitting surface is a planar array of a large number of passive elements each of which is capable of producing controllable phase shifts, amplitude attenuation, etc. effects independently of the incident electromagnetic wave by means of a smart controller, and by adjusting the phase shift of all the elements on the smart reflecting surface, the reflected signal can be added coherently to the desired receiver to increase the power of the received signal or in the form of negative gains to unintended receivers to avoid interference and enhance privacy. A more common solution today is to interconnect the intelligent reflecting surface with a communication system, the phase shift of each reflecting element being controlled by the communication system, so that the power of the received signal is enhanced. At this time, in addition to the original channel in which the transmitting end directly reaches the receiving end, two channels in which a signal reaches the intelligent reflection surface from the transmitting end and reaches the receiving end by being reflected from the intelligent reflection surface are generated. Most of the literature and literature today considers calculating and optimizing the reflection coefficients of the reflecting elements given the precise instantaneous channel state information of these three channels. In fact, the intelligent reflecting surface does not have the function of receiving and analyzing signals, and the instantaneous channel state information of each channel is difficult to accurately estimate, so that the intelligent reflecting surface is difficult to be applied to a practical system.

Disclosure of Invention

The invention provides a method for adjusting the reflection phase of the intelligent reflection surface in real time by using a neural network to assist the wireless communication by considering the practical application scene of the intelligent reflection surface and the defects of the prior art.

The invention specifically adopts the following technical scheme to solve the technical problems:

the uplink data transmission comprises the following steps:

step 1, sending data frame from user end to base station end, where the user end has N antennas, the base station end has M antennas, each data frame includes L time slots, where the first L time slots_pOne slot is a pilot signal, the remainder (L-L)_p) And each time slot is a data signal, and partial power of a data frame is reflected by an intelligent reflecting surface with K reflecting units and reaches a base station end.

Step 2, the base station receives the pilot frequency part in the current data frame and utilizes L_pPilot frequency time slot, obtaining channel state information matrix H_t∈C^M×NT is the sequence number of the current data frame, and C represents a complex number field. Meanwhile, the base station measures the pilot signal power, and the measurement result is marked as P_r。

Step 3, at the base station end, for the t data frame, H_t,H_t-1,θ_(t-1,n)，θ_(t-2,n)The intelligent reflection surface reflection phase theta within the t data frame time is predicted as the input of the trained neural network, and the prediction result is recorded as theta_(t,n)And b is 1,2, and K, n is the serial number of the current reflection unit.

And step 4, the base station controls and adjusts the phase configuration of the intelligent reflection surface according to the predicted phase configuration, and the intelligent reflection surface is used for enhancing and receiving the data part in the data frame.

Step 5, the base station receives the data part of the t-th data frame,carrying out signal detection, measuring the power of data signal, and recording the measurement result as P_cCalculating the power gainWhere G (x) is a monotonically increasing function with respect to x.

Step 6, the base station sends the channel state information H_t,H_t-1And reflection phase configuration theta of the intelligent reflection surface when receiving pilot frequency_(t-1,n)，θ_(t-2,n)As input, the prediction result θ obtained_(t,n)As an output, a new sample is created, denoted M_t。

Step 7, the base station according to the new sample M_tConstructing a new sample set R_t. For all constructed samples M_kK is 1,2,.. times, T, extracting by probability W times, constructing a sample set with W number of samples, and the probability that the kth sample is extracted may be

Where τ (T-k) is a decay function, a monotonically decreasing function with respect to (T-k).

Step 8, the base station uses the constructed sample set R_tFurther training the existing neural network model, wherein the maximum iteration number of the training isWhere F (x) is a monotone decreasing function for x and the trained neural network is saved.

And 9, repeating the steps 1-8 when receiving the next data frame.

The downlink data transmission comprises the following steps:

step 1, sending data frame from base station end to user end, where the base station end has N antennas, the user end has M antennas, each data frame includes L time slots, where the first L time slots_pOne slot is a pilot signal, the remainder (L-L)_p) The time slots being part of a data signal, data frameThe sub-power is reflected by an intelligent reflecting surface with K reflecting units and reaches a user side.

Step 2, the user end receives the pilot frequency part in the current data frame and utilizes L_pPilot frequency time slot, obtaining channel state information matrix H_t∈C^M×NT is the sequence number of the current data frame, and C represents a complex number field. Meanwhile, the user measures the power of the pilot signal, and the measurement result is marked as P_t。

Step 3, the user receives the data part of the t-th data frame, carries out signal detection, sends a data frame to the base station end, and sends the pilot signal power P_tAnd channel state information matrix H_tAnd sending the data to a base station end.

Step 4, at the base station end, for the t data frame, H_t,H_t-1,θ_(t,n)，θ_(t-1,n)The intelligent reflection surface reflection phase theta within the t data frame time is predicted as the input of the trained neural network, and the prediction result is recorded as theta_(t+1,n)N is the serial number of the current reflection unit. Calculating power gainWhere G (x) is a monotonically increasing function with respect to x.

And 5, controlling by the base station, and adjusting the phase configuration of the intelligent reflection surface according to the predicted phase configuration for enhanced reception of a data part in the data frame.

Step 6, the base station sends the channel state information H_t,H_t-1And reflection phase configuration theta of the intelligent reflection surface when receiving pilot frequency_(t,n)，θ_(t-1,n)As input, the prediction result θ obtained_(t+1,n)As an output, a new sample is created, denoted M_t。

Step 7, the base station according to the new sample M_tConstructing a new sample set R_t. For all constructed samples M_kK is 1,2,.. times, T, extracting by probability W times, constructing a sample set with W number of samples, and the probability that the kth sample is extracted may be

Where τ (T-k) is a decay function, a monotonically decreasing function with respect to (T-k).

Step 8, the base station uses the constructed sample set R_tFurther training the existing neural network model, wherein the maximum iteration number of the training isWhere F (x) is a monotone decreasing function for x and the trained neural network is saved.

And 9, repeating the steps 1-8 when receiving the next data frame.

By adopting the technical scheme, the invention can produce the following technical effects:

the invention provides a method for configuring auxiliary wireless communication reflection phase by utilizing a neural network to adjust an intelligent reflection surface in real time, which can use a neural network model to train the neural network by adopting instantaneous channel state information acquired in real time and power gain of signals, and adjust the phase allocation of the intelligent reflection surface in real time in the process of receiving data so as to play a role in optimizing the quality of received signals. Compared with the prior art, the method has the following advantages:

(1) the reflection phase configuration only requires knowledge of the channel information from the transmitter to the receiver and does not require knowledge of the channel incident from the transmitter to the intelligent reflective surface and reflected back through the intelligent reflective surface to the receiver, thus reducing the number of variables to be estimated.

(2) In the configuration of the intelligent reflection surface phase, the selection of the neural network model can be changed, the better neural network model can effectively improve the quality of the received signal, the selection can be made between the complexity and the effect of the model according to the requirement of actual engineering, and a space for improving the configuration method is provided.

(3) The neural network parameters updated in real time can enable the predicted optimal configuration to be better adapted to the current scene, and the intelligent reflecting surface can obtain a better optimization effect in the scene with continuously changing channel information.

(4) In the scheme of the reflection phase configuration, the information transmission performance is better than that under the method of randomly configuring the phase.

Drawings

FIG. 1 is a flow chart of a method for configuring a reflection phase for assisting wireless communication by adjusting an intelligent reflection surface in real time through a neural network according to the present invention.

FIG. 2 is a flow chart of a method for updating a neural network sample set by using a neural network to adjust the configuration of the reflection phase of the assisted wireless communication of the intelligent reflection surface in real time.

Fig. 3 is a graph illustrating a variation of the reflected power indicator with the iteration number i in embodiment 1 of the present invention.

Fig. 4 is a graph illustrating a variation of the reflected power indicator with the iteration number i in embodiment 2 of the present invention.

Detailed Description

The following describes embodiments of the present invention with reference to the drawings.

The invention is applicable to the following scenes: the intelligent reflection surface assisted wireless communication system comprises an uplink communication scene and a downlink communication scene, wherein the uplink communication scene and the downlink communication scene comprise a sending end, a receiving end and one or more intelligent reflection surface modules; the transmitting end is provided with N antennas, the receiving end is provided with M antennas, all intelligent reflection surface modules contain reflection units with the total number of K, all the intelligent reflection surface modules are connected with the receiving end through links, and the receiving end is used for connecting reflection coefficient phase theta of each reflection unit N_nN is 1, 2. The transmitting end transmits data to the receiving end, and the instantaneous channel impulse responses of all the paths are recorded as H e C^M×NThe transmission signal goes through two paths to the receiving end: one is a direct arrival path, and the instantaneous channel impulse response is denoted as h_dE C, where C represents a complex field; the other is reflected by the intelligent reflection surface unit and reaches the receiving end, and the instantaneous channel impulse response is recorded as h_rC, which contains a cascade of three parts: sending end to each reflection sheetChannel impulse response h of element_in∈C^K×NAdditional phase shift θ per reflection unit_nN1, 2.. N, the channel impulse response h experienced by the reflection unit reflecting the signal to the receiving antenna_out∈C^M×KThus, h_r＝h_outΘh_inWhere Θ ∈ C^K×KIs formed byA diagonal matrix is formed.

During communication, a signal received by a receiving end in any time slot can be represented as:

y＝Hx+z＝(h_d+h_r)x+z＝(h_d+h_outΘh_in)x+z

where y ∈ C, x ∈ C, and z ∈ C are a reception signal, a transmission signal, and a noise signal, respectively. h is_d，h_out，h_inCannot be adjusted depending on the physical propagation channel, but the intelligent reflective surface assists the wireless communication system to actively adjust Θ to change the resultant equivalent channel h_rE.g., C, different Θ will result in different quality of the received signal. Therefore, the key of the system design is how the receiving end dynamically adjusts the reflection phase θ of the reflection unit according to the received signal and the channel state information_nN is 1, 2.. times.n, so that the average power of the received signal reaches a maximum value. The existing literature assumes channel state information h_in∈C^{K ×N}And h_out∈C^M×KIt is known that in practical application, it is difficult to obtain the two pieces of channel state information, and only the channel state information H e C from the transmitting end to the receiving end can be obtained^M×NIf adopted, is based on h_in∈C^K×NAnd h_out∈C^M×KThe known method for configuring the reflection phase is difficult to apply to practical application. Homogeneous channel state information h due to reflected paths_in∈C^K×NAnd h_out∈C^M×KChannel state information H ∈ C hidden from transmitting end to receiving end^M×NIn the method, one neural network which can be suitable for all the neural networks is difficult to train by using the existing neural networkA model of a channel scene.

Aiming at the problem, the invention provides the configuration method for adjusting the reflection phase of the auxiliary wireless communication of the intelligent reflection surface in real time by utilizing the neural network, and considering that the application scene of the intelligent reflection surface is indoor, the change of the channel state information is small in a long time, and the channel state information H belonging to C can be obtained^M×NAnd receiving the power of data, adjusting a sample set of the neural network in real time, continuously training the neural network and finishing the optimization of the phase configuration of the intelligent reflection surface, and continuously optimizing the power of the received signal while obtaining better parameters of the neural network model.

Specifically, as shown in fig. 1, a method for configuring a reflection phase for assisting wireless communication by adjusting an intelligent reflection surface in real time through a neural network includes the following steps:

step 1, sending data frame from user end to base station end, where the user end has N antennas, the base station end has M antennas, each data frame includes L time slots, where the first L time slots_pOne slot is a pilot signal, the remainder (L-L)_p) And each time slot is a data signal, and partial power of a data frame is reflected by an intelligent reflecting surface with K reflecting units and reaches a base station end.

Step 2, the base station receives the pilot frequency part in the current data frame and utilizes L_pPilot frequency time slot, obtaining channel state information matrix H_t∈C^M×NT is the sequence number of the current data frame, and C represents a complex number field. Meanwhile, the base station measures the pilot signal power, and the measurement result is marked as P_r。

Step 3, at the base station end, for the t data frame, H_t,H_t-1,θ_(t-1,n)，θ_(t-2,n)The intelligent reflection surface reflection phase theta within the t data frame time is predicted as the input of the trained neural network, and the prediction result is recorded as theta_(t,n)N is the serial number of the current reflection unit.

And step 4, the base station controls and adjusts the phase configuration of the intelligent reflection surface according to the predicted phase configuration, and the intelligent reflection surface is used for enhancing and receiving the data part in the data frame.

Step 5, the base station receives the data part of the t-th data frame, performs signal detection, and measures the power of the data signal, and the measurement result is recorded as P_cCalculating the power gainWhere G (x) is a monotonically increasing function with respect to x.

Step 6, the base station sends the channel state information H_t,H_t-1And reflection phase configuration theta of the intelligent reflection surface when receiving pilot frequency_(t-1,n)，θ_(t-2,n)As input, the prediction result θ obtained_(t,n)As an output, a new sample is created, denoted M_t。

Step 7, the base station according to the new sample M_tConstructing a new sample set R_t. For all constructed samples M_kK is 1,2,.. times, T, extracting by probability W times, constructing a sample set with W number of samples, and the probability that the kth sample is extracted may be

Where τ (T-k) is a decay function, a monotonically decreasing function with respect to (T-k).

Step 8, the base station uses the constructed sample set R_tFurther training the existing neural network model, wherein the maximum iteration number of the training isWhere F (x) is a monotone decreasing function for x and the trained neural network is saved.

And 9, repeating the steps 1-8 when receiving the next data frame.

The downlink data transmission comprises the following steps:

step 1, sending data frame to user end by base station end, wherein base station end has N antennas, user end has M antennas, each data frameComprising L time slots, wherein the first L_pOne slot is a pilot signal, the remainder (L-L)_p) Each time slot is a data signal, and partial power of a data frame is reflected by an intelligent reflection surface with K reflection units and reaches a user side.

Step 2, the user end receives the pilot frequency part in the current data frame and utilizes L_pPilot frequency time slot, obtaining channel state information matrix H_t∈C^M×NT is the sequence number of the current data frame, and C represents a complex number field. Meanwhile, the user measures the power of the pilot signal, and the measurement result is marked as P_t。

Step 3, the user receives the data part of the t-th data frame, carries out signal detection, sends a data frame to the base station end, and sends the pilot signal power P_tAnd channel state information matrix H_tAnd sending the data to a base station end.

Step 4, at the base station end, for the t data frame, H_t,H_t-1,θ_(t,n)，θ_(t-1,n)The intelligent reflection surface reflection phase theta within the t data frame time is predicted as the input of the trained neural network, and the prediction result is recorded as theta_(t+1,n)N is the serial number of the current reflection unit. Calculating power gainWhere G (x) is a monotonically increasing function with respect to x.

And 5, controlling by the base station, and adjusting the phase configuration of the intelligent reflection surface according to the predicted phase configuration for enhanced reception of a data part in the data frame.

Step 6, the base station sends the channel state information H_t,H_t-1And reflection phase configuration theta of the intelligent reflection surface when receiving pilot frequency_(t,n)，θ_(t-1,n)As input, the prediction result θ obtained_(t+1,n)As an output, a new sample is created, denoted M_t。

Step 7, the base station according to the new sample M_tConstructing a new sample set R_t. For all constructed samples M_k，k＝1，2，...，T, extracting for W times according to the probability to construct a sample set with the number of samples W, wherein the probability of extracting the kth sample can be

Where τ (T-k) is a decay function, a monotonically decreasing function with respect to (T-k).

Step 8, the base station uses the constructed sample set R_tFurther training the existing neural network model, wherein the maximum iteration number of the training isWhere F (x) is a monotone decreasing function for x and the trained neural network is saved.

And 9, repeating the steps 1-8 when receiving the next data frame.

Therefore, the invention only uses the channel state information obtained by the receiving end, and continuously updates the sample set of the neural network, thereby continuously optimizing the parameters of the neural network, adjusting the configuration of the reflection coefficient of the reflection element and achieving the purpose of enhancing the signal power.

To verify that the method of the present invention can rapidly update the reflection coefficient configuration, the following two examples are presented for verification.

Example 1:

the embodiment shows how to realize the method for utilizing the neural network to adjust the intelligent reflecting surface in real time to assist the wireless communication reflection phase configuration. In this embodiment, an uplink data transmission scenario is adopted, the number K of units of the intelligent reflection surface is 64, the units are regularly distributed in space, the number of antennas at the transmitting end is 1, the number of antennas at the receiving end is 3, and h is an integer_d，h_out，h_inIs a Rayleigh distributed channel, assuming that the channel state information H ∈ C in each data frame^M×NNo change, channel state information h after each transmission of one frame_d，h_out，h_inSlight change occurs, the power of the transmitted signal is kept unchanged, and the signal is transmittedThe emitted energy reaches 50% through the direct path, and 50% through the reflection of the intelligent reflection surface. Each frame comprises 32 time slots, wherein the first 8 time slots transmit pilot signals, the remaining 24 time slots transmit data signals, and a total of 10 frames of signals are transmitted.

The neural network is pre-trained, the reflected power of the initial phase configuration is higher than that of the random configuration phase, after receiving a new pilot signal, the predicted optimal phase configuration is obtained by using the training of the existing neural network parameters, and then the power gain G obtained by the receiving end is obtained_tUpdating a neural network model sample set, further training the neural network, and specifically updating the data set as follows:

s1: calculating power gain coefficient G obtained by receiving end_tWherein G is_tIs composed ofP_cFor the t-th pilot, the power of the received data part, P, after receiving the pilot signal and updating the phase configuration_rG (x) is a monotonically increasing function with respect to x for the received power of the pilot signal in the t-th data frame.

S2: for the t-th pilot received, construct sample M_tThe channel state information matrix H when the tth and t-1 pilots are to be transmitted_t∈C^M×N，H_t-1∈C^M×NAnd the phase configuration of the intelligent reflecting surface when receiving the pilot frequency when transmitting the tth and t-1 pilot frequenciesComposition sample M_tThe phase configuration calculated by the trained neural network modelComposition sample M_tForming a sample M of the training neural network_t。

S3: for all constructed samples M_kK 1,2, T, extracting W times with probability to construct a unitThe probability that the kth sample is extracted can be W

Where τ (T-k) is a decay function, a monotonically decreasing function with respect to (T-k).

S4: further training the existing neural network model parameters according to the selected neural network model by using the constructed sample set, wherein the maximum iteration number of each time isWhere F (x) is a monotonic decreasing function with respect to x.

By updating the data set, the parameters of the neural network are continuously optimized and adjusted during the transmission and reception of the signals. Recording the average power indicator J of the pilots received in each frame_t＝|h_d|²+|h_outΘ_th_in|²This is compared to the average power index obtained by randomly generating the reflection phase and plotted to obtain fig. 3. As shown in fig. 3, after the reflection coefficient configuration method provided by the present invention is used, since the neural network model is pre-trained, the average reflection power obtained by the receiving end rapidly increases in the initial stage of receiving signals, and tends to be stable after the 5 th frame, and better received power is maintained, and the reflection power obtained by the reflection coefficient configuration method of the present invention is about 1.4 times of the average reflection power when the phase is randomly configured, which illustrates the effectiveness of the present invention.

Example 2:

the embodiment shows how to realize the method for utilizing the neural network to adjust the intelligent reflecting surface in real time to assist the wireless communication reflection phase configuration. In this embodiment, an uplink data transmission scenario is adopted, the number K of units of the intelligent reflection surface is 64, the units are regularly distributed in space, the number of antennas at the transmitting end is 1, the number of antennas at the receiving end is 3, and h is an integer_d，h_out，h_inIs a Rayleigh distributed channelSuppose that the channel state information H ∈ C in each data frame^M×NNo change, channel state information h after each transmission of one frame_d，h_out，h_inThe power of the transmitted signal is kept unchanged by slight change, 50% of transmitted energy reaches through the direct path, and 50% of transmitted energy reaches through the reflection of the intelligent reflection surface. Each frame comprises 32 time slots, wherein the first 8 time slots transmit pilot signals, the remaining 24 time slots transmit data signals, and a total of 10 frames of signals are transmitted.

The neural network is not pre-trained, the reflected power of the initial phase configuration is lower than that of the random configuration phase, after a new pilot signal is received, the predicted optimal phase configuration is obtained by using the existing neural network parameter training, and then the power gain G obtained by the receiving end is obtained_tUpdating a neural network model sample set, further training the neural network, and specifically updating the data set as follows:

s1: calculating power gain coefficient G obtained by receiving end_tWherein G is_tIs composed ofP_cFor the t-th pilot, the power of the received data part, P, after receiving the pilot signal and updating the phase configuration_rG (x) is a monotonically increasing function with respect to x for the received power of the pilot signal in the t-th data frame.

S2: for the t-th pilot received, construct sample M_tThe channel state information matrix H when the tth and t-1 pilots are to be transmitted_t∈C^M×N，H_t-1∈C^M×NAnd the phase configuration of the intelligent reflecting surface when receiving the pilot frequency when transmitting the tth and t-1 pilot frequenciesComposition sample M_tThe phase configuration calculated by the trained neural network modelComposition sample M_tForming a sample M of the training neural network_t。

S3: for all constructed samples M_kK is 1,2,.. times, T, extracting by probability W times, constructing a sample set with W number of samples, and the probability that the kth sample is extracted may be

Where τ (T-k) is a decay function, a monotonically decreasing function with respect to (T-k).

S4: further training the existing neural network model parameters according to the selected neural network model by using the constructed sample set, wherein the maximum iteration number of each time isWhere F (x) is a monotonic decreasing function with respect to x.

By updating the data set, the parameters of the neural network are continuously optimized and adjusted during the transmission and reception of the signals. Recording the average power indicator J of the pilots received in each frame_t＝|h_d|²+|h_outΘ_th_in|²This is compared to the average power index obtained by randomly generating the reflection coefficient and plotted to obtain fig. 3. As shown in fig. 3, although the neural network model is not pre-trained after the reflection coefficient configuration method provided by the present invention is used, the average reflection power obtained by the receiving end rapidly increases in the initial stage of receiving signals, and tends to be stable after the 5 th frame, and better received power is maintained, and the reflection power obtained by the reflection coefficient configuration method of the present invention is about 1.1 times of the average reflection power when the phase is randomly configured, which illustrates the effectiveness of the present invention.

In conclusion, the method is applied to an indoor scene with slowly-changed channel state information, and can quickly predict the optimized optimal phase configuration of the intelligent reflection surface by continuously optimizing neural network model parameters under the condition that information of a channel which is incident from a transmitter to the intelligent reflection surface and is reflected to a receiver through the intelligent reflection surface is not required to be obtained, so that the power of a received signal is optimized.

The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.